Online program Overview Session: MIC-15

Abstract

Spatial resolution and sensitivity are the two most critical characteristics of a positron emission tomography (PET) scanner, as they determine the accuracy and precision of reconstructed images. Achieving uniform ultra-high spatial resolution across the field of view (FOV) while maintaining sufficient sensitivity requires PET detectors capable of resolving ultra-fine, high-aspect-ratio crystals and extracting precise depth-of-interaction (DOI) information to mitigate parallax errors. In this study, we developed and evaluated two dual-ended readout detectors for small-animal brain PET applications. Each detector comprised a 36 × 36 array of lutetium yttrium oxyorthosilicate (LYSO) or gadolinium aluminum gallium garnet (GAGG) coupled to two linearly-graded silicon photomultipliers (LG-SiPMs). The crystal element size of the two crystal arrays was 0.20 × 0.20 × 9 mm³, while the pitch of the two crystal arrays was 0.25 mm. Despite the crystal element aspect ratio reaching 45:1, most crystal elements in both detectors were resolved, yielding average energy resolutions of 37.3 ± 5.6% and 31.7 ± 5.7%, respectively, for the LYSO and GAGG arrays, resepectively. The mean coincidence timing resolutions (CTRs) of the central 32 × 32 crystal elements, measured using a 200–1000 keV energy window to select events, were 1025 ± 40 ps (LYSO array) and 1205 ± 62 ps (GAGG array). Additionally, the average DOI resolutions across crystal elements were 1.48 ± 0.10 mm and 1.47 ± 0.10 mm, respectively, for the LYSO and GAGG arrays, respectively.

Abstract

High-repetition-rate (HRR) operation is a key limitation in laser-driven particle and photon sources, requiring diagnostics compatible with repetition rates of order 1 Hz or higher. We investigate standard industrial CMOS imaging sensors as soft X-ray detectors, analysing their single-photon response in the 5.5--12.6 keV range. The response is linear, and the sensors are calibrated in terms of generated electron--hole carriers versus incident photon energy. X-ray interactions in pixel sensors produce single-pixel (SPE) and multi-pixel events (MPE), which are analysed independently; removing MPEs significantly improves spectral resolution. This work is a first step toward the systematic study of commercial CMOS sensors under exposure to high-energy particles and photons, including response, radiation tolerance and dynamic range. Their compact size, low cost, and intrinsic compatibility with high-repetition-rate operation make them suitable for laser-plasma, nuclear and particle physics, and biomedical applications.

Abstract

To address spatial resolution and sensitivity limitations in whole-body PET for small-organ imaging, such as heart, breast and prostate, this study developed an asymmetric dual-ended readout for TOF-DOI PET detectors.

The asymmetric readout design utilizes a 4×4 SiPM array (6×6 mm² pixels) with individual channel readout at the front end, for good timing performance . At the back end, an 8×8 SiPM array (3×3 mm² pixels) is read out using a diode-based row-column summing multiplexing circuit to facilitate event position decoding while reducing the number of readout channels. Specifically, the multiplexing scheme reduces the back end channels from 64 to 16, lowering the total channels per detector from 80 to 32. LYSO arrays with three crystal dimensions were evaluated: an 8×8 array (3.10×3.10×20 mm³ crystals), a 13×13 array (1.88×1.88×20 mm³ crystals), and a 16×16 array (1.50×1.50×20 mm³ crystals).

Detector performance was evaluated in an array-to-array coincidence setup, with two identical detector modules placed face-to-face. Flood images show all crystals in the three evaluated arrays can be clearly resolved under both non-multiplexed and multiplexed readout conditions. The average energy resolutions for the 8×8, 13×13, and 16×16 arrays are 12.49%, 11.71%, and 10.92%, respectively, which are comparable to those obtained with non-multiplexed readout . A preliminary evaluation of CTR for the 8×8 array yields 327 ps (non-multiplexed) and 365 ps (multiplexed), indicating limited timing degradation due to multiplexing.

The proposed asymmetric multiplexed readout design maintains acceptable spatial, energy, and timing resolution while substantially reducing electronic complexity . This configuration is suitable for developing cost-effective, high-performance organ-specific PET scanners . Future work will focus on optimizing the MPPC operating overvoltage and PETsys threshold settings, as well as implementing an interleaved multiplexing circuit to improve CTR performance.

Abstract

Time-of-flight (TOF) and depth-of-interaction (DOI) information improve the spatial resolution and signal-to-noise ratio of PET imaging, but typically requires numerous readout channels, increasing system complexity and cost.

In this work, we developed a TOF-DOI PET detector using a novel dual-ended interleaved multiplexing (DiMux) readout. The detector consists of an 8×8 LYSO crystal array with polished surfaces, where each crystal measures 3.1 × 3.1 × 20 mm³ with a 3.2 mm crystal pixel pitch, and is optically isolated using BaSO₄ reflectors. Both ends are one-to-one coupled to 8×8 SiPM arrays and read out by two complementary interleaved multiplexing circuits, which use Schottky barrier diode-based charge division. In this approach, each crystal is encoded as a row signal on one end and a column signal on the opposite end, with multiplexed signals digitized by PETsys TOFPET2 ASICs. This dual-ended, complementary design enables accurate position decoding using a conventional center-of-gravity (CoG) method, eliminating the need for deep learning-based decoding.

Flood images with clearly resolved crystals are generated from the summed signals of both ends. The detector achieves a average DOI resolution of 2.97 mm and a coincidence time resolution (CTR) of 257 ps, which further improves to 234 ps after excluding inter-crystal scattering (ICS) events. Compared with a non-multiplexed dual-ended reference detector, the DiMux scheme reduces the number of readout channels by a factor of four while maintaining comparable TOF and DOI performance.

These results demonstrate that the DiMux readout provides a high performance, low complexity, and scalable solution for clinical PET imaging systems.

Abstract

The purpose of this study was to design a single-ended readout PET detector with BGO providing DOI, to investigate the effects of lateral surface treatments and inter-crystal coupling conditions on detector performance, and to improve the performance using deep learning. A detector consisting of an 8 × 8 array of 3 mm × 3 mm × 20 mm BGO, each one-to-one coupled with a SiPM, was modeled using DETECT2000. Photon distributions were estimated for four detector configurations: polished + air (PA), polished + glue (PG), unpolished + air (UA), and unpolished + glue (UG). The x-, y-, and z-positions were analytically estimated using photon distribution information. To improve positioning accuracy, a deep learning model combining CNN and MLP was implemented. The CNN extracted features from photon distributions, and the MLP estimated the fine position using CNN-derived features concatenated with analytically calculated positions. The difference between the ground-truth and the estimated positions was evaluated using the mean absolute error (MAE). The accuracy of DOI estimation was assessed by evaluating the DOI identification accuracy in 20 mm long BGO divided into three equal segments. For x- and y-position estimations, the MAE values showed minimal variations (< 0.1 mm) across the four detector configurations, regardless of the estimation methods. For DOI estimation, the UA and UG achieved lower MAE (3.8 and 3.6 mm) and higher DOI identification accuracies (63% and 66%) than the other detector configurations. By applying deep learning, the MAE values for UA and UG were reduced to approximately 1 mm, and the DOI accuracies increased to 92% and 91%, respectively. These results indicate that three-level DOI identification can be achieved using photon distribution information obtained from a simple single-ended readout detector. In particular, the UA provided superior DOI estimation performance, and the incorporation of deep learning further enhanced DOI positioning accuracy.

Abstract

To investigate potential applications of indirect photon-counting detectors, we developed a model to simulate pulse trains with pile-up for a combination of scintillator and silicon photomultiplier (SiPM). We generated pulse trains for various input rates, and the output count rate (OCR) and the spectral degradation factor (SDF) were analyzed. This was performed for two scintillators and counting algorithms, various cut-off frequencies of the applied low-pass filter, event energy thresholds, and a mono- and a polychromatic source. The results were compared to experiments. For the OCR, the model and measurements produced very similar results for all investigated parameter combinations. While the SDF values of the simulations tended to show some differences compared to the measurements, the general trend was similar.

Abstract

This work investigates a collimator-less scintillator-based detector that uses penumbra maps to reconstruct the spatial origin of radioisotopes without photon losses from collimation, motivated by the need for in-vivo dosimetry monitoring during targeted radionuclide therapy (TRT) and the limitations for its broader clinical adoption. Monte Carlo simulations were performed in GATE to evaluate the method's viability, using a 218 keV mono-energetic point-like source to replicate the first gamma emission on Ac-225's decay scheme. Given its favorable photoelectric fraction for this energy, BGO was chosen as the detector's material on the simulations. Geometrical parameters such as crystal dimensions, inter-crystal spacing, and array layout were systematically varied. Results show that inter-crystal spacing significantly shapes the photon distribution pattern, the larger 0.4mm value produced a distinct visual artifact. Depth profiles vary with source position in a predictable manner and provide information regarding the radiation's penetration depth. The implemented dual-layer checkerboard geometry demonstrated complementary detection behavior between layers, improving spatial discrimination over single-layer designs. These findings support the feasibility of collimator-less gamma imaging and show that geometric optimization shapes the detector's spatial response.

Abstract

CTR in PET detectors is a quadrature combination of scintillator photon statistics, intrinsic detector physics, and front-end circuit electronics, but conventional measurement reports only the combined FWHM, leaving the relative contributions indeterminate. This hinders rational optimization between circuit-side and detector-side improvements. We introduce a covariance-based framework that decomposes the CTR noise budget from a single coincidence measurement by simultaneously extracting timestamps from two independent readouts—the timing chain and the energy chain. The covariance Cov(ΔT_T, ΔT_E) isolates the common-mode intrinsic contribution, while variance differences identify chain-specific circuit noise in closed form, requiring no calibration beyond standard CTR acquisition. Applied to a 4 condition of LSO and GAGG:Ce (3×3×20 mm³) with Broadcom NUV-MT and NUV-DJ SiPMs, the framework extracts a circuit noise floor σ_nT = 11.2–18.5 ps per detector across all conditions, supporting its interpretation as a detector-independent electronics floor. Intrinsic contributions cluster sharply by crystal type (LSO 28–31 ps, GAGG 49–56 ps), exceeding the analytical CRLB by 1.42–1.62× due to DOI and transport effects. A walk asymmetry ρ(ΔT_E, dE)/ρ(ΔT_T, dE) = 1.42–1.82× reflects front-end pulse-shape physics. The decomposition quantifies circuit-improvement leverage: LSO configurations retain 11–15% potential CTR reduction, while GAGG configurations operate within 2–6% of their photon-statistics bound. The framework provides a single-measurement diagnostic for prioritizing circuit-side versus detector-side investments in PET front-end development, and also provides the first quantitative noise-budget characterization of NUV-DJ SiPMs.

Abstract

Stacked scintillator configurations to obtain depth-of-interaction (DOI) information, combined with small crystal pitch, are commonly used to improve spatial resolution and maintain uniform performance across the field of view. Such configurations are prone to inter-crystal scatter (ICS), where energy deposition is shared across multiple crystals, leading to event mis-assignment and degradation of detector performance, including incorrect DOI estimation. Correction of these effects typically benefits from individual channel readout, which is often limited in highly multiplexed systems. In this work, we investigate ICS in a dual-layer LSO detector with 1.6 mm pitch, employing a 10×10/9×9 stacked crystal configuration coupled to a 4×4 SiPM array. The detector uses a highly multiplexed readout scheme with only three output channels (x, y, and total energy), imposing additional constraints on event discrimination. To analyze ICS effects, a hybrid simulation framework was developed, pairing Geant4-based modeling of gamma interactions with custom light transport and readout model. The simulations reproduce characteristic structures observed in experimental data, including patterns in floodmaps and energy-dependent light spread distributions. Based on these results, a light-pattern-based metric was introduced to identify ICS events using an energy-dependent threshold derived from single-interaction event distributions. This approach indicates that 32.9 ± 5.5% of events per crystal are affected by ICS and assigned to a non-optimal crystal, including mis-assignment between layers. Reassignment of these events to the most likely interaction crystal increased effective sensitivity by 4.1% and improved energy resolution from 10.6 ± 1.8% to 10.2 ± 1.8% FWHM. These results demonstrate that ICS as major effect in finely pixelated, DOI-capable detectors and that its identification and partial correction is feasible under highly multiplexed readout conditions.

Abstract

We present a modular FPGA-based data acquisition and real-time processing architecture for SiPM-based PET detectors, developed in the context of compact, low-cost PET imaging. The system combines a multiplexed readout of a SiPM array with free-running high-speed ADCs and on-chip FPGA processing to extract event energy, timestamp, and interaction position in real time. The approach emphasizes compactness, scalability, and low-latency operation, while reducing channel count and system complexity. First measurements demonstrate correct operation and clear crystal identification in flood maps, highlighting both the potential and current limitations of SiPM multiplexing, such as the limited CTR. This work aims to enable cost-efficient, flexible, and scalable PET imaging systems.

Abstract

Photon-counting computed tomography (PCCT) is a rapidly emerging (medical) imaging technology. PCCT commonly relies on finely pixelated room-temperature semiconductor detectors (RTSD). Scintillation detectors have long been considered too slow to handle the high X-ray fluence rates encountered in clinical PCCT. However, the fast signals provided by modern scintillators and silicon photomultipliers (SiPMs) allow us to reconsider this paradigm. SiPM-based scintillation detectors offer excellent image quality and are known to operate reliably in clinical time-of-flight positron emission tomography (TOF-PET) scanners. Recent research indicates that the combination of ultrafast SiPMs and scintillators could potentially provide a robust, scalable, and affordable X-ray photon-counting detector (PCD) technology. Here, we evaluate a novel type of silicon photomultiplier (SiPM) for this application. The SiPM is based on small (< 15 m) single-photon avalanche diodes (SPADs), which have an innovative design that enables a fast recharge time (< 7 ns) as well as a high geometric efficiency (fill factor), resulting in high photodetection efficiency (PDE) (~ 50%) over a broad wavelength range. These benefits can be exploited to mitigate the classical trade-off between dynamic range and PDE, which is of high interest for X-ray photon-counting applications. We perform experiments with different design variations of the new SiPM, coupled to small pixels of commercially available scintillators (e.g. LYSO:Ce, LaBr3:Ce). We report on the pulse shape, count-rate capability, and spectral performance achieved with different combinations of SPAD design and scintillator.

Abstract

We present the design and characterization of a 11.6 million pixel, 8 µm pitch, low noise (1.5 e rms), high dynamic range (93.5 dB), backside-illuminated (BSI) CMOS image sensor with microlenses (M) coupled to a scintillator for X-ray detection. Characterisation with a microfocus X-ray source shows high MTF, in excess of 30-line pairs per mm (lp/mm), while operating at an air kerma of 20.9 mGy per image.

Abstract

This work presents the development of a semi-monolithic laser-engraved PET detector with dual-ended readout for continuous Depth of Interaction (DOI) estimation. The detector is based on a monolithic LYSO scintillator structured using Subsurface Laser Engraving (SSLE) techniques, enabling controlled light sharing between neighboring pixels while pre-serving the simpler readout and calibration typically associated with pixelated detectors, together with the accurate interaction vertex determination characteristic of monolithic designs. The main objective of this work is to determine the optimal placement of a reduced number of photodetectors in order to maximize DOI performance while minimizing the system’s complexity. The study is based on the direct experimental measurement of light sharing inside the scintillator crystal. For this purpose, two 61- channel SiPM matrices are coupled to both faces of a LYSO scintillator matrix featuring a honeycomb laser-engraved pattern of 181 crystals. Taking advantage of the fact that the existing electronics can process up to 64 channels, the light distribution measured on one side of the detector is combined with the signal collected by only three selected photodetectors on the opposite side to estimate the DOI and evaluate different configurations for these additional channels. Using fully instrumented matrices on both faces allows the evaluation of different three-SiPM configurations and provides a detailed understanding of scin-tillation light-sharing behavior. The proposed approach aims to identify compact dual-ended configurations capable of achieving continuous DOI estimation with reduced readout complexity for future high-resolution PET detectors.

Abstract

We report updates on the development and current progress of PETVision, a modular flat-panel positron emission tomography (PET) system designed to achieve ultra-high Time-of-Flight (TOF) resolution. Traditional cylindrical PET scanners often face limitations due to high crystal volume costs and geometries that restrict simultaneous medical intervention. By leveraging unprecedented Coincidence Time Resolution (CTR), the PETVision architecture enables limited-angle, flat-panel geometries that reduce material costs while maintaining high image resolution.

The Photo-Detection Module (PDM) of the PET system is utilizing a 2.5D integration configuration, where Silicon Photomultiplier (SiPM) arrays and readout ASICs are coupled to opposite sides of a passive interposer. The sensing layer employs state-of-the-art NUV-HD-MT technology, featuring custom "glass-less" Through Silicon Vias (TSVs) to achieve a Photon Detection Efficiency (PDE) of up to 65% and a Single Photon Time Resolution (SPTR) with Full Width at Half Maximum (FWHM) as low as 20 ps. Readout electronics development is proceeding in two stages: an initial 32-channel FastIC+ BGA chiplet demonstrating a CTR of (85±1) ps FWHM with 2×2×3 mm3 LYSO:Ce:0.2%Ca scintillator crystals. This is followed by the 64-channel Positron Emission Tomography Integrated Circuit (PoETIC) ASIC, which targets an electronics jitter contribution below 12 ps r.m.s. to support the 10-picosecond challenge. This modular approach aims to provide a high-performance, cost-effective PET imaging system suitable for diverse clinical and research applications, including simultaneous PET/intervention procedures.

Abstract

Recently developed Broadcom NUV-DJ SiPMs have improved photon detection efficiency (PDE) by placing the high-field region deeper than that of NUV-MT SiPMs. PDE is one of the major factors affecting coincidence timing resolution (CTR), and its performance varies depending on the scintillation light wavelength. Therefore, this study comparatively evaluated the CTR performance of LYSO:Ce, GAGG:Ce, and BGO scintillators, which have different peak emission wavelengths, coupled with NUV-MT and NUV-DJ SiPMs. The detectors consisted of 6 × 6 mm² NUV-MT and NUV-DJ SiPMs coupled with 3 × 3 × 20 mm³ LYSO:Ce and GAGG:Ce crystals and 3 × 3 × 3 mm³ BGO crystals using optical grease. Timing and energy signals were processed using a high-frequency (HF) readout board and acquired using an oscilloscope. CTR measurements were performed as a function of overvoltage for each detector. In addition, quartile-based rise-time analysis was conducted for the BGO detectors to investigate the effect of prompt Cherenkov photon detection on timing performance as indirect indicators. The detectors coupled with NUV-MT SiPMs achieved CTRs of 109.0 ps (LYSO:Ce), 182.8 ps (GAGG:Ce), and 210.8 ps (BGO) at an overvoltage of 12 V, whereas those coupled with NUV-DJ SiPMs achieved CTRs of 103.0 ps (LYSO:Ce), 172.4 ps (GAGG:Ce), and 187.2 ps (BGO), respectively. Compared with the NUV-MT SiPMs, the NUV-DJ SiPMs improved CTR by 5.2%, 4.5%, and 11.2% for LYSO:Ce, GAGG:Ce, and BGO, respectively. The quartile analysis further showed improved timing performance for fast rise-time events when using the NUV-DJ SiPMs. These results indicate that the improved PDE of the NUV-DJ SiPM enhances prompt photon detection and contributes to improved CTR performance for BGO scintillators. Therefore, the proposed detector configuration may be beneficial for future TOF-PET and BGO-based timing detector applications.

Abstract

Monte Carlo simulations of optical photon transport are the gold standard for modeling PET detectors, but their step-by-step nature makes large-scale optical systems computationally prohibitive. Generative models are an attractive alternative, and we previously demonstrated conditional generative adversarial networks for single crystals and crystal arrays, in which physics knowledge enters the model through architectural extensions and a physics-informed loss. In this work, we propose a fundamentally different way of incorporating physics. Instead of acting on the loss, prior physical information is injected directly into the conditioning pathway of the generative model. A lightweight, geometry-agnostic light transport estimator (GALTE) propagates photons through the segmented crystal geometry under a simplified physical model and produces a compact descriptor of the expected sensor-level light response. This descriptor replaces the bare emission coordinates as conditioning input, leaving the generative model to capture only the residual stochastic structure that the analytical prior cannot resolve. We instantiate the framework with a conditional GAN whose architecture and trainable parameter count matches the published OptiGAN baseline, so that observed differences can be attributed to the conditioning rather than to model capacity. On a 3×3 BGO array, the GALTE-conditioned model preserves the distributional fidelity of the OptiGAN baseline across the full crystal array while reproducing physically interpretable structures of the joint position–momentum distribution more closely, particularly in regimes of high reflection multiplicity. Most notably, the GALTE-conditioned model reaches its best performance after roughly one third of the training epochs required by the OptiGAN baseline. Physics-informed conditioning thus opens a path toward generative models that combine physical priors with machine learning.

Abstract

Time-of-flight positron emission tomography (PET) timing calibration is especially challenging for light-sharing detectors, where biases arise from interaction-dependent light transport, silicon photomultiplier response, and effects of readout electronics. Conventional skew and walk corrections are robust but leave high-dimensional event signatures underused. We introduce an in-system approach that learns detector-wise corrections from coincidence data without requiring pair-specific models. A shared network conditioned by detector embeddings estimates single-interaction offsets from singles observables, yielding a structured model with linearly scaling detector-specific parameters. We evaluate the method on two prototype PET systems: a 24-module semi-monolithic detector scanner and a 28-module multi-layer segmented-crystal detector scanner. For both, data were acquired at 450 randomized 22Na point-source positions and split by source position for fivefold cross-validation. We studied a residual multilayer perceptron, a geometry-aware transformer, and a lightweight edge multilayer perceptron. Across scanners and energy windows, learned models improved coincidence timing resolution over time-skew and time-walk correction. The residual multilayer perceptron achieved the best timing resolutions, reaching 431ps and 382ps for scanners 1 and 2. Transformer and edge variants performed similarly with smaller models and generalized to detector pairs excluded during training. Feature ablation identified light spread as the dominant auxiliary cue, with local position and energy adding further improvements. These results establish detector-wise learning as a practical route to scalable timing calibration for PET systems with complex light-sharing detector topologies.

Abstract

Monte Carlo simulation (MCS) is essential for PET and SPECT system modeling and optimization. However, existing CPU-based toolkits such as GATE and Geant4 remain too slow for iterative system design. We present a GPU-accelerated SPECT simulation module within the United-Imaging Healthcare (UIH) Molecular-imaging Innovation Platform (uMIP), extending its validated PET capability to single-photon imaging with native support for both multi-pinhole and parallel-hole collimator geometries. The key innovation is a closed-form analytic ray-aperture intersection algorithm for multi-pinhole collimators, which eliminates iterative geometric traversal and enables massively parallel GPU execution. Its speed is approximately 2~3 orders of magnitude faster than a CPU-based simulation toolkits. Quantitative validation against Geant4 demonstrates excellent agreement for both multi-pinhole and parallel-hole configurations: point-source projection profiles across six field-of-view positions yield NRMSE <0.80% and Pearson correlation >90%, while transmitted-photon spatial distributions and energy spectra show consistent results at both 140 keV and 440 keV. The proposed framework provides a practical, high-fidelity simulation platform for rapid SPECT collimator design and imaging protocol optimization.

Abstract

Accurate Monte Carlo simulation of Positron Emission Mammography (PEM) scanners is essential for optimizing detector design, reconstruction algorithms, and dosimetry in dedicated breast imaging. This work presents a GATE 9.4 simulation of the Naviscan PEM scanner validated against clinical sensitivity measurements at Hamad Medical Corporation, Doha, Qatar. Scanner geometry was implemented from published specifications, with detector efficiency derived from LYSO crystal attenuation via the Beer-Lambert law (approximately 70% for 13 mm crystals, mu = 0.86 cm-1 at 511 keV). A cylindrical aluminum sleeve phantom in five configurations (zero to four 2.5 mm sleeves) was simulated using a 0.58 ns coincidence timing window and a 350-650 keV energy window, matching the hospital protocol. True coincidences were counted following clinical convention, including scattered events within the energy window. The base-case configuration achieved 96.8% agreement with hospital measurements. Mean agreement across all five configurations was 70.3%, with the correct exponential attenuation trend reproduced across increasing sleeve thicknesses. Systematic underestimation at higher attenuation is consistent with known GATE aluminum cross-section limitations and the absence of scatter recovery corrections. All parameters were derived from first principles without empirical tuning. Preliminary results demonstrate a rigorous, reproducible simulation framework for PEM, with complete validation to be presented at the conference.

Abstract

This paper presents an optimization framework for Halbach magnet arrays targeting PET-compatible portable low-field MRI. Positioning an MRI system within existing PET rooms offers a practical alternative to PET-in-MRI inserts, as it simplifies compliance with radioisotope regulations. However, achieving high field uniformity in a 200-magnet Halbach array requires optimizing discrete rotation angles within an exponentially large search space. To address this, we developed a genetic algorithm (GA) utilizing adaptive mutation and dynamic elitism. To prevent premature convergence caused by rapid diversity loss in high-dimensional spaces, we investigated five population seeding strategies, including a novel Segmented Stochastic Progression (SSP) method using stratified complexity sampling. Performance was evaluated via fitness convergence and Hamming distance using progressive mutation step-size ranges (sigma = 6 to 1 and 10 to 1). Our results demonstrated that heuristic seeding maintained population diversity 1.3 to 2.6 times more effectively than random initialization. Specifically, the SSP method provided superior global search stability, yielding highly ordered magnetic flux streamlines across a 10 cm target region within a 20 cm diameter ring. Furthermore, mean time to solution (MTTS) remained consistent (0.94–1.00 relative to baseline), indicating the search quality was improved without increasing the computational overhead. This approach offers a robust, transferable methodology for complex electromagnetic design optimizations.

Abstract

This paper presents a GATE Monte Carlo model and its preliminary validation of the uMI Panorama 35™ PET system, which combines a 35 cm axial field-of-view with sub-200 ps time-of-flight performance. The model was validated for coincidence time resolution (CTR) and against NEMA NU 2-2018 measurements of sensitivity and noise equivalent count rate. The coincidence time resolution of 197 ps obtained with the model is comparable to 194 ps measured experimentally, and a system sensitivity of 21.0 versus 23.5 kcps/MBq. It also reproduced count-rate performance, with a peak noise equivalent count rate of 615 kcps at 43.7 kBq/mL, compared with 687 kcps at 43.2 kBq/mL experimentally. These results show the preliminary work on its validation and focuses mainly on the main counting performance characteristics of the scanner. Future work will extend validation to additional image tests, including spatial resolution and image quality.

Abstract

PET image quality depends on accurate identification of the gamma-ray interaction point. RETIMAGER proposes laser-engraved (LE) pixelated LYSO blocks, where inter-crystal transparency can be tuned through engraving density. This controlled light sharing can preserve crystal identification while providing XY position information. Previous GOSS optical simulations combined with artificial intelligence (AI) predicted the transparency range expected to optimize XY resolution for 1.25 mm pitch hexagonal crystals. Here, we extend this simulation-guided strategy to experimental data by analysing four LE matrices with different engraving densities and effective transparencies. For each event, the seven largest SiPM signals were used to characterize the light-sharing pattern. A χ² classifier was developed from reference patterns associated with the two main configurations imposed by the hexagonal geometry: crystals above one SiPM and crystals close to a vertex shared by three SiPMs. GOSS simulations with different transparency values were processed with the same procedure and used to select a representative configuration and define the ROI criterion for extracting the χ² reference distributions. This criterion was transferred to the experimental matrix, separating the expected 1-SiPM-like and 3-SiPM-like populations. These results show that simulation-guided light-sharing analysis supports classification of the experimental response and future application of AI-based XY identification to LE matrices, aiming to determine the transparency that maximizes detector performance and enables intracrystal positioning.

Abstract

Dark-field X-ray imaging (DFXI) measures ultra-small-angle scattering from sub-resolution density fluctuations and provides contrast complementary to attenuation mammography. We present a wave-optics simulation of a compact single-grid spatial-harmonic imaging (SHI) DFXI system for breast microcalcification typing, focusing on type-II hydroxyapatite (HAp) and type-I calcium oxalate dihydrate (COD). Instead of a generic monodisperse-sphere model, the simulation uses a calibrated γ(ξ)-based microstructure interface: HAp is represented as a Gaussian-random-field medium, whereas COD is represented as a compact-crystal scatterer with weak surface-dominated dark-field response. For a 17.5 keV source, 20 μm focal spot, 47 μm grid period, and 2.0 m source-to-detector distance, the operating autocorrelation length is ξG=249 nm. Simulations with simplified Bakic-style Voronoi backgrounds over BI-RADS A--D show that large HAp clusters remain detectable under anatomy-limited conditions and that the dark-field-to-absorption ratio, r=ΔIdf/Δ(1-T), separates HAp-like and COD-like clusters more robustly than either channel alone. These results support compact tabletop SHI-DFXI for specimen-level microcalcification characterization and surgical-margin assessment.

Abstract

Total Body Positron Emission Tomography (TBPET) scanners offer a substantial sensitivity increase compared to traditional scanners, leading to lower doses or acquisition times. In parallel, interest in modelling such tomographs through Monte Carlo simulations has raised. In this work, a faster framework to simulate TBPET scans is presented by modelling the Siemens Biograph Vision Quadra scanner. SimPET platform is used for simulation, which combines the use of SimSET and STIR. To assess its performance, spatial resolution, sensitivity and image quality NEMA tests were run. Spatial resolution shows reasonable agreement, with an average deviation of 0.6 mm. Although our model overestimates the experimental sensitivity, as expected for a simple cylinder model, a calibration factor is proposed. Concerning image quality, recovery coefficients align with experimental data, while background variability is higher due to reduced maximum ring difference. To sum up, our approach represents a faster alternative (projected x40 speed-up), which could make feasible the large-scale databases simulation when fully adjusted and optimized.

Abstract

Large-animal preclinical studies are essential for translating cardiovascular imaging technologies toward clinical applications; however, dedicated SPECT systems optimized for large-animal cardiac imaging remain limited. In this study, we propose a full-ring large-animal cardiac SPECT system composed of six detector heads and investigate the impact of multi-pinhole multiplexing on system performance. Two collimator configurations are compared: a low-multiplexing reference design with 6 pinholes per plate and a high-multiplexing system with 36 pinholes per plate. The 36-pinhole configuration is generated using a Halton sequence, serving as a specific high-multiplexing representative rather than an optimized arrangement. System performance is evaluated using sensitivity maps and multiple digital phantoms, including a Defrise disk phantom for axial sampling assessment, a Derenzo phantom for spatial resolution evaluation, and a hot-sphere phantom for contrast analysis. All studies employ analytic projection and the Maximum Likelihood Expectation Maximization reconstruction. Results demonstrate that the high-multiplexing system achieves a higher peak sensitivity (2.04 × 10⁻³ vs. 3.87 × 10⁻⁴), improved hot‑rod resolution (3.5 mm clearly resolved vs. 4.5 mm marginally resolved), and a smaller detectable hot‑sphere diameter (4 mm vs. 5 mm). In conclusion, high-multiplexing multi-pinhole design may benefit spatial resolution and detectability in stationary SPECT systems, although projection-overlap-induced artifacts must be carefully managed.

Abstract

Scintillator length of 20 mm has been considered a standard in positron emission tomography (PET); however, recent advances in time-of-flight (TOF) performance necessitate a re-evaluation of this assumption. While shorter scintillators improve coincidence timing resolution (CTR) at the expense of detection efficiency, enhanced TOF information may compensate for this loss in terms of image quality. In this work, we experimentally investigated the effects of crystal thickness on CTR and performed simulations of a clinical PET scanner to evaluate image quality. We used fast-LGSO crystals with a cross section of 4 × 4 mm2 and varying lengths, each coupled to a silicon photomultiplier (SiPM; Hamamatsu Photonics) that had an effective area of 4.0 × 4.0 mm². The timing and energy signals of the SiPM were extracted using a custom-made high frequency readout circuit and digitized with a CAMAC digitizer. The measured CTRs of 10 mm, 15 mm, and 20 mm fast-LGSO crystals were 121 ps, 129 ps and 134 ps, respectively. Based on the experimental results, we investigated the effects of crystal thickness on the image quality of a clinical PET scanner (ring diameter= 675 mm, axial coverage= 158 mm) by Geant4 simulation. Image quality was evaluated using an IQ phantom based on the NEMA NU 2 standard, and metrics including contrast recovery and background variability. Among the three crystal thicknesses, the 20 mm thickness provided the highest contrast while keeping the background variability. At this stage, we have concluded that longer scintillators appear to provide better image quality. We are still investigating the performance of even longer lengths such as 25 mm and 30 mm, and their performance will be discussed at the conference.

Abstract

Monte Carlo (MC) methods are powerful tools for simulating physical processing such as estimating in positron emission tomography (PET) scattered and true coincidences in PET, with GATE considered the gold standard. However, the high computational costs can limit its clinical applicability. SimSET offers a substantially faster alternative while retaining the key photon-interaction physics required for accurate PET simulation. In this work, we present a systematic concordance study between SimSET and GATE for the GE Omni Legend PET scanner geometry. The two simulators are compared across four metrics: sensitivity sinograms for the true coincidences, 2D energy spectra, trues-only reconstructed images, and scatter fraction, using a uniform cylinder and a realistic [18F]PE2I brain phantom as test cases. Strong agreement was observed across all metrics, with scatter fractions differing by less than 0.03 percentage points (25.65% vs. 25.63%). SimSET was found to be approximately ~59x times faster than GATE for almost equivalent simulation conditions. These results demonstrate that SimSET reproduces GATE outputs with high fidelity on this state-of-the-art scanner geometry, establishing it as an efficient and accurate platform for MC-based scatter estimation in clinical PET imaging workflows.

Abstract

The purpose of this study is to develop an accurate event positioning algorithm for a compact scintillation camera to enhance its intrinsic resolution and useful field-of-view (FOV). The compact camera consists of a 50×50×7mm³ continuous NaI(Tl) crystal directly coupled to a Hamamatsu H12700 position-sensitive photomultiplier tube (PSPMT) with an 8×8 multi-anode readout array. Two kinds of reference libraries of detector responses for event positioning were built by analytical model of scintillation light distribution in a monolithic crystal and GENT4-based GATE Monte Carlo (MC) simulation, respectively. Each entry of the discrete grids-based library contains a detector response distribution corresponding to a known interaction position in crystal volume.  Then a hybrid event positioning framework combining least-square estimation (LSE) and k-nearest-neighbour (k-NN) was developed and experimentally validated using Tc-99m flood sources. Preliminary test results show that the proposed hybrid event positioning method with MC simulation-based reference library exhibits good positioning linearity and edge-event positioning accuracy to enhance detector’s spatial resolution and useful FOV as compared to the conventional centroid positioning method. Experimental results also demonstrate that the hybrid algorithm provides robust and computationally efficient event positioning. In conclusion, the approach combing hybrid algorithm with MC simulation-based reference library offers a practical and accurate solution for next-generation monolithic gamma-camera systems requiring high throughput and high intrinsic resolution. Further experimental evaluations are continued under investigations.

Abstract

Positron range (PR) degrades spatial resolution in positron emission tomography (PET), with its efect being radionuclide- and medium-dependent. Current accurate PR simulators are computationally intensive and cannot be used in clinical applications. This work presents the Hybrid Monte Carlo (HMC) PR simulator, a GPU-accelerated condensed-history code in CUDA Fortran for 3D annihilation imaging in heterogeneous voxelized geometries derived from CT images and μ-maps. Pre-computed interaction tables covering 24 reference-tissue materials and support for arbitrary β⁺ emitters up to 3.55 MeV end-point energies (⁶⁸Ga, ⁸²Rb, etc.) allow both GPU and CPU implementations. Maximum speeds reach 4.77·10⁵ (Intel Core i9-14900KF) and 3.32·10⁸ (NVIDIA GeForece RTX 5080) histories/s, surpassing prior CPU-only PR simulators by up to 4 orders of magnitude. Close agreement with PenEasy 2024 is demonstrated: point-source simulations reach FWHM residuals are below 0.1 mm at 10⁷ histories, and voxelwise R²=0.982 and PSNR≈30.8 dB are achieved in a realistic myocardium phantom while decreasing the simulation time from 6 hours to 1 second (using HMC GPU). The resulting annihilation maps enable PET resolution modeling and patient-specific  PR correction within clinically practical time frames.

Abstract

Multi-pinhole collimation in single-photon emission computed tomography (SPECT) trades aperture penetration against sensitivity and spatial resolution. Closed-form analytic sensitivities with penetration are known for the circular pinhole and the rectangular pinhole (recthole), but not for the recently introduced minkhole: a Minkowski-sum aperture lofting a circular waist into a rectangular exit. We close that gap with three knife-edge closed forms&mdash;an Anger three-term penetration expansion, an exact on-axis Stieltjes form in elementary moment integrals, and a Stieltjes-corrected hybrid&mdash;anchored to a two-dimensional level-set integral as the point-source-exact reference. The three-term expansion shares the universal penetration-order scaling of the circular-pinhole and recthole results; the minkhole&rsquo;s first-order coefficient carries a disk-rectangle Steiner cross-term that vanishes at the recthole limit and is structurally absent from the cylindrical circular pinhole. The on-axis Stieltjes route recovers the Metzler&ndash;Moore recthole closed form term-by-term as the waist radius vanishes. Validation against a three-dimensional ray-tracing reference on a 12-panel benchmark grid gives worst relative errors of 1.01% (Anger), 0.69% (hybrid), and 0.39% (level-set exact). The hybrid enables sub-millisecond sensitivity evaluation for minkhole-based SPECT system-matrix construction.

Abstract

Integration of Compton cameras (CC) with PET can be advantageous for multitracer imaging and to improve the resolution for simultaneous β⁺-γ emitters. This technology is showing potential to become a viable option. The combination of PET and SPECT faces challenges due to collimation in SPECT and poor performance at high photon energies. CCs overcome this limitation and are more suitable for high-energy and multiple gamma emitting radiotracers. The IRIS group of IFIC (Valencia) has shown via simulations that combined CC and PET imaging of a Derenzo-like phantom with ⁸⁹Zr, ¹²⁴I, and both ¹⁸F and ⁹⁵Tc is feasible, demonstrating the ability of Compton-PET to image multiple radionuclides. In continuation of this work, a performance evaluation of an experimental Compton-PET prototype is ongoing. An identical geometry is also studied through GATE simulations. The geometry consists of four identical detectors, of which two are combined as a CC and the other two are placed as PET detectors facing each other at either side of the geometrical center of the system. Image reconstruction is performed using a custom algorithm developed by the IRIS group. We present reconstructed images of a ²²Na source array obtained with the simulated Compton-PET system. Successful imaging results pave the way for subsequent laboratory testing of the Compton-PET system, enabling simultaneous β⁺-γ and multitracer radionuclide imaging.

Abstract

Proton Computed Tomography (pCT) is an emerging imaging modality designed to reduce uncertainties in proton therapy planning by directly mapping the Relative Stopping Power (RSP) of tissues. This study provides a comprehensive performance evaluation of the INFN pCT prototype, focusing on experimental dosimetric characterization and the determination of noise scaling laws. Utilizing a silicon tracker and a YAG:Ce calorimeter, measurements were conducted at the APSS Trento center with a 211.2 MeV proton beam. Dosimetry performed with TLD-100 thermoluminescent dosimeters revealed an absorbed dose of 1.4–1.5 mGy per10⁸ trigger events, resulting in a total dose of 11.6 mGy for a high-quality 9-rotation scan. The investigation into image noise demonstrated that the system operates in a purely statistical regime, with the standard deviation of RSP following the inverse square root of the product of dose and voxel volume. No significant additive noise floor was detected. Analysis identified a 1 mm³ voxel volume as an optimal benchmark, constrained by Multiple Coulomb Scattering physical limits. At this resolution, measured noise levels ranged from 0.3% to 0.7%, significantly outperforming the 1% clinical requirement. These results validate the system’s ability to provide high-accuracy RSP maps at low patient doses.

Abstract

In PET/CT scanners the x-ray source is turned off during the PET acquisition for several reasons, including preventing saturation of the PET detectors due to pulse pile-up from lower-energy scattered x-rays. We conducted simulation and experimental studies to evaluate if there are operating conditions where simultaneous PET and CT imaging is feasible.

Simulation studies were used to estimate the flux and energy distribution of scattered x-rays for a dedicated breast PET/CT scanner using a tungsten x-ray tube operated at 65 kVp. The x-ray tube spectrum was used for tracking of x-rays scattered at roughly 90 deg into a PET detector. We also evaluated using metal filters in front of the PET detector to preferentially absorb the lower energy scattered x-rays. We then conducted experimental studies of 511 keV photon detection by a LYSO scintillator in the presence of scattered x-rays. The x rays were generated from a mammography scanner operated at 49 kVp. The scattering object was a water phantom in the mammography field of view. The PET detector was located at 90 deg to the x-ray beam direction. The outputs were evaluated in terms of timing and energy spectra.

The simulation studies indicated that for the CT and PET scanner geometry, the scattered x-ray flux was overall reduced by 90 %. The addition of 1 mm of Ag or Cu in front of the PET detector reduced the flux of scattered x-rays incident on the PET detector by 99.5% or 96.5%, while reducing the flux of 511 keV photons by only 9% or 16%. The experimental studies demonstrated the expected dominance of the scattered x-rays on the PET detector output. The use of 1 mm of Cu filtration restored the photopeak signal, except for the predicted 16% loss. Both simulations and experimental studies indicate that simultaneous PET and x-ray imaging may be possible under certain conditions of x-ray tube technique, scanner geometry, and pre-detector PET filtration.

Abstract

As innovative imaging tools, both X-ray dark-field and X-ray fluorescence imaging show significant potential for medical applications. However, each modality has technical trade-offs that can limit its effectiveness when used in isolation. While molecular imaging plays a crucial role in modern diagnostics by enabling early detection and personalized therapy, existing in vivo methods often suffer from low sensitivity, limited resolution, or high ionizing radiation exposure.

Abstract

In this paper we will introduce a system for two-dimensional imaging of positron emission. It has been developed for the investigation of plant roots grown in rhizotrones. A rhizotron is a narrow, transparent flower pot that forces the roots to grow in a virtually two-dimensional manner. For this reason, a 3D reconstruction, as done in a PET scanner, is not necessary, and for a 2D projection two opposing planar detector heads are sufficient. The heads cover an area of 134x134 mm² and each contains 4096 LYSO scintillation crystals of 1.85x1.85x10 mm³ each. Digital Silicon Photomultipliers (Philips DPC) detect the scintillation events which are stored as list mode data on disk and are already used to instantly visualize an online image of the activity distribution. We will present first images and characterize the system based on evaluation measurements.

Abstract

Zebrafish play an important role in biomedical research as a key vertebrate model for disease studies and drug screening; however, PET imaging has not yet been translated to this animal model. To fill this gap, the PET prototype MERMAID has been specifically conceived for imaging adult zebrafish and other small aquatic animals. Initial studies have proven the feasibility of the concept, as well as the need for anatomical reference information. For the latter, we are developing a cone beam CT system to be integrated into MERMAID. Challenges are the low soft tissue contrast in zebrafish and its small thickness, the attenuation of the water-filled imaging chamber, and potentially high radiation doses for the target animal. Although synchrotron radiation and contrast agents have both proved successful in providing high resolution and contrast zebrafish CT images, neither is suitable for our case, nor phase-contrast CT imaging techniques. The current CT setup consists of a cesium iodide flat panel detector and a handheld x-ray tube with a tungsten filament. For image reconstruction simultaneous algebraic reconstruction technique (SART) adapted from the reconstruction toolkit (RTK) is used. This setup allows a resolution smaller than 0.4 mm. We have successfully imaged several characterization phantoms as well as adult zebrafish ex-vivo, allowing visualization of the skeleton and small soft tissue changes in both air and water. Further work will include image registration, soft-tissue contrast enhancement, and integration with the PET component.

Abstract

This study presents an initial evaluation of a dual-panel PET prototype with TOF-DOI capabilities for high-resolution local human organs imaging. The system consists of two flat panel detectors, offering an open geometry suitable for small anatomical regions. Each panel has a 212 × 212 mm² active area and consists of dual-layer, multi-resolution LYSO detectors. An interleaved multiplexing circuit was implemented to reduce the readout channels by a factor of four, with signals processed by TOFPET2 ASIC electronics.

At the detector level, clear flood images were obtained with well identified crystals. The average energy resolutions were 14.60% and 14.09% for the top and bottom layers. The bottom layer achieved an average DOI resolution of 3.37 mm. The best coincidence timing resolution was 280 ps for bottom–bottom events after excluding inter-crystal scatter events. With a panel distance of 30 cm, the spatial resolutions at the center of the FOV were 1.15, 1.53, and 3.63 mm FWHM in the three orthogonal directions. The peak sensitivity was 4.29% at the FOV center and 2.14% at a 5 cm axial offset.
For phantom imaging, 1.6-mm rods were clearly resolved at the FOV center and remained discernible at a 5 cm off-center when four DOI bins were used in reconstruction. These initial results indicate that the dual-panel TOF-DOI PET prototype is promising for high-resolution imaging of localized human anatomy and small lesions detection. Its open and flexible geometry may also support dose monitoring in proton therapy. Further work will focus on correction methods, a more comprehensive performance evaluation, and practical imaging demonstrations.

Abstract

This work presents a Monte Carlo-based performance evaluation of a preclinical PET system based on a four-layer depth-of-interaction (DOI) detector. The detector with optimized layer thickness design comprises two LYSO arrays and two BGO arrays. The first and third layers adopt 60×30 LYSO arrays with pixel dimensions of 0.78×0.78×4 mm3 and 0.78×0.78×6 mm3, respectively, the second and fourth layers employ 40×20 BGO arrays with pixel dimensions of 1.21×1.21×4 mm3 and 1.21×1.21×10 mm3, respectively. System spatial resolution and sensitivity were characterized via Monte Carlo simulations following the NEMA NU 4 protocol. The single-ring system achieves the peak sensitivity of 16.6% and an average spatial resolution of 0.69 mm across the field of view (FOV). The complete dual-ring complete system could further improve the sensitivity to >25%. Compared with the two-layer detector system, the proposed four-layer DOI design exhibits significantly improved sensitivity while maintaining high and uniform spatial resolution. Experimental validation using the four-layer detector with total size of 51.6×25.8×25.5 mm3 demonstrated CNN classification accuracies of 84% for the LYSO layers and 76% for the BGO layers. Flood images using 45° oblique collimated incidence showed distinct event clustering patterns between upper- and lower-layers Future work will focus on developing the single-ring validation prototype and further improving the DOI layer discrimination accuracy.

Abstract

We report preliminary experimental results on the 3D imaging point spread function response performance of the Ultra-High Performance Brain PET (UHB-PET) system employing semi-monolithic LYSO slabs, machine-learning–based event positioning, and list-mode 3D point spread function (PSF) reconstruction. Detector calibration was previously performed using a slit-collimated 22-Na point source translated across monolithic and depth dimensions. A previous multilayer perceptron (MLP) estimated monolithic and depth positions of interaction, while axial positioning was derived via a previous center-of-gravity algorithm.

List-mode coincidence events were registered encoding time, energy, time-of-flight, module pair and within-module 3D coordinates, discretized into virtual detector pixels (50 transaxial × 50 axial × 20 depth) and mapped to global cylindrical coordinates after depth-of-interaction correction. A list-mode MLEM algorithm with an optimized Gaussian PSF model kernel width was employed. The system 3D imaging PSF was evaluated from 3-min acquisitions by placing a 0.25 mm 22-Na point source across 25 axial positions spanning the full axial FOV along the transaxial FOV center.

3D PET images free of visual artifacts were obtained. A uniform 3D PSF with mean relative absolute deviation from FOV center of 5.83%, 5.87% and 7.32% for both FWHM and FWTM was observed in radial, tangential and axial PSF components, respectively, across all axial locations demonstrating an axially homogeneous 3D imaging PSF for UHB-PET.

Ongoing work includes implementation of classification-based MLP positioning with improved detector resolution, incorporation of experimentally derived spatially variant PSF kernels, and validation with an extended set of phantoms to further optimize 3D imaging PSF response and its uniformity in all directions across the entire FOV.

Abstract

Our PET/MR breast insert currently under development is designed to be compatible with a whole-body 3T PET/MRI system. The proposed insert comprises a ring geometry based on dual layer LSO detector blocks coupled via FFC cables to SiPM arrays. Such integration in the confined and RF-shielded PET/MR environment introduces challenges related to thermal stability, which are addressed using data-driven temperature monitoring and correction. In an isothermal setup, the photopeak remains stable (898.7 ± 0.5). Under a temperature gradient, a drift from 901.1 to 914.1 (~ 1.1 to 8.5 keV) within 30 min is observed, resulting in a shifted and broadened energy spectrum (photopeak 907.8, FWHM 76.9). After scaling energy values via the derived ratio of the peak shift to the reference, the photopeak is restored to 899.1 with reduced FWHM (75.7). In parallel, combined operation of insert and whole-body PET systems requires alignment of independent timestamp domains to enable mixed-coincidence reconstruction. Two detector blocks were operated with independent local clocks, a synchronized reference was established and desynchronized by introducing a delay. A data-driven algorithm estimates this offset via a sliding-window coincidence search (±100 ns), where the offset is determined by the maximum coincidence counts within the timing window. After correcting for the delay, time difference spectra within a narrow ±3 ns time window could be restored while preserving a 430 ps coincidence resolution time. These results demonstrate that data-driven correction and synchronization enable stable PET/MR operation and potentially enable mixed-coincidence imaging across independent systems. Future work will focus on scaling to multi-ring systems, enhancing in-situ temperature sensing during MR operation, and evaluating mixed-coincidence imaging between independent PET systems enabled by the proposed synchronization approach.

Abstract

Conventional brain PET imaging of small rodents necessitates anesthesia to prevent motion artefacts. However, anesthesia may alter brain function and preclude the simultaneous assessment of brain function and animal behavior. The Motion-Adaptive Molecular and Behavioral Observation (MAMBO)-PET system aims to overcome these limitations by enabling brain imaging in awake, freely moving mice. The system consists of four detector panels. Two panels have been assembled, while the remaining two are in the final stages of assembly. Here, we report the initial performance of the two-panel MAMBO-PET system. The mean detector panel energy resolution ranged from 14.9% to 15.0%. The FWHM of the radial and tangential PSF profiles at the center of the FOV were 0.9 mm and 5.7 mm, respectively. Future work includes the NEMA characterization of the full 4-panel system.

Abstract

Our group is developing Dynamic Cardiac SPECT (DC-SPECT) featuring laser-processed CsI: Tl scintillators with converging pixels to avoid depth-of-interaction blurring. This work reports, for the first time, the initial imaging performance of the DC-SPECT scanner featuring laser-processed CsI: Tl detector modules with converging pixels. A 50×50×8.05 mm³ CsI: Tl crystal was laserprocessed into a 25×25 converging-pixel array. The crystal was coupled to an 8×8 multi-pixel photon counter array , read out by a custom 64-channel analog front-end board and a 12-bit waveform digitizer. Three decoding algorithms were evaluated: MLE-Bicubic, MLE-Nearest, and a deep neural network (NN) decoder. A ⁵⁷Co point source and a ⁹⁹ᵐTc three-line source phantom were imaged using list-mode MLEM reconstruction. For the point source, FWHM values of 6.24, 6.58, and 5.67 mm were obtained for MLE-Bicubic, MLE-Nearest, and NN at 20 iterations, with NN yielding the sharpest response. At 50 iterations, FWHM values improved to 4.63, 4.98, and 5.02 mm, with MLE-Bicubic achieving the sharpest response. For the three-line source phantom, sources located near the center of the field of view yielded mean FWHM values as low as 3.55 mm (NN) and 4.04 mm (MLE-Bicubic). Sources near the field-of-view edge exhibited broader responses, consistent with the known degradation of detector sensitivity toward the detector periphery. The converging-pixel architecture effectively eliminates depth-of-interaction ambiguity without additional hardware, further supporting the practical viability of this detector design. Future work will focus on optimization of reconstruction parameters, and further phantom and cardiac imaging studies toward clinical deployment of the DC-SPECT system.

Abstract

Accurate geometric calibration is essential for high-resolution 3D PET image reconstruction because detector element mispositioning can introduce spatial misalignment of lines-of-response (LORs) and image blur. This challenge is particularly important in rotating flat-panel PET scanners, where opposing detector panels rotate during acquisition to improve angular coverage and enable fully 3D reconstruction. In this work, we developed and evaluated an image-based calibration method for estimating linear offsets between detector centers and the axis of rotation in the Radialis PET Imager, an organ-targeted rotating flat-panel PET scanner. The calibration was performed using a ²²Na point source acquired with twelve step-and-shoot detector rotations from -90° to +75°. Four point source experiments were performed at different source locations. An iterative algorithm was implemented to estimate linear offsets by minimizing distances between reconstructed point sources from different detector angles and the target image acquired at 0°. The estimated offsets were fed into transformation matrices used for composite image reconstruction of list mode data from all detector angles. Performance was evaluated using spatial resolution measurements from composite images, Hoffman brain phantom scans, and a clinical ¹⁸F-Florbetaben brain scan. Offset calibration improved the average spatial resolution (FWHM) from 3.3 mm to 2.9 mm, with axial FWHM improving from 5.2 mm to 3.0 mm. Hoffman phantom images reconstructed with offset calibration demonstrated improved contrast and visualization of internal structures. The calibration of ¹⁸F-Florbetaben clinical brain images demonstrated improved visualization of anatomical features, including the septum pellucidum, as well as improved contrast and reduced apparent cortical tracer uptake relative to image reconstructed without offset calibration.

Abstract

Compact PET scanners based on highly integrated front-end ASICs require stable thermal conditions to preserve detector gain, timing response, energy resolution, and inter-module uniformity. This is especially relevant in low-footprint PET systems, where high channel density and closed mechanical integration limit conventional forced-air cooling.

This work presents a thermal stabilization strategy designed for a new family of clinical and preclinical PET scanners: BuckyPET, BrainPET and UltraPET. The solution combines adaptive SiPM bias compensation with active liquid cooling of the ASIC electronics using a cascade proportional--integral (PI) controller. The outer loop regulates an exponentially filtered high-percentile aggregate of ASIC temperature deviations from a learned baseline, while the inner loop controls the coolant temperature through the chiller command. Saturation, setpoint slew-rate limiting and anti-windup mechanisms are included to improve robustness under variable conditions.

The proposed strategy maintained the ASIC thermal deviation within the tolerance band for different ambient conditions and high data throughput acquisitions. Preliminary measurements show improvements of up to 15% in energy resolution and up to 8% in coincidence time resolution compared with worst-case thermal drift conditions. The architecture enables compact, silent, and patient-friendly scanner designs, while preserving reproducible PET performance.

Abstract

Correcting counting loss phenomena caused by pulse pile-up and dead time is essential for the medical application of an energy-resolving photon counting detector (ERPCD). At the IEEE 2025 MIC conference, we proposed a counting loss correction algorithm applicable to objects, but it was limited to an effective atomic number (Z_eff) of 7–13. In this study, we propose a novel counting loss correction algorithm that extends the target object to Z_eff = 3–20 to cover biological samples, and evaluate the quantitative performance of the Z_eff images related to the algorithm. The novelty is to estimate the effect of counting loss on objects with Z_eff = 3–20 using experimentally derived counting loss correction functions for polymethyl methacrylate (PMMA, Z_eff = 6.5) and aluminum (Al, Z_eff = 13). Using the correction function for each Z_eff, the measured counting rate can be converted to the expected counting rate under ideal conditions without counting loss. Since the Z_eff of the object is unknown at the beginning of the analysis, corrections for Z_eff = 3–20 are tentatively performed. The corrected counting rate is then applied to the material identification algorithm, which was proposed in a previous study, and Z_eff is derived. Finally, the Z_eff of an object is determined when the entire algorithm works properly. To demonstrate our method, reference samples with known Z_eff (PMMA, Al, bilayer structures of PMMA and Al, and Si powder) and tooth samples were measured using a prototype ERPCD system equipped with a cadmium zinc telluride (CZT) semiconductor detector. When applying the correction to the measurement of reference samples, the accuracy of the Z_eff image was +/-1 under general radiography conditions. For a dental sample, tooth structure and resin can be observed with proper Z_eff. This result demonstrates that counting loss correction is important for the clinical application of ERPCD.

Abstract

Photon-counting CT (PC-CT) enables acquisition of energy-resolved images. Analyzing these images provides potential for quantitative diagnosis based on material composition. This study aims to develop an algorithm that simultaneously generates effective atomic number (Z_eff) and effective mass density (ρ_eff) images from virtual monoenergetic images (VMIs) acquired by clinical PC-CT. To demonstrate the feasibility of our algorithm, phantom experiment and clinical lung datasets were analyzed as representative applications. In our algorithm, Z_eff and ρ_eff values are estimated voxel-wise from two VMIs. The CT values are converted to μ values, and their ratio is converted to the Z_eff value using a reference curve created from a known database of mass attenuation coefficient (μ/ρ). In addition, the tentative mass densities as a function of atomic number are estimated by fitting μ/ρ data of each atomic number to the measured μ values, and the ρ_eff is uniquely determined using the Z_eff information. The algorithm was validated using an in-house low-density phantom and clinical chest CT images from 10 patients. We analyzed Z_eff and ρ_eff values in diseased and normal lung tissue regions. From the phantom study, the measured ρ_eff values were consistent with the reference density of the samples. In clinical analysis, lung diseases exhibited characteristic changes in ρ_eff. A metastatic lung nodule showed a similar Z_eff value compared to normal lung tissue, but its ρ_eff value was clearly higher. During pneumonia follow-up, ρ_eff values decreased by approximately 56% after one month of disease regression, while Z_eff values remained nealy unchanged. Our algorithm generates Z_eff and ρ_eff images from PC-CT data, allowing quantitative interpretation of conventional CT values. These results suggest that ρ_eff provides a useful quantitative indicator for lung lesion assessment.

Abstract

Metal artifacts in computed tomography (CT) degrade image quality and reduce quantitative reliability, particularly in the presence of metallic objects. Conventional metal artifact reduction (MAR) methods, such as sinogram interpolation and normalized correction, rely heavily on accurate metal segmentation and often fail to preserve fine anatomical structures. Although dual-energy CT (DECT) provides additional spectral information for material characterization, its potential has not been fully utilized in prior-guided iterative MAR frameworks. In this study, we propose a prior-guided iterative MAR method for DECT that incorporates virtual monoenergetic image (VMI) priors for improved artifact suppression and structural preservation. Metal regions are first identified in the image domain using residual-based metal segmentation from initial dual-energy reconstructed images, and the resulting metal mask is forward-projected to generate the metal trace in the projection domain. Dual-energy projection data are then corrected in the projection domain using prior-guided interpolation to reduce metal-induced inconsistencies. Material decomposition is then performed to generate VMIs, which provide structurally consistent priors for subsequent image-domain reconstruction refinement. The final images are obtained by iterative prior-guided reconstruction that enforces data consistency while incorporating VMI-based edge guidance and artifact-suppression regularization to suppress residual artifacts and preserve anatomical boundaries. Quantitative evaluations on simulated datasets demonstrate that the proposed method significantly reduces root mean square error and improves structural similarity compared with normalized MAR. These results indicate that the proposed framework effectively suppresses metal artifacts while maintaining structural fidelity and quantitative accuracy.

Abstract

Accurate BMD  measurement is essential for osteoporosis diagnosis and fracture risk assessment. Wide fan-beam  DXA systems provide faster scan acquisition than pencil-beam systems; however, their geometric characteristics can introduce magnification errors when the distance between the X-ray source and the subject changes. These distortions affect bone mineral content (BMC), projected bone area, and ultimately BMD accuracy. The iNSiGHT C510 (OsteoSys Co., Korea) employs a 25° fan beam with a SID distance of 812 mm, substantially wider than the 4.5° narrow fan beam geometry of the PRIMUS, which operates at a 634 mm SID. To compensate for magnification-induced distortion, the iNSiGHT C510 incorporates the geometric correction within the LUCIDA software platform. The correction algorithm was evaluated using a custom lumbar spine phantom elevated from 0 to 200 mm above the tabletop in 10 mm increments, with 10 repeated measurements obtained at each height. Without correction, the iNSiGHT C510 showed BMD deviations ranging from approximately +40% to −40%. After applying the correction algorithm, residual BMD variability was reduced to within ±5%, comparable to the ±4–5% variation observed with the PRIMUS under identical conditions. Correlation analysis between corrected iNSiGHT C510 measurements and corresponding PRIMUS values demonstrated strong agreement, with a coefficient of determination (R²) of 0.934 and a standard error below 2%. These results indicate that the proprietary correction algorithm effectively compensates for magnification errors inherent to the wide fan-beam geometry, enabling the iNSiGHT C510 to achieve clinically equivalent BMD accuracy suitable for routine bone density assessment.

Abstract

Computed tomography (CT) has made significant advances in reducing X-ray exposure while preserving image quality. A novel project to improve CT imaging involves subsituting anti-scatter grids with filtering out scattered photons using time-of-flight (ToF) measurements. A first approach resorting to ToF filtering was developed to eliminate the noise due to scattered photons, thereby improving the contrast-to-noise ratio. Additionally, ToF-CT has the potential to exploit both detected ballistic and scattered photons for reconstructing both 3D images of absorption and scattering using model-based image reconstruction.
In this approach, an X-ray radiation propagation model must be resorted to. 
For this purpose, a transient radiative transport equation is presented herein. It is divided into ballistic and scattering propagation components. The present study focuses on the ballistic component, with the aim of evaluating various numerical methods to achieve better numerical temporal accuracy.

Abstract

The easyPET/CT is a multimodal imaging system that integrates positron emission tomography and computed tomography (CT), using LYSO:Ce scintillator crystals coupled to silicon photomultipliers. CT acquisition is done in photon-counting (PC) mode, enabling energy discrimination and improving contrast-to-noise ratio. The radiation source used is a ²⁴¹Am radioactive source (2.89 MBq), which simplifies system integration. This work presents the first characterization of the integrated CT modality of the easyPET/CT of  through simulation, evaluating CT number calibration, noise, and uniformity. A non-linear system response between pixel intensity and Hounsfield units was observed and successfully corrected using a second-order polynomial fit (R²=0.99). Noise and uniformity assessment revealed peripheral non-uniformity attributed to reconstruction artifacts. Spatial resolution characterization is ongoing. These results establish a foundation for a compact, portable PCCT system with strong potential for multimodal preclinical imaging applications.

Abstract

This work characterizes the LYSO:Ce/SiPM detector array of the easyPET/CT, a compact multimodal preclinical imaging platform. The characterization of the 16×2 array of LYSO:Ce crystals, coupled one-to-one to SiPMs was assessed per crystal in terms of energy resolution (ER), count rate, and sensitivity using a ²⁴¹Am radioactive source (59.5 keV). The mean ER was 33.8 ± 1.8% FWHM at 59.5 keV, competitive with current LYSO:Ce-based photon-counting detector systems reported in the literature. The mean count rate and sensitivity were 62.4 ± 5.3 photons/s and 0.7 ± 0.1 counts•photon^-1, respectively, both displaying a uniform spatial distribution across the array. The overall uniformity of the detector response validates the shared PET/CT hardware architecture as a viable approach for photon-counting CT, confirming that using the PET detector for CT operation preserves detector integrity. Future work will address  pile-up characterization at higher count rates to further define the dynamic range of the system.

Abstract

Attenuation correction (AC) is crucial for artefact-free PET/MR images. Currently, CT-derived PET attenuation maps (AMs) are based on transformation schemes from conventional CT scanners with data from polyenergetic X-ray spectra. It is known that converting from polyenergetic CT image data to monoenergetic 511 keV linear attenuation coefficients (LACs) introduces systematic errors in the LACs employed for AC in PET imaging. To address this issue, we investigated spectral CT to obtain more accurate AMs. Spectral CT offers advantages over conventional CT scanners by providing the possibility of material decomposition. It also allows to reconstruct virtual monochromatic images (VMI) from the inherently multiple polyenergetic data sets of CT.  The spectral information can also be used to obtain virtual non-contrast (VNC). In our approach, we use angiographic spectral CT scans with an administration of iodinated contrast agents and computed VNC images from the obtained dataset. The linear transformation scheme from monochromatic CT images to LACs at 511 keV was obtained from VNC images for different tissue types using an electron density phantom and a photon counting (PC) CT scanner. The use of VNC images effectively simplifies the transformation from HUs to LACs requiring only a single linear function. It further minimizes residual deviations of the LACs for soft-tissues mimicking materials. Deriving the LACs from VNC images instead of standard CT images may lower the systematic errors of CT-based attenuation correction for PET scanners equipped with dual energy CT systems.

Abstract

Metal-induced artifacts in spectral photon-counting computed tomography (SPCCT) degrade image quality and, critically, corrupt material decomposition (MD) accuracy by introducing spectral distortions across energy bins. This paper presents a generalized optimization pipeline applied to a residual dense U-Net, developed through a two-stage process combining manual grid search and Optuna-based automated machine learning (AutoML) to tune the optimizer, learning rate, loss configuration, and dense block depth. The pipeline is evaluated using a dual-phase framework that covers energy analysis and MD performance on hydroxyapatite and iodine density images. RDU-Net+ achieves a mean signal-to-noise ratio (SNR) of 35.34, root mean squared error (RMSE) of 0.0045, and structural similarity index (SSIM) of 0.954, while improving hydroxyapatite classification accuracy to 92.37\% and iodine accuracy to 91.86\%. All improvements over the baseline are statistically significant ($p < 0.001$). The proposed pipeline provides a structured, reproducible optimization framework for deep learning-based metal artifact reduction in SPCCT.

Abstract

X-ray fluorescence computed tomography (XFCT) enables nanoparticle imaging and biological function imaging without radioisotopes. We successfully improved image quality by applying Deep Image Prior (DIP), a self-supervised deep learning technique that does not require training datasets, to XFCT images. However, conventional DIP, which uses randomly generated input images, has problems such as reduced spatial resolution and deviation from the true image distribution. To address these problems, we propose CT-DIP, a modified DIP framework in which the network input is replaced with projection images acquired from simultaneous X-ray transmission computed tomography (XTCT). Monte Carlo simulations using Geant4 were performed to generate both XFCT and XTCT projection images of a mouse brain phantom, with iodine used as a contrast agent to mimic cerebral blood flow. Image reconstruction was carried out using the ordered subsets expectation maximization (OS-EM) algorithm after DIP denoising. Visual evaluation and quantitative evaluation using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) demonstrated that CT-DIP outperformed conventional DIP, Gaussian filtering, and only OS-EM reconstruction. Furthermore, CT-DIP reduced training time from 75.3 minutes to 4.2 minutes, representing an approximately 18 times speedup. These results suggest that CT-DIP has strong practical potential for XFCT imaging in low-count environments, and future work will focus on validation with real measurement data.

Abstract

The performance of fine-pitch cadmium zinc telluride (CZT) photon-counting detectors is strongly governed by experimentally measurable charge transport and leakage current characteristics. We present the fabrication and experimental characterization of a pixelated CZT detector array with a nominal pixel pitch of 441 × 441 μm², targeting high-channel-count architectures for computed tomography (CT). CZT material properties were characterized using a custom-built time-of-flight (TOF) measurement system to extract electron mobility and lifetime under controlled electric field conditions. Dark current behavior was systematically measured as a function of bias voltage and pixel geometry to assess operational limits imposed by material quality. Experimentally derived transport and leakage parameters were used to perform a detector-level trade-off analysis linking CZT material properties, pixel pitch, and bias field. The results establish experimentally grounded constraints and design guidelines for scaling CZT detector arrays toward advanced photon-counting CT systems.

Abstract

Cross-modality synthesis from cone-beam CT (CBCT) to CT is central to online adaptive radiotherapy and accurate dose recalculation. Although planning MRI provides valuable soft-tissue information, daily CBCT captures the patient’s updated anatomy on the treatment couch. Its direct use remains limited by scatter artifacts and imperfect Hounsfield Unit (HU) calibration. This study investigates whether aggregating sequential on-treatment CBCTs improves deep learning-based synthetic CT (sCT) generation compared with single-fraction CBCT or static MRI priors alone.

We developed a conditional GAN with Residual-in-Residual Dense Blocks to generate high-fidelity sCTs. The model was evaluated on 190 patients with brain metastases, split into 152 training cases and 38 independent test cases. Each patient had a planning CT used as ground truth, a planning T1-weighted MRI, and up to three daily CBCT scans. Four input settings were compared: 1 CBCT, 1 CBCT+MRI, 3 CBCTs, and 3 CBCTs+MRI.

Adding temporal and multimodal information improved sCT accuracy. The single-CBCT baseline achieved 67.79 HU MAE and 29.12 dB PSNR, while the full 3 CBCTs+MRI configuration reached the best performance, with 59.81 HU MAE and 30.82 dB PSNR. Importantly, temporal accumulation alone using 3 CBCTs achieved 60.01 HU MAE, outperforming the static multimodal 1 CBCT+MRI setting at 65.98 HU. This suggests that longitudinal CBCT redundancy provides stronger density and attenuation constraints than a single CBCT complemented by a rigid planning MRI prior.

Sequential CBCT aggregation, especially when fused with MRI, improves HU calibration and reduces daily setup artifacts, supporting robust automated sCT generation for online adaptive radiotherapy.

Abstract

Photon counting computed tomography (PC-CT) enables energy-resolved imaging and material-specific analysis through individual X-ray photon detection. In this study, dynamic hepatocyte-specific imaging was performed in healthy and liver tumor mouse using a developed two-dimensional (2D) PC-CT system consisting of a 16 × 64-channel YGAG scintillator array coupled with MPPC array, with a voxel size of 1×1×1 mm³. Gd-EoB was intravenously administered, and time-resolved imaging was conducted before and after contrast injection. Reconstructed CT images in the 50–60 keV energy range and estimated Gd concentration maps were evaluated around the liver and bladder regions. Dynamic enhancement was observed after Gd administration in both healthy and liver tumor mice, with differences in temporal enhancement behavior and estimated Gd concentration tendency between the two models. Although the tumor region was not clearly distinguishable because of its small size and current image quality limitations, temporal changes in CT value and estimated Gd concentration were still observable. MRI comparison demonstrated a positive relationship between MRI signal intensity and PC-CT-derived Gd concentration, although quantitative tendencies differed because of differences in imaging mechanisms and contrast sensitivity between MRI and PC-CT. These results demonstrate the feasibility of scintillator-based PC-CT for dynamic hepatocyte-specific imaging in vivo.

Abstract

Material decomposition is commonly based on the R-value, defined as the ratio of attenuated intensities measured at low-high X-ray energies. However, materials with similar compositions, such as fat and muscle, can exhibit overlapping R-values, limiting the accuracy of material decomposition. This study investigates the feasibility of improving material decomposition using scatter-induced arrival time difference. Scattered photons travel longer paths than primary photons, resulting in delayed arrival times. Photons with arrival time delay exceeding a threshold were distinguished as scattered photons. The scatter ratio (SR), defined as the ratio of scatter counts in low to high energy regions, was used for an additional identifier for material decomposition. DXA system was simulated using GATE with a detector consisting of 240 × 24 LYSO pixels with dimensions of 3 × 3 × 5 mm³, an X-ray source operated at 83 kVp and 0.2 mA with a 0.025 mm Ce filter, and the source-to-detector distance set to 1000 mm. Fat, muscle, and bone phantoms were placed at the center between the source and detector. The arrival time delay threshold was set to 20 ps since the maximum time difference of direct photons within the 5 mm pixel depth is 16 ps under ideal timing resolution of 0 ps. Each material was represented in feature space using the R-value and SR. Material separability was evaluated using a separability index based on the difference between feature means and standard deviations. In the feature space, fat, muscle, and bone formed distinct clusters. SR derived from photon arrival time difference improved the separability index for all material pairs. For fat–muscle, the separability index increased from 13 to 17, achieving a 25% increase. For fat–bone and muscle–bone pairs, separability improved by 84% and 62%, compared with the R-value. These results demonstrate the feasibility of improving the material decomposition in DXA using scatter-induced arrival time difference.

Abstract

Limited-angle tomography and few-view tomography are extremely ill-conditioned problems. A usual way to reconstruct an image in limited-angle tomography or in few-view tomography is via an iterative algorithm that minimizes a two-term objective function. The objective function consists of a data fidelity term and a Bayesian term. A common Bayesian term is the total variation (TV) norm of the image. This paper proposes a different way to reconstruct the image in limited-angle tomography and few-view tomography by iteratively deconvolving the point spread function (PSF). Two PSFs are considered: the PSF of the filtered backprojection (FBP) and the PSF of pure backprojection. The method of using the PSF of pure backprojection gives better results.

Abstract

To address the high computational complexity and low efficiency caused by traversing all field-of-view (FOV) voxels during image reconstruction in full-ring cascade gamma coincidence imaging systems (CGCIS), we propose a location priors algorithm to significantly accelerate image reconstruction. By restricting system matrix calculations, forward projection and back-projection operations to the ROI. This reduces the computational complexity from O(N_x×N_y×N_z) to O(m×n×l), thereby significantly improving reconstruction efficiency. Experimental results demonstrated that, compared to conventional iterative algorithm, the proposed location priors algorithm achieved an approximate 37.05-fold acceleration in reconstruction speed while preserving high imaging quality. Ultimately, this advancement provides robust technical support for real-time, dynamic imaging monitoring utilizing theranostic nuclides.

Abstract

We present a gradient-based SPECT reconstruction framework that treats the physics-based forward model as a frozen differentiable operator and optimizes the 3D activity volume through backpropagation, an approach directly analogous to Google’s DeepDream algorithm. The forward model is implemented as a PyTorch module comprising gantry rotation via trilinear interpolation, photon attenuation correction, and depth-dependent detector blur via per-slice convolution with measured point response function (PRF) kernels. Reconstruction minimizes the Mighell modified chi-square, a Poisson-appropriate cost function, using adaptive gradient descent. We evaluate four first-order optimizers (Adam, AdamW, RMSprop, SGD) on a hot/cold sphere phantom acquired on a Siemens Symbia Intevo with LEHR  collimation and demonstrate that RMSprop achieves the fastest cost reduction. Preliminary reconstructions at 20 iterations recover the phantom geometry with visible hot sphere contrast, validating both the forward model accuracy and the gradient-based reconstruction pipeline.

Abstract

This work presents a modular software integration framework that bridges physics-based PET reconstruction with modern deep learning regularization. While deep learning has shown significant promise in medical imaging, integrating learned priors into established reconstruction pipelines remains a challenge. We address this by developing a direct interface between the DeepInverse library and the Synergistic Image Reconstruction Framework (SIRF), which wraps the Software for Tomographic Image Reconstruction (STIR) engine. This integration enables the direct deployment of learned priors, such as the Deep Residual UNet (DRUNet) and Weakly Convex Ridge Regularizers (WCRR), within accurate physics-based forward models. To demonstrate the framework's flexibility, we integrate and evaluate three reconstruction strategies using learned regularization: PnP-PD3O (WCRR), PnP-SVRG (DRUNet), and PnP-HQS (DRUNet). Performance is validated using a simulated 2D thorax phantom, comparing the proposed framework against Ordered Subsets Maximum A Posteriori One Step Late (OSMAPOSL), corresponding to standard Ordered Subsets Expectation Maximization (OSEM) reconstruction when no prior is applied, and OSMAPOSL with Relative Difference Prior (RDP) baselines. Quantitative results demonstrate improved image fidelity using learned regularization. As a representative example, the PnP-SVRG approach achieved a Peak Signal-to-Noise Ratio (PSNR) of 45.90 dB, compared to 39.52 dB for baseline OSEM, showing improved reconstruction quality using this unified interface. The framework suppresses noise while preserving structural details. By providing an open-source Python-based bridge, this work lowers the barrier for evaluating state-of-the-art inverse-problem techniques in realistic clinical PET imaging scenarios.

Abstract

Lithium-ion batteries (LIBs) are widely used in electric wheelchairs, and their safety is a critical concern due to the risk of thermal runaway caused by internal defects. In practical applications, batteries are operated as multi-cell modules, and periodic inspection is required as their condition degrades over time. X-ray computed tomography (CT) is widely used for non-destructive testing. However, CT inspection of compact multi-cell modules is challenging since overlapping attenuation from neighboring cells can lead to reconstruction artifacts and reduce defect contrast. In this study, CT reconstruction strategies for multi-cell LIB modules were evaluated using a Geant4-based simulation framework. Projection data were generated under the simulated conditions, and various combinations of projection processing and reconstruction methods were compared. Reconstructed image quality and defect detectability were quantitatively evaluated using the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and detectability index (d′). The results show that the combination of unfiltered projection data, OS-SART reconstruction, and Gaussian smoothing achieved the highest mean d′ across the nine cells. Compared with the baseline reconstruction, SNR, CNR and d′ improved by 233%, 119% and 117%, suggesting improved defect-to-background separation. These results indicate that reconstruction strategy significantly affects defect detectability, and optimization of reconstruction methods is necessary. In future work, experimental validation will be conducted on a prototype radiography system.

Abstract

PET reconstruction aims to recover the activity distribution from coincidence measurements, while accurate quantitative imaging relies on comprehensive modeling of the physical processes spanning positron annihilation, photon propagation, and detector response. Existing deep learning approaches commonly rely on end-to-end supervision using iterative reconstructions as reference labels, which limits quantitative accuracy and does not explicitly account for the physical degradation mechanisms governing PET measurements. In this work, we propose Wiener Wavelet Net (WieNet), a degradation aware reconstruction framework that models the geometric time of flight probabilistic (GTP) image as a blurred and bias contaminated observation of the underlying activity distribution. A two-stage learnable Wiener filtering strategy is introduced to address TOF induced spatial blurring through frequency domain deconvolution, while a wavelet attention fusion module performs multi energy guided scatter correction by selectively incorporating reliable low frequency information. A 3D denoiser suppresses residual stochastic noise, and an absolute activity calibration module restores the quantitative intensity scale using patient and acquisition parameters. Experiments on clinical and simulated datasets demonstrate that WieNet provides improved scatter compensation and quantitative accuracy compared with fully corrected OSEM and other baseline methods. These findings suggest that degradation aligned reconstruction integrating blur restoration, scatter correction, and quantitative calibration offers a principled alternative to purely supervised deep learning PET reconstruction.

Abstract

Depth-of-interaction (DOI) information is essential to reduce parallax-related spatial blurring in PET systems using thick scintillation crystals, especially in compact preclinical scanners. We present an image-guided DOI estimation method for a preclinical PET scanner with dual-ended readout, aimed at DOI calibration from both signals. The method follows a superiterative reconstruction strategy. An initial image is reconstructed with a fixed DOI assignment and then used as a spatial prior. Then, from the list-mode file, for each event LOR indicated by a crystal-pair, several DOI candidates are compared for the said crystal pair, generating several possible DOI-LORs. Each candidate LOR is forward-projected onto the previous image, and the value of such projection is used as a compatibility metric. The DOI combination with the highest compatibility is selected to update the DOI-LOR for the given event. To assess consistency, the dual-ended energy ratio was represented as a function of the estimated DOI. The ratio-DOI distributions show a clear depth-dependent trend, indicating that the proposed method provides a DOI response correlated with detector light sharing.

Abstract

In cardiac-gated SPECT studies, the acquired data counts can exhibit significant inter-subject variability. Consequently, the reconstructed images are often processed with post-reconstruction filtering in which the smoothing level needs to be adjusted individually among subjects for optimal noise suppression. In this study, we investigate an operator-interpolation approach to achieve an adaptable level of smoothing with a deep learning (DL) denoising network for processing cardiac-gated images, enabling real-time, patient-specific adjustments without retraining the DL network. We assessed the performance of this approach on a set of 20 clinical gated acquisitions using a DL network pre-trained at two nominal smoothing levels on image data from 862 studies. The results indicate that the proposed approach could achieve nearly identical denoised output images compared to retraining the DL network for a specific target smoothing level, with an average relative mean-squared-error (rMSE) below 0.15% and the structural similarity index measure (SSIM) above 0.982.

Abstract

The application of organ-specific imaging on axially long field-of-view (LAFOV) scanners is currently emerging, where an axially shorter CT acquisition defines the target region of imaging. Nevertheless, the PET system naturally acquires additional measurements, where high-polar-angle data contain contributions from both the CT-defined region and the region extending beyond it. Estimation of correction factors, including attenuation and scatter, requires knowledge of the object in the extended region using traditional approaches.

The modern scanner has TOF resolution of a few centimeters, and the TOF local tomography property can be exploited for this task. High-polar-angle PET data, truncated in the TOF domain, can be used, as they contribute to the targeted region. This paper investigates this approach with the goal of faster reconstruction of the target region of interest. Scatter correction is performed using quantitative knowledge of the object obtained from reconstruction of the subset of data for which correction factors are fully known (residual scatter approach). Attenuation can be obtained by a similar approach or by the ML-ACF method using full or TOF-truncated emission data. Alternatively, the attenuation map can be estimated from the full-extent emission data compressed into a single backprojection image using a fast DL method.

Preliminary investigations on the Siemens Vision Quadra show that the sensitivity of the targeted region can be restored with accurate quantification.

Abstract

Cardiac PET imaging is significantly affected by respiratory and cardiac motion, leading to image blurring and reduced quantification accuracy, particularly in motion-sensitive myocardial regions. Although dual-gating techniques can mitigate motion artifacts, their clinical adoption remains limited due to substantial noise amplification and degraded image quality. In this work, we propose a clinically feasible dual-gating reconstruction framework that integrates deep-learning–based denoising with motion-compensated reconstruction to address this fundamental limitation.

The proposed method was evaluated using both XCAT simulations with known ground truth and clinical datasets from 18 patients undergoing Flyrcado and Rubidium (Rb) cardiac PET imaging. A deep-learning denoising model is applied to enhance gated images and improve the robustness of motion estimation. Motion fields are estimated using a symmetric normalization approach and incorporated into a novel multi-pose reconstruction framework, where motion information is directly embedded into forward and backward projections for efficient motion-aware reconstruction.

Simulation results demonstrate accurate recovery of myocardial defects that are blurred or indistinguishable in standard reconstructions without motion correction. Clinical results show improved image sharpness, enhanced myocardial delineation, and preserved count statistics across both tracers, indicating consistent performance under realistic imaging conditions.

By effectively addressing the noise–motion trade-off while maintaining computational efficiency, the proposed framework enables practical clinical implementation of dual-gated cardiac PET imaging. This approach has the potential to improve diagnostic confidence and quantitative accuracy in routine cardiac PET applications.

Abstract

Ultrasound-modulated optical tomography (UOT) is a hybrid biomedical imaging modality that uses a laser beam and ultrasound (US) to reveal optical properties deep within scattering living tissues. It exploits the acousto-optic effect and is based on the detection of photons whose frequency is shifted by the US frequency. Traditionally, focused US waves are used to directly sample the optical image pixel by pixel. To improve the signal-to-noise ratio of the optical image and reduce the acquisition time, US plane waves can be used, resulting in a tomographic problem. However, the finite numerical aperture of the US probe introduces a limited-angle problem, resulting in a loss of lateral resolution. To mitigate this issue, structuration of the plane waves (SW) has been successfully tested for long acquisition times. The objective of the study is to evaluate the performances of Least Squares with Total Variation (LS-TV) tomographic reconstruction in SW UOT for shorter acquisition times, using simulated data with realistic noise levels.
We show first that reducing the number of configurations of the US plane wave structuration from 1353 initially to 41 maintains reconstruction quality for a fixed total acquisition time. Then, we increase gradually the acquisition frame rate from 0.002 Hz initially to 3 Hz. Analysis with adapted regularization parameter reveals a trade-off between frame rate, noise in the reconstructed image, and contrast recovery coefficient (CRC) for an optical absorber mimicking a 1.35 mm radius tumor. Noise increases with frame rate while CRC remains stable below 0.19 Hz and drops significantly above 1.5 Hz.
This study shows for the simulated setup that LS-TV reconstruction enables SW UOT imaging up to a frame rate of 1.5 Hz. It also shows that noise is the primary limiting factor for high-speed UOT rather than sparsity in undersampled SW data.

Abstract

Sparse-view CT reconstruction can reduce radiation dose and acquisition time, but it leads to a severely ill-posed inverse problem. Recent denoiser-based methods such as SNORE provide effective image priors for this setting, yet remain computationally expensive at full resolution. In this work, we propose a multiresolution extension of Prox-SNORE for sparse-view CT. The method solves a sequence of reconstruction problems from coarse to fine resolution levels, using each intermediate solution to initialize the reconstruction at the next finer level. This strategy improves the conditioning of the optimization in the early stages and accelerates convergence at high resolution. Experiments on 30-view CT show that the proposed approach achieves a better time–quality trade-off than standard and annealed Prox-SNORE, while reaching comparable final reconstruction quality. These results suggest that multiresolution optimization is a simple and effective way to make stochastic denoiser-based reconstruction more practical for accelerated CT reconstruction.

Abstract

Positron range introduces a spatially variant, density-dependent blur in positron emission tomography (PET) images of high-energy emitters such as ⁶⁸Ga and ⁸²Rb, degrading quantification at tissue interfaces. We present a graphics-processing-unit (GPU) implementation of a heterogeneous maximum-likelihood expectation-maximization (MLEM) positron range correction with a joint expectation-maximization/total-variation (EM-TV) update, built on an exact density-weighted forward/adjoint pair derived from a recently published multi-branch range parametrization supporting eight isotopes. The forward operator combines near-field volume integration with far-field Fibonacci-sphere ray traversal of a CT-derived density volume; the adjoint is the matched transpose, and the iterative update conserves total counts exactly. The code corrects a clinical-scale 192x192x335 ⁶⁸Ga-PSMA volume in 2 minutes (1 iteration, 8 subsets, RTX 3090) and produces visibly sharper structures on a ⁶⁸Ga mouse acquisition with a preclinical scanner in 45 seconds. The code is publicly available at https://github.com/jlherraiz/PRC_OSEM.

Abstract

Joint Maximum Likelihood Activity and Attenuation (MLAA) provides a promising CT-less attenuation correction framework but suffers from global scaling ambiguity. Intrinsic Lu-176 background radiation provides patient-specific transmission information to reduce this issue, while introducing a highly nonlinear and ill-conditioned joint likelihood. We propose the Separable Cubic Surrogate with Trust Region (SCS-TR) algorithm, utilizing an analytical Lipschitz constant and a vectorized solver to ensure strictly monotonic convergence and avoid local optima. The reconstruction is initialized with Maximum Likelihood Transmission Reconstruction (MLTR) results. GATE simulations studies demonstrate that SCS-TR effectively recovers clear bone structures and smooth uniform regions without explicit regularization. The proposed method recovered clear bony structures and uniform soft-tissue regions without explicit regularization, and maintained structural fidelity under 20% count statistics.

Abstract

Respiratory, cardiac and patient motion during positron emission tomography imaging create artifacts in reconstructed images and can bias their interpretation. In this study, we propose a data-driven, organ-based method to correct motion. Our framework utilizes deep learning-based organ segmentation on no motion corrected images, followed by motion gating derived from centroid of tracer distribution computed within the segmented regions of interest on histo-images. Motion vector fields between the gated images are estimated using the Lucas-Kanade algorithm. Motion is subsequently corrected for in the corresponding gated sinograms and the motion-corrected image is reconstructed from the motion-free sinograms. We demonstrate the effectiveness of our method by applying it to the liver, enabling accurate estimation and correction of respiratory motion. Qualitative and quantitative results show that our method significantly improves the quality of the reconstructed images compared to no motion correction, and that the organ-specific approach also outperforms motion correction based on motion estimation on the whole abdominal tracer uptake. The average mean and maximum lesion intensities increased by 23.4% and 26.9%, respectively. Combined with better contrast observed on the reconstructed image, these findings suggest that the organ-based approach can recover some true signal that would otherwise be lost due to motion. The proposed framework paves the way for more precise, organ-focused motion correction strategies, enabling the reconstruction of high-quality images tailored to specific anatomical regions.

Abstract

In cascaded gamma photon coincidence imaging, coincidence events offer superior spatial localization accuracy. However, their low cascade branching ratio limits detection efficiency compared to single-photon events, resulting in elevated statistical noise and prolonged acquisition times.

Since simultaneously acquired single-photon events also encode the spatial distribution of the radionuclide, we propose a multi-information joint iterative reconstruction algorithm. This method fuses coincidence and single-photon events to construct a joint cost function.

Given the incomplete spatial sampling of single-photon events, linearly integrating these modalities via the MLEM algorithm causes high-count single-photon data to overwhelm the high-resolution coincidence features. This leads to image distortion and edge blurring.

To overcome this, we introduce a Maximum A Posteriori (MAP) estimation framework. Discarding the conventional image-domain spatial smoothing prior, the algorithm utilizes single-photon information simultaneously for direct image reconstruction and projection-domain regularization. Within this framework, a data weighting factor, α, governs the integration of single-photon data in the likelihood term, while a prior penalty factor, β, modulates the anti-distortion penalty.

For ultra-low dose or rapid scanning protocols, increasing α leverages the high detection efficiency of single photons to bolster effective statistical counts, suppress Poisson noise in count-starved coincidence events, and enhance sensitivity. Concurrently, synergizing with β to counteract image distortion yields substantial improvements in image quality and the contrast-to-noise ratio.

Abstract

Full-ring multi-aperture single-photon emission computed tomography (SPECT) can provide excellent spatial resolution, but geometric misalignments of collimator apertures and detector blocks can substantially degrade image quality. We developed and evaluated a detector-block-level geometric calibration method for full-ring SPECT systems using analytical ray-tracing simulation. The simulated system used the geometry of Alpha-SPECT, consisting of 34 detector modules arranged on a ring with a 40.4-cm radius. Each module contained a 2 (transaxial) × 8 (axial) detector-block array, and each block comprised 40 × 40 CZT crystals with dimensions of 0.5 × 0.5 × 6 mm³. Each module was paired with an 18-mm-thick tungsten collimator containing sixteen 2-mm-diameter pinholes, each focused onto one detector block and separated from the detector surface by 3.1 cm. The proposed calibration estimated source-position offsets and local pinhole and detector-block misplacements by fitting three source-shift parameters and four pose-perturbation parameters for each pinhole–block pair. Simulated projection were generated by ray tracing from a 3 × 3 × 3 grid of 140-keV point sources with 2.5-cm spacing, after applying realistic random perturbations to source positions and system geometry. For each pinhole–block pair, nonlinear least-squares fitting minimized the discrepancy between measured and estimated block projections using independent multistart global optimization followed by two-stage local refinement. Across all detector blocks, the estimated geometric perturbations closely matched the known applied perturbations. Reconstruction of a small Jaszczak phantom using the calibrated geometry showed significantly improved image quality compared with reconstruction using the perturbed geometry. These results demonstrate that detector-block-level calibration can accurately recover local geometric misplacements and support robust high-resolution full-ring multi-aperture SPECT imaging.

Abstract

Image reconstruction in Positron Emission Tomography (PET) is conventionally performed using Maximum Likelihood Expectation Maximization (MLEM) or Ordered Subsets Expectation Maximization (OSEM). While highly successful clinically, these approaches are limited in their ability to incorporate general regularization functions, particularly non-smooth penalties such as Total Variation (TV). In practice, most clinical reconstruction workflows combine OSEM with post-processing Gaussian filtering, which may lead to oversmoothing or residual noise in low-count acquisitions. In this work, we propose a Chambolle–Pock Primal-Dual Hybrid Gradient (CP-PDHG) framework for PET reconstruction compatible with both sinogram and list-mode representations. The proposed formulation introduces a stacked operator representation enabling simultaneous treatment of complete sinogram and list-mode likelihood terms while naturally incorporating convex regularization functions. We derive closed-form dual updates for the Poisson likelihood and implement Huber-regularized reconstruction as a differentiable approximation to TV. The method was evaluated on simulated IQ phantom data with count levels ranging from 10⁶  to 10⁸ detected coincidences and on clinical long axial field-of-view (LAFOV) PET acquisitions acquired on a Siemens Biograph Vision Quadra system. In noiseless simulations, CP-PDHG achieved convergence rates comparable to MLEM, while in low-count scenarios it demonstrated substantially improved noise behavior. Regularized CP-PDHG consistently outperformed MLEM with Gaussian post-filtering in both simulated and clinical low-count reconstructions, showing improved lesion conspicuity and reduced noise propagation.

Abstract

Positron emission tomography (PET) attenuation correction (AC) is essential for achieving quantitative accuracy and is conventionally performed using computed tomography (CT). However, CT-based AC introduces challenges such as image misregistration and additional radiation exposure. Although deep learning-based CT-free AC methods have shown promising performance, concerns remain regarding the reliability and physical consistency of purely data-driven attenuation estimates. In this study, we investigate a hybrid CT-free AC framework that combines deep learning-based attenuation estimation with maximum likelihood reconstruction of attenuation and activity using time-of-flight information (TOF-MLAA). As an initial step toward this integration, we focus on evaluating the TOF-MLAA performance for attenuation-map reconstruction using emission data alone. Monte Carlo simulation was conducted using a realistic digital brain phantom and a brain-dedicated PET scanner geometry. The results demonstrate that TOF-MLAA can reconstruct attenuation maps with anatomically plausible structures and physically meaningful spatial distributions, despite residual noise and quantitative inaccuracies. As a next step, tracer-agnostic deep learning-derived attenuation estimates will be incorporated into the TOF-MLAA framework. This integration is expected to improve the stability and quantitative performance of attenuation reconstruction while enhancing the reliability and physical consistency of deep learning-generated attenuation maps.

Abstract

Dead-time correction is important for quantitative PET imaging; however, typical methods are approximate, as global dead-time correction is applied using system-level count-rate information. Even with more refined methods, such as block-wise correction, limitations remain. On the other hand, one-to-one coupled detectors, which provide independent signal readout for each crystal, may exhibit crystal-dependent variations in count rate and dead-time behavior. In this study, therefore, we developed a crystal-wise dead-time correction method for the brain-dedicated PET system VRAIN, in which each scintillator crystal is coupled to an individual SiPM. Singles count rates for each crystal were calculated from singles event data acquired during measurement of an ¹⁸F-filled phantom over a wide range of radioactivity levels and the rates were used to estimate dead time. Dead-time factors representing sensitivity reduction due to count loss were modeled using both non-paralyzable and paralyzable models. The factors were estimated for arbitrary acquisition time frames and incorporated into the forward projection model of a list-mode OSEM reconstruction algorithm. To evaluate the proposed method, pool phantom measurements were performed and compared with a detector-wise correction method. The proposed method showed improved linearity with radioactivity compared with the detector-wise approach. Comparable performance was observed between the two dead-time models, while the non-paralyzable model showed slightly better linearity. These results demonstrate the feasibility of crystal-wise dead-time correction for improving quantitative accuracy in brain PET systems with one-to-one coupled detectors.

Abstract

The easyPET.3D system is a preclinical positron emission tomography (PET) scanner featuring a non-conventional acquisition geometry based on the dual-rotation of two opposing detector modules, combining a fan rotation for angular sampling and an axial rotation for volumetric coverage. Its performance was previously evaluated following the NEMA NU 4-2008 standard, demonstrating consistent spatial resolution across the field-of-view. However, such assessment relies on filtered back projection, an analytical reconstruction method that may not accurately represent the performance of systems with unconventional geometries, where its application can be challenging, thereby limiting the direct applicability of the standard protocol.  In this work, the easyPET.3D spatial resolution was further investigated using a micro-Derenzo phantom and iterative reconstruction (LM-MRP), enabling a more realistic and application-oriented evaluation. A quantitative metric based on the smallest resolvable rod diameter was introduced to formalize the commonly adopted smallest resolvable rod criterion and provide a simple and reproducible indicator under practical imaging conditions. The analysis was performed for the three predefined acquisition modes of the system (fast, medium, slow), previously established based on optimized operating parameters. These modes differ in axial rotation step angle and acquisition time, reflecting distinct sampling conditions while preserving the same hardware configuration. This work explores a more application-oriented approach for assessing spatial resolution in preclinical PET systems with non-conventional geometries.

Abstract

The NEMA NU 2 protocol provides the standard for positron emission tomography (PET) spatial resolution evaluation and benchmarking, yet its reproducibility and repeatability have not been rigorously evaluated for modern PET systems. In this study, we investigated the variability of NEMA spatial resolution measurements using the uEXPLORER total-body PET scanner under controlled experimental conditions. A total of 195 point source measurements were acquired to determine an estimate of NEMA spatial resolution uncertainty along with the effects of capillary size, activity concentration, and acquired counts. Spatial resolution was computed following the NEMA standard using in-house code and a vendor provided tool for filtered back projection reconstruction. Despite controlled experimental conditions, statistically significant differences were observed in 16 of 86 comparisons indicating limited reproducibility across experiments. From our measurements we determined that a single NEMA spatial resolution measurement exhibits an uncertainty of approximately 0.2 mm (≈5%), improving to 0.1 mm (≈3%) with repeated acquisitions. There was no observed impact to spatial resolution from capillary size (0.3–1.0 mm) or total counts acquired, while a potential dependence on activity concentration was identified but requires further validation. These findings highlight a discrepancy between reported precision in the literature and practical measurement variability. We recommend incorporating repeated measurements and reporting uncertainty to ensure meaningful benchmarking. Revisions to the NEMA protocol may be necessary to better reflect the performance of modern PET systems

Abstract

In PET, the large dynamic range of activity in the scanner field-of-view (FOV) represents a significant challenge in accurately detecting and quantifying all incoming photons and coincidences. Block pileup and deadtime, well-known effects that increase non-linearly with the activity exposure, are unavoidable for all types of detector hardware. Thus, the observed global sensitivity at the block-level, as well as the detection efficiency at the crystal-pixel level, can be strongly impacted. The adequate characterization and correction of these sources of bias is imperative to obtain artifact-free quantitative images from the reconstruction process.

Abstract

In this work we apply the fully automated motion correction algorithm using open source Software for Tomographic Image Reconstruction (STIR) with automated motion detection for [¹⁸F]-Flutemetamol ([¹⁸F]-FMM) tracer for β-amyloid accumulation in the brain cortex region. The visual results show significant improvement in contrast in the Pons and Cerebellum region with decrease blurring at the edges around the cortex region of the brain. The quantitative results shows statistically significant agreement of centiloid, a metric for early diagnosis Alzheimers' disease, between the motion corrected [¹⁸F]FMM tracer and the noisy longest frame without detectable motion using the single sided Wilcoxon test at 5% compared to no motion correction.

Abstract

Spontaneous patient motion degrades image quality and compromises quantitative accuracy in brain PET imaging. In this work, we present and evaluate three data-driven approaches for rotational motion detection using list-mode PET data: a point-cloud-based principal component analysis (PCA) method operating on LOR midpoint coordinates, a PCA method applied to backprojected images, and a convolutional neural network (CNN)-based approach. Controlled measurements were performed using an anthropomorphic head phantom on a custom holder allowing isolated single-plane rotations, with ground truth determined via MRI at each position. The PCA-based methods achieved an angular precision of 1-3 degrees, with the backprojection-based implementation additionally reducing signal cross-talk between projection planes. A full comparison of all three methods will be presented at the conference.

Abstract

Positron Emission Tomography (PET) detector systems are continuously being developed and optimized to enhance sensitivity and spatial resolution. Standards, such as those defined by the National Electrical Manufacturers Association (NEMA), are essential for reliable performance assessment. However, little has changed between the first (NU 2-1994) and the latest (NU 2-2024) NEMA standard on spatial resolution assessment of PET scanners. Also, the small animal PET scanner standard (NU 4) has not been updated since 2008. Reconstructing point sources with outdated filtered back-projection remains mandatory. Although Maximum Likelihood Expectation Maximization (MLEM) algorithms are favored for achieving higher resolution and lower noise, their use for spatial resolution assessment is discouraged, since reconstructed point sources can become arbitrarily small with many iterations [1]. In this work, we propose a protocol for rigorous and accurate estimation of the spatial resolution of MLEM-reconstructed images. Our method relies on the Derenzo phantom [2], yet, any hot-rod phantom may be used. Spatial resolution is assessed in two steps. First, Modulation Transfer Function (MTF) values are extracted at spatial frequencies defined by the rod sizes. Then, a parametric MTF model is fitted to derive the resolution from the associated Point Spread Function (PSF). We show that this method converges, removing the dependence on the number of iterations. We further demonstrate that it yields spatial resolution values consistent with theoretical expectations. Therefore, we recommend updating the upcoming NEMA NU 2 and NU 4 standards to reflect established iterative reconstruction practices and enable a robust assessment of spatial resolution across PET systems.

Abstract

In this work, random and scatter corrections were developed for N-PET, a dedicated brain PET system, using the correction framework supported by CASToR. Each detector module contains dual-layer LYSO scintillator arrays with four DOI encoding bins. Random coincidences were estimated using the delayed window method, while scatter coincidences were estimated using a GATE Monte Carlo simulation model. Both random and scatter estimates were incorporated as additive terms in the iterative reconstruction model. To improve the statistical stability of sparse LOR-level estimates, each 2×2 neighboring crystal array was grouped as one block. For each block-pair, the random and scatter counts from all corresponding crystal-level LORs were summed and evenly redistributed to the LORs within the same block-pair.

Random correction was evaluated experimentally with an image quality phantom containing three hot spheres with diameters of 13 mm, 22 mm, and 37 mm. The results showed that random correction reduced background noise outside the phantom, and normalization and attenuation correction further improved image uniformity and hot sphere visibility. Scatter correction was preliminarily evaluated using GATE simulation. The proposed method increased the CRC of the 13 mm hot sphere from 0.49 to 0.60, which was comparable to the scatter-free reconstruction result of 0.59. These results show that the proposed list-mode-based random and scatter correction method improves the reconstructed image quality of the N-PET system. Future work will validate the scatter correction with experimental data and further optimize the accuracy of both corrections.

Abstract

Accurate patient attenuation maps (μ-maps) are essential for precise PET quantification. In this study, we compared four methods for generating brain µ-maps in TOF PET/MRI scans and evaluated their impact on reconstructed PET images. We analyzed MRI segmentation, MLAA-based attenuation correction, and two deep learning (DL) approaches: convolutional neural networks (CNN) and denoising diffusion probabilistic models (DM). Two clinical studies utilizing [¹⁸F]-FDG and [¹⁸F]FE-PE2I tracers on a Siemens BIOGRAPH One PET/MRI (Siemens Healthineers) were examined. The MR-based technique occasionally misclassified skull bone and paranasal sinuses as soft tissue. Both DL networks produced reasonable µ-maps for [¹⁸F]-FDG but struggled with the [¹⁸F]FE-PE2I data, which was absent from the training set. While MLAA-based µ-maps assisted in recovering missed bone and sinuses, they were more susceptible to noise. For [¹⁸F]-FDG, µ-map variations did not cause major PET quantification differences, though MLAA showed a slight overcorrection. Conversely, for [¹⁸F]FE-PE2I, CNN-based attenuation led to an activity underestimation of up to 10% in the striatum. This study highlights the importance of µ-map selection, and future work will utilize PET/CT for a systematic "ground truth" comparison.

Abstract

Quantitative PET requires attenuation correction (AC) to compensate for the loss of annihilation photons due to absorption and scattering within the patient's body [1]. The current clinical gold standard is CT, but it increases radiation dose, hardware costs, and registration complexity [2]. These limitations motivate the exploration of CT-less AC approaches in PET imaging [3]. One such method is Detector Scattered Attenuation Correction (DSAC), which uses detector-scattered photons to estimate the attenuation map from the PET data itself [4]. In DSAC, detector-scattered (DS) coincidences are used. These coincidences define lines of response (LORs) in which a gamma ray undergoes Compton scattering in one detector module and then deposits its remaining energy in another detector module after passing through the patient. As these photons pass through the phantom, attenuation information is encoded through photoelectric absorption and scattering. In this comparative study, we use GATE simulations [5] with two different materials, plastic and LYSO, to estimate attenuation maps via DSAC. The results show that plastic scintillator based JPET Scanner [6,7] yield a significantly higher DS-coincidence fraction (47%) than LYSO (1.2%), representing a 6.3-fold increase in useful statistics despite the lower raw registration efficiency of plastic. Single Slice Rebinning (SSRB) reconstructions show lower absolute error, better uniformity, and reduced cupping for plastic, while the phantom remains visible for LYSO despite its lower DS statistics. Our study concludes that DSAC is feasible for PET imaging with both scintillator types, with plastic offering stronger DS statistics and improved preliminary quantitative metrics

Abstract

Accurate randoms correction in PET is typically achieved via delayed-window (DW) or singles-product (SP) methods, both requiring dedicated hardware unavailable on many research PET systems. We present Non-imaging Pairs Random Estimation (NPR), a calibration-free estimator that uses detector pairs whose lines of response (LORs) lie outside the reconstruction field of view (FOV) as a random-coincidence baseline. Per-pair randoms are estimated in three closed-form steps from prompt list mode data alone. Validated against GATE Monte Carlo ground truth on data acquired across 38 simulations spanning random fractions 9–99%, NPR achieves per-pair MAE = 0.45%, matching DW (0.50%) and SP (0.35%) for spatially distributed sources. Under a 5×-singles-asymmetry hot-sphere stress test, NPR retains 1.96% MAE (matching DW, which is also robust) while SP fails by +230%/-33%. A 20-pair FPGA configuration supporting NPR45 (neighbor pairs only, MAE = 0.50%) is demonstrated on R-PET with an ¹⁸F capillary phantom.

Abstract

In iterative image restoration, the total variation (TV) norm is the most common regulator to control noise and enforce sharp edges in the images. This paper proposes an exponential version of the TV norm. This exponential version of the TV norm behaves similar to the L_p norm for 0 < p < 1 and it promotes sharp image edges better than the TV norm. This exponential TV is easy to implement. We use computer simulations to compare the effectiveness of preserving sharp edges by the TV regulator and by the exponential version. The image restoration task is image denoising. The exponential regulator is demonstrated as more effective in obtaining piecewise-constant images than the popular TV regulator.

Abstract

Walk-through positron emission tomography (WT-PET) is an emerging upright imaging geometry based on two vertically opposing flat detector panels. Although this design may improve patient comfort and acquisition efficiency, respiratory motion remains a major source of image degradation. This study evaluates respiratory motion correction strategies for WT-PET using XCAT anthropomorphic simulations. A 30s acquisition was simulated in a WT-PET geometry and reconstructed using an in-house MLEM-based algorithm. Data-driven amplitude- and phase-based respiratory gating were followed by non-rigid motion correction using Advanced Normalization Tools (ANTs) and Fast Algorithms for Motion Correction (FALCON). Overlapping-frame reconstruction using 0.9375s windows shifted every 0.3125s was also evaluated to assess the effect of temporal smoothing. Performance was quantified against a motion-free reference using normalized root mean square error (NRMSE), diaphragm edge width, and line profile analysis. ANTs phase-based 4-gate correction achieved the best performance, reducing NRMSE from 64% to 6% and restoring the diaphragm edge width to the motion-free reference level (25.10 mm vs. 25.79 mm). FALCON phase-based 4-gate correction reduced NRMSE to 12% but showed limited recovery of diaphragm sharpness, with an edge width of 28.40 mm compared with 28.62 mm for no motion correction. Overlapping-frame ANTs correction reduced NRMSE to 15% and yielded an intermediate edge width of 27.13 mm, although line profiles indicated residual local displacement and smoothing. These findings show that data-driven respiratory gating combined with deformable registration can substantially improve image quality in upright WT-PET.

Abstract

Modern Positron Emission Tomography (PET) scanners are increasingly complex, incorporating a large number of modular detectors that require meticulous synchronization. Precise calibration of detector parameters, such as energy response, depth-of-interaction (DOI), and time-alignment, is essential to maintain image quality. This work presents a robust multi-parametric correction pipeline applied to two LYSO based, DOI capable PET scanners: a preclinical  (12 cm transaxial FOV) device and a extended Field of View clinical geometry (23 cm transaxial FOV) scanner, targeting four primary sources of systematic degradation: crystal-by-crystal energy centroid offsets, individual crystal gain non-uniformities, time alignment, and DOI-dependent energy and time variations. Calibration data were acquired from reference point sources and specialized phantoms to ensure statistical reliability. The timing alignment pipeline, which iterates recursively over crystal- and detector-level centroids, improved the Coincidence Time Resolution (CTR) by 50% through the effective mitigation of electronic offsets and time-walk effects. Furthermore, DOI-dependent energy corrections linearized the crystal response, improving energy resolution from 25.0% to 12.6% FWHM. These results demonstrate that a modular multi-parametric calibration framework significantly optimizes timing and energy performance across heterogeneous PET architectures, providing a scalable solution for high-sensitivity molecular imaging.

Abstract

We investigated whether a positron emission
tomography-derived latent biomarker can serve as a compact
prognostic signal after longitudinal modeling. The study used
3,458 scans from 1,624 participants in the Alzheimer’s Disease
Neuroimaging Initiative. A Bayesian mixed-effects model was
fitted to the biomarker trajectory and evaluated in two practical
settings: forecasting future visits for previously observed
subjects and estimating short-term dementia risk from predicted
biomarker values. For subjects already observed during training,
future visits were predicted with a mean root mean squared error
of 0.298 and a mean absolute error of 0.250. For entirely new
subjects, a change-from-baseline formulation reduced prediction
error to a mean root mean squared error of 0.217 and a mean
absolute error of 0.171 after an initial scan. A training-only
logistic calibration applied to held-out predictions reached an
area under the receiver operating characteristic curve of 0.826
for future dementia risk estimation. These results suggest that
the latent biomarker can move beyond representation learning
and support short-term longitudinal forecasting and clinically
interpretable risk stratification.

Abstract

Data augmentation techniques applied in state-of-the-art image reconstruction with deep learning in the positron emission tomography (PET) modality are performed in the reconstructed image domain and subsequently forward-projected to the sinogram domain. In this work, we present a method for applying affine rotation and scaling transformations to sinogram images obtained from a series of simulations in GATE. The proposed algorithms were validated on 100 2D sinogram slices from Nema-like and Jaczack-like phantoms. The preliminary results from this study indicate that the proposed methods are visually indistinguishable from those obtained by applying the transformation in the image domain and then forward projecting to the sinogram domain. The quantitative evaluation was performed using the metrics Zero Mean Normalized Cross-Correlation (ZNCC), which yielded mean values over 0.95, and the Gradient Normalized Cross-Correlation (GNCC), which yielded mean values over 0.66, for both the rotation and scaling transformations. These preliminary results show that the implemented algorithms could be considered the first of their kind, designed to tackle data augmentation directly in the sinogram domain.

Abstract

Multiplexed positron emission tomography (mPET) is an emerging method that enables two or more tracer distributions to be separated from data acquired in one PET scan.  The technique requires high detection efficiency for triple coincidences resulting from a prompt-gamma and two annihilation photons emitted from a positron+prompt gamma emitter. In this study, we focus on multi-hit events occurring within the same coincidence timing window inside the same detector, which are typically discarded as multi-photon hits or inter-crystal scatter (ICS). We propose a recovery method for both annihilation and prompt-gamma events from these multi-hit events, thereby improving the detection sensitivity for triple coincidences. We first generated a new charge‑to‑digital converter (QDC)–based energy window derived from single‑hit 511 keV events for each crystal. This crystal‑specific window was then applied to each hit in multi‑hit events, and then each hit was classified as either an ICS, annihilation photon, or prompt‑gamma event. Evaluation with a ²²Na point source placed at the center of the scanner showed increases of 3.5% in annihilation photon singles, 10.8% in prompt‑gamma singles, 1.1% in double coincidences, and 9.1% in triple coincidences. These results indicate that the proposed method recovers useful events and enhances triple‑coincidence sensitivity for mPET.

Abstract

We present the Hyperion DAQ (data acquisition platform), a versatile MRI-compatible infrastructure for PET detector data acquisition. It combines firmware on a sensor-tile FPGA and a mainboard FPGA with a software stack that splits between a Data Acquisition and Processing System (DAPS) host — receiving and processing the ASIC data over a 10 Gig Ethernet downstream link — and a separate control computer that coordinates the platform components. The firmware handles ASIC configuration, capture and per-packet annotation; the higher-level calibration routines run in software. We have integrated the PETsys TOFPET2 ASIC as the first front-end digitization chip and built an end-to-end calibration suite on top of the platform: breakdown determination via bias sweeps, signal-level and threshold calibration via discriminator sweeps, a charge-to-digital converter (QDC) linearization sampled by on-tile test pulses and stored as a per-bin lookup table, and a time-to-digital converter (TDC) calibration driven by the natural random trigger stream from a 511 keV source. All four routines run in-system on the assembled scanner. We further report on the ASIC’s test-pulse insensitivity window, showing that its position depends on the end phase of the test pulse rather than its start phase — a finding relevant to anyone calibrating the TDC with a test-pulse-based code-density method. On a Visor Max detector pair (8×8 6 mm Broadcom SiPMs read out by a single TOFPET2 ASIC, 64 LYSO crystals per tile), three ²²Na sources, 5 V over-voltage and 17°C cooling, we measure an energy resolution of 10.7 % FWHM at 511 keV and a coincidence time resolution of 248 ps FWHM. An end-to-end evaluation of the data-rate handling of the firmware and software acquisition chain will be presented at the conference. The platform is being prepared for the next generation of Hyperion-based PET/MRI scanners.

Abstract

Accurate lesion-level dosimetry is essential for evaluating therapeutic response in radiopharmaceutical therapy; however, its reliability in SPECT/CT imaging is limited by partial volume effects, which can significantly underestimate activity concentration in small or heterogeneous lesions. These limitations become particularly important when comparing radionuclides with different emission characteristics, such as the widely used ¹⁷⁷Lu and the emerging ¹⁶¹Tb, which emits additional low-energy electrons that may enhance microscopic dose deposition. In this study, we performed a lesion-level dosimetric analysis of ¹⁶¹Tb-FAPI and ¹⁷⁷Lu-FAPI using a Monte Carlo simulation framework based on patient-specific SPECT/CT data. Two iterative partial volume correction methods, Richardson–Lucy and reblurred Van Cittert, were applied to assess their impact on absorbed dose estimation. Lesions were segmented on SPECT images, and voxel-wise dose maps were generated using GATE simulations, followed by lesion-level dose aggregation. Relative dose differences were calculated by comparing corrected and non-corrected images. Results show that partial volume correction significantly affects lesion dosimetry for ¹⁷⁷Lu-FAPI, with dose increases reaching approximately 30.6% in bone metastases and 23.2% in liver lesions, while changes in corresponding normal tissues remained minimal. In contrast, ¹⁶¹Tb-FAPI exhibited more stable behavior, with lesion-level variations generally limited to below 10% and moderate reductions observed in lung lesions (≈ −15%). Both correction methods produced consistent trends, with slightly lower magnitude changes for the reblurred Van Cittert approach. Overall, these findings indicate that lesion-level dose estimation for ¹⁷⁷Lu is highly sensitive to image correction, whereas ¹⁶¹Tb provides more robust dosimetric estimates.

Abstract

Deep learning-based attenuation correction (DL-AC) offers potential advantages over conventional CT-based attenuation correction (CT-AC) in PET imaging, including avoidance of additional radiation exposure and reduced susceptibility to PET–CT misalignment. However, in DL-AC application scenarios, when CT-derived attenuation maps (CT-μ) are unavailable as reference images, quality control (QC) of deep learning-derived attenuation maps (DL-μ) often relies on manual inspection. To address this limitation, we propose a patch-level 3D convolutional neural network (3D CNN) framework for automated anomaly detection and localization in DL-μ without requiring CT during inference. During training, CT-μ was used to assist patch-level annotation of DL-μ, and artifact-free CT-μ patches without observer-identified abnormalities were incorporated as normal samples to help the model learn normal attenuation patterns. During inference, only DL-μ was used, and sliding-window prediction was used to aggregate patch-level anomaly probabilities into a voxel-wise anomaly probability map. On the validation set (51 cases), the framework yielded patch-level performance with an AUC of 0.9984, an F1-score of 0.9802, a sensitivity of 97.06%, and a specificity of 98.86%. These results suggest that the proposed framework can provide spatial cues for DL-μ, potentially supporting efficient evaluation and reducing reliance on manual QC inspection.

Abstract

Artificial intelligence approaches require training on large amounts of data to achieve acceptable levels of performance as well as generalizability. However, there is a lack of available annotated datasets for nuclear medicine imaging modalities as they are a relatively low-volume clinical modality, especially when compared to other anatomical imaging modalities, such as CT and MRI. We aim to leverage a deep generative modeling approach based on latent diffusion models to generate synthetic PET/CT images to address this issue. ¹⁸F-FDG PET/CT and ¹⁸F-DCFPyL PET/CT scans of 668 patients with various types of cancer, including lung cancer, melanoma, lymphoma, and prostate cancer were used to train a latent diffusion model. The data were randomly partitioned into training and test sets using a 90%/10% split. A hierarchical, latent diffusion modeling approach was trained in multiple stages. A patch volume autoencoder was trained to learn a discrete latent representation of local image volume patches using a vector quantized-variational autoencoder. A diffusion model was then trained to generate image volumes sampled from random noise in the latent space. Image quality metrics, including peak signal to noise ratio (PSNR) and mean absolute error (MAE), were evaluated across the training and test sets. On the training set, the approach yielded PSNR values of 40.85 ± 5.50 dB and 36.77 ± 5.30 dB and MAE values of 0.015 ± 0.01 and 0.021 ± 0.011 between the generated and original PET and CT image volumes, respectively. On the test set, the approach yielded PSNR values of 40.00 ± 6.19 dB and 35.48 ± 5.75 dB and MAE values of 0.017 ± 0.012 and 0.030 ± 0.075 between the generated and original PET and CT scans, respectively. The approach achieved relatively high PSNR and low MAE values, respectively, indicating that the approach generated high fidelity PET and CT images. The present study demonstrated a proof of concept for generating realistic PET/CT volumetric scans.

Abstract

Quantum-inspired denoising is employed for clinical PET/CT, PET/MR, and SPECT imaging by treating images analogous to disordered materials and analyzing the eigenfunctions of an associated Hamiltonian. Quantum localization is quantified by the participation ratio, enabling filtering of non‑localized, high‑participation modes that dominate the mid‑frequency range and are associated with diffusive noise. By suppressing these modes and reconstructing from the remaining localized modes, noise is reduced without training data or manual regularization. The framework is applied to 46 multi‑vendor, multi‑tracer patient datasets from Massachusetts General Brigham to characterize scanner‑dependent noise structure and its impact on image appearance. Across the cohort, quantum‑inspired filtering yields consistent decreases in blind image quality scores (PIQE and BRISQUE) together with increases in entropy and Laplacian variance, indicating lower perceived noise and enhanced structural detail in the denoised clinical images.

Abstract

Patient motion remains a common challenge in dynamic cardiac positron emission tomography (PET), as it degrades image quality and complicates reliable evaluation of motion correction methods. Realistic motion simulation requires motion models that capture the variability observed in clinical data, but this remains difficult when simulations rely on only a limited set of motion patterns. In this work, we propose a data driven framework for motion model selection and simulation in dynamic cardiac PET. We characterize motion trajectories using physically interpretable features and analyze them using clustering and trajectory similarity based on dynamic time warping to identify a compact set of representative motion models. These models are integrated with patient-specific motion scaling for frame-wise simulation of motion-corrupted PET data. The proposed framework produces stable normalized error levels across 1308 patient–model evaluations while using only a small number of representative motion models. The results suggest that the selected models effectively capture dominant motion behaviors observed in the dataset and enable realistic PET motion simulation.

Abstract

Patient head motion during positron emission tomography (PET) brain imaging degrades system spatial resolution and quantitative accuracy; even sub-millimeter displacements can blur fine structures in highresolution systems. We present the validation of an OptiTrack PrimeX 41 fourcamera near-infrared stereocamera system for use with a novel γβ⁺ brain-imaging prototype targeting ≤0.200mm spatial accuracy and ≤250ms temporal synchronization with list-mode data acquisition. System calibration achieved a mean residual of 0.075mm, below the manufacturer’s 0.1 mm specification. Tracking accuracy was characterized over 360,000 Euclidean-distance measurements using a Capintec brain phantom displaced by Zaber precision stages (repeatability ≤0.050mm) as ground truth, yielding RMSE of 0.018–0.038mm and maximum outliers ≤0.135mm across three orthogonal directions that were at least 5× better than the 0.200mm target. Three temporal synchronization strategies were evaluated; a network-based UDP broadcast produced a correctable 17.2±2.4ms transmission delay, yielding a combined synchronization uncertainty of 57ms versus the 250ms requirement. Performance was maintained in a clinically realistic configuration with a mock 40-cm imaging bore and supine patient positioning.

Abstract

Tracer kinetic modeling using the two-tissue (2T) compartment model remains challenging in dynamic SPECT imaging, particularly under realistic noise conditions. In this study, we propose a noise-robust Bayesian parameter estimation method that incorporates physiological prior information for the 2T model in dynamic SPECT imaging. To validate the proposed method, we conducted both simulation study and dynamic ^99mTc-4BOH clinical study. The proposed method was compared with the conventional nonlinear least-squares (NLS) method. In the simulation study, the proposed method yielded lower relative bias in the estimated kinetic parameters (K₁, k₂, k₃, and k₄) than the NLS method under both low and high noise conditions. In the clinical ^99mTc-4BOH study, K₁ estimated using the proposed method demonstrated significantly stronger agreement and lower relative bias with respect to head-to-head ¹³N-NH₃ PET myocardial blood flow (MBF), which was used as the reference standard, than the NLS method. In conclusion, the proposed method substantially improved the accuracy and precision of parameter estimation for the 2T model compared with the NLS method.

Abstract

We propose a kinetic-consistent conditional flow matching framework for dynamic [11C raclopride PET SRTM parametric imaging. Conventional indirect SRTM fitting is vulnerable to frame-wise count noise, causing unstable voxel-wise R1 , k2 , and BPND maps. Using 105 scans from 35 subjects, count-reduced datasets were generated at 50%, 25%, and 10% of the original counts. The model refines reduced-count indirect SRTM estimates toward full-count SRTM pseudo-reference maps using kinetic consistency, count/noise conditioning, and physiological constraints, reducing BPND  deviation, high-percentile errors, and kinetic-parameter disagreement compared with indirect SRTM and denoising and regularization-based baselines

Abstract

In proton therapy, range uncertainties can result in insufficient dose delivery to the tumor or unintended dose delivery to organs at risk (OARs). To address this, several range verification methods have been investigated, and prompt gamma imaging is considered a promising method due to the strong correlation between the dose profile and prompt gamma distribution at the distal dose fall-off location. Based on prompt gamma imaging, a treatment monitoring system was developed to measure the in vivo beam range and detect range shifts during the treatment. The developed system consists of a 72-slit parallel collimator, 144 CsI(Tl) scintillation detectors, a dedicated field-programmable gate array (FPGA)-based data acquisition (DAQ) system, and a dedicated program with a graphical user interface (GUI). The DAQ system calculates pulse height and timing of each signal from 144 detectors. The dedicated program reconstructs 1D prompt gamma distribution using pulse data and determines the beam range of each spot using a dedicated range estimation algorithm [1]. The system's performance was evaluated using a clinical proton beam prescribed to deliver 0.5 Gy to a spherical clinical target volume (CTV) with 50 mm diameter in a homogeneous PMMA phantom. To evaluate range shift detection performance, range shifts were induced at both the global range shift (1–10 mm) and local range shift (5 and 10 mm) conditions. At the same time, the computation time for range deviation analysis was evaluated. The system detected global range shifts with an accuracy of 1 mm and identified the presence and location of local range shifts exceeding 5 mm. Furthermore, the time required for range deviation analysis per layer was found much shorter than energy layer switching time. These results demonstrate the system can verify the in vivo beam range during treatment in clinical practice.

Abstract

Proton range verification (PRV) in proton therapy is an outstanding clinical need. Prompt-gammas (PG) emitted along the proton track constitute a strong signal to determine the Bragg Peak position. However, the high energy of PGs (1-7 MeV), the high particle flux and the presence of significant background radiation make prompt-gamma imaging (PGI) challenging.
We report on the development of pure Cherenkov pixelated detectors for PRV. The detector consisted of an 8×8 pixelated PbF2 crystal array coupled to an 8×8 S14161-3050HS-08 SiPM array from Hamamatsu, read out by a row-column signal processing board. Two analytical event reconstruction algorithms were evaluated: rise-to-the-power with k=2 (RTP2) and k=3 (RTP3).
The detector was exposed to both a sealed 228Th source and a 67.5 MeV proton beam at a current of 100 pA. Floodmaps were obtained using both RTP2 and RTP3 methods to resolve the 8×8 matrix of the crystal under laboratory source irradiation. Similar results were obtained under proton beam irradiation, showing the robustness of this detector concept.
In this contribution we demonstrate for the first time the ability of a pixelated pure Cherenkov detector to provide spatial resolution information in clinically relevant proton beam conditions. The prompt emission of Cherenkov light and exclusive sensitivity to highly energetic PGs offer a very high count rate capability, effective background rejection, and capacity to multiplex signals, enabling to use less channels and low power readout electronics. Future work will focus on quantitative floodmap quality metrics, detector modeling to optimize the detector geometry, and acquisition of additional data in the proton beam line at higher proton energies and beam currents.

Abstract

Accurate proton range verification (PRV) in proton therapy (PT) is an unmet clinical need and is essential to exploit the full potential of PT to maximize treatment effectiveness and minimize toxicity. We present a Monte Carlo (GATE 10) study of a Cherenkov-based prompt gamma imaging system combining an array of PbF2 crystal detectors with a MURA coded-aperture collimator for real-time PRV. The detector response was studied through detected Cherenkov photon distributions generated by prompt-gammas with energies ranging from 0.5 MeV to 7.0 MeV in steps of 0.5 MeV. A threshold of 15 detected Cherenkov photons was applied to reject low-energy background, which effectively suppressed 511 keV annihilation gammas and sub-MeV contributions while retaining the prompt gammas of interest. Image reconstruction was performed using the Maximum Likelihood Expectation Maximization algorithm, with the system matrix computed with 4.4 MeV point-like sources. The imaging performance was validated on isotropic gamma point sources in both 1D and 2D coded-mask configurations across a field of view of ∼130 mm. A direct comparison between discrimination based on energy deposition (E > 1 MeV) and discrimination based on the number of detected Cherenkov photons (nckv ≥ 15) showed comparable performance, with position residuals consistent with zero within statistical uncertainties in both cases. For the proton beam measurements, background rejection was performed exclusively via the Cherenkov photon count threshold. Under realistic proton beam conditions, the distal fall-off position was reconstructed with a systematic residual below 1.5 mm and a statistical precision of less than 2 mm for beam intensities above 10^8 protons, corresponding to approximately 10 ms of irradiation at a standard proton current of 2 nA (1.25 × 10^10 protons/s). This work presents the concept and shows the feasibility of using Cherenkov-based coded-aperture imaging for PRV by means of a simulation framework.

Abstract

Purpose: We are developing a Personalized Remote Radiation Tracking Portable Organ Dosimetry Device (PRRT PODD) for at-home monitoring of patients receiving radiopharmaceutical therapy (RPT). We report on an FPGA-based data acquisition system capable of wireless communication in support of this goal.

Methods: Our acquisition system supports16-radiation detectors and outputs an energy histogram for each using a time-over-threshold (ToT) energy measurement. The electronics consists of a custom analog board that converts each detector's analog signal into a digital ToT signal; an FPGA board for signal processing; and a microcontroller board that serves as the system controller. The microcontroller communicates with the FPGA via UART to retrieve energy histogram data, manages system configuration parameters, and provides Bluetooth Low Energy connectivity to a mobile Android app. The electronics fits within the PODD’s base assembly and can operate on cell phone battery power. Each detector has a small thermistor for individual probe temperature monitoring and to characterize MPPC gain versus temperature. An Android app was developed to control data acquisition and wireless transfer of data to a remote workstation. The detector and electronics acqusition system was tested using Tc-99m, Na-22 and Lu-177 radioactive sources at clinically relevant counting rates.

Results: Sixteen channel operation of our FPGA-based data acquisition system has been validated using a variety of radionuclides. The output properties of the probes have been characterized versus temperature to enable post acquisition normalization of signals.

Conclusion and future work: A compact, wireless 16-channel gamma spectroscopy system for remote monitoring of Lu-177 RTP was designed and tested. Future work includes implementing real-time temperature compensation based on continuous temperature readings, ensuring stable operation without post-processing.

Abstract

Metal implants can degrade CT image quality and introduce artifacts that affect dose calculation in particle therapy. This study investigates attenuation images (µ‑MLAA) reconstructed from time‑of‑flight PET data using maximum‑likelihood reconstruction of attenuation and activity for metal artifact reduction. The linearity of a calibration curve relating CT number to relative electron density was first evaluated using a PMMA phantom with metal and tissue‑equivalent inserts. Metal artifacts were then studied in a water phantom containing a Ti6Al4V total endoprosthesis (TEP) implant. µ‑MLAA attenuation images reconstructed from PET list mode data and CT images reconstructed using iterative metal artifact reduction (iMAR) were evaluated using gamma index analysis (10%, 1 mm) relative to an artifact‑free reference slice. The calibration curve remained linear in the presence of metal inserts (R² > 0.999 ). For the TEP phantom, µ‑MLAA images achieved a higher gamma pass rate (96.3%) than iMAR CT (70.8%). Using the calibration curve, the relative electron density of the TEP shaft was estimated as n_{e, rel} = 3.58±0.12  corresponding to a 3.5% deviation from reported reference values. These results indicate the potential of µ‑MLAA images for reducing metal artifacts and supporting implant characterization in radiotherapy imaging.

Abstract

PET-based range verification in ion-beam therapy requires decay correction for a depth-dependent mixture of positron emitters. We introduce a range-verification-oriented carbon-ion PET workflow that partitions beam-axis reconstructions into axial slices and fits post-beam-off time-activity curves with four fixed-half-life components (¹¹C, ¹⁵O, ¹⁰C, ¹³N); the small ¹³N term (~0.4%) was grouped with ¹¹C for comparison. The experiment contained 120 reconstructed 10-s frames, and corresponding GATE simulation modeled 221.48 MeV/u carbon ions with 2.0 x 10_⁷ primary particles and seven positron-emitting nuclides. Across 35 accepted PET slices (R² >= 0.9), fitted fractions were 82.6%, 12.1%, 4.9%, and 0.4%, showing clear ¹¹C dominance. After converting fitted amplitudes to observation-window counts, plateau PET fractions were 77.5%, 14.5%, and 8.0%, versus grouped simulation values of 63.2%, 27.1%, and 9.7% for ¹¹C+¹³N, ¹⁵O+¹⁴O+¹⁷F, and ¹⁰C. Carbon-ion PET agreement is therefore nuclide dependent: ¹⁰C showed the smallest residual bias after correction, whereas long-lived carbon/nitrogen activity remains high and oxygen-group activity remains low. This nuclide-aware comparison reveals isotope-group-specific discrepancies that total-activity range metrics cannot localize.

Abstract

The IRIS group at IFIC-Valencia, as part of the FARO (Feasibility of Adaptive Surface-Guided Radiation Oncology) project, funded under the VLC-BIOMED 2024 call (AP2024VLC-01), has developed a prototype to ensure the reproducibility of patient positioning, movement management in oncology applications and correlate the patient position across different scans. The system, based on Lighthouse v2 technology, uses base stations that emit infrared light and detectors that record these pulses. Experimental measurements at La Fe University and Polytechnic Hospital have been carried out using a torso phantom coupled to a motor that emulates different respiratory patterns. The results show that our non-invasive prototype is capable of calculating static position and reproducing respiratory movement with millimeter accuracy.

Abstract

Accurate in vivo dose range verification is critical in proton therapy due to uncertainties in beam range. Prompt gamma imaging (PGI) is promising because PG emissions correlate with the Bragg peak. However, conventional detectors can suffer from signal pile-up under high-intensity proton beams. Cherenkov detectors are attractive for PGI due to their selective gamma sensitivity, ultrafast response, and high count-rate capability. In this work, a PGI system based on PbF₂, a pure Cherenkov-emitting material, was simulated, and a multilayer perceptron (MLP) model was employed to predict proton range and one-dimensional dose distributions. The detector comprises a 32 × 32 PbF₂ crystal array coupled to a silicon photomultiplier (SiPM) array, with a knife-edge slit collimator providing one-dimensional PG projections. Proton beams from 70 to 120 MeV in 10 MeV intervals were simulated in a PMMA phantom using Geant4, and 1000 PG profiles per energy were generated via Poisson resampling. The MLP was trained on 70–120 MeV beams (80% training, 20% validation), while intermediate energies from 75 to 115 MeV (in 10 MeV intervals) were reserved for testing. Predicted dose distributions agree well with ground truth, achieving sub-millimeter proton range accuracy and mean relative errors (MRE) of 0.6–2.4%. These results demonstrate the feasibility of combining Cherenkov-based PGI with machine learning for accurate dose and range prediction, supporting future real-time, beam-by-beam verification in proton therapy.

Abstract

Post-thyroidectomy whole-body imaging after Iodine-131 administration in differentiated thyroid cancer is performed to verify biodistribution, identify thyroid bed remnants, detect locoregional and distant metastases, and support patient-specific dosimetry. The clinical standard, planar scintigraphy with a high-energy parallel-hole collimator, suffers from photon collection efficiency below 1% and depth ambiguity that can mask small remnants. We present a GATE/Geant4 Monte Carlo comparison of a clinical planar gamma camera and a compact handheld silicon/cesium iodide Compton camera at the 364 keV photopeak of Iodine-131 This translates into shorter acquisition times in clinical practice. Preliminary results indicate a sensitivity advantage of 30% for the Compton camera over the gamma camera in planar mode. The handheld geometry enables point-of-care bedside imaging of high-activity post-therapy patients, with potential to transform post-radioiodine therapy monitoring and dosimetry in thyroid cancer.

Abstract

Patient-specific dosimetry in peptide receptor radionuclide therapy (PRRT) using ¹⁷⁷Lu-DOTATATE commonly requires multiple quantitative SPECT/CT acquisitions to estimate time-integrated activity and absorbed dose distributions. Although multi-time-point imaging improves kinetic characterization, it increases clinical workload and patient burden. Single-time-point (STP) dosimetry approaches have therefore been proposed as practical alternatives, although conventional analytic STP methods may have limited accuracy in heterogeneous activity distributions.

In this work, CNN-based voxel dosimetry frameworks for ¹⁷⁷Lu-DOTATATE SPECT/CT were compared against conventional Hänscheid-based STP dosimetry. Quantitative SPECT/CT datasets from 44 paediatric treatment cycles were retrospectively analysed. Reference voxel-level absorbed dose maps were generated using multi-time-point dosimetry and GATE Monte Carlo simulation. Conventional STP dosimetry was performed using the Hänscheid formalism and a GATE-derived dose voxel kernel (DVK). A 3D U-Net architecture was used to estimate voxel-level absorbed dose maps from co-registered activity and density distributions using either a single 96 h activity input or paired 4 h and 120 h activity inputs.

CNN-based dosimetry demonstrated reduced organ-level absorbed dose deviations across the liver, spleen, kidneys, and tumour volumes relative to conventional STP dosimetry. Mean voxel-wise NMAE and NRMSE values decreased from 45.0 % and 54.3 % for conventional Hänscheid-STP dosimetry to 20.6 % and 27.4 % for CNN-STP dosimetry using a single 96 h activity input. The CNN model using paired 4 h and 120 h activity inputs further reduced these values to 18.8 % and 25.1 %, respectively. Improved preservation of heterogeneous tumour absorbed dose distributions was additionally observed in dose-volume histogram analysis.

Abstract

Positron emission tomography has been routinely used over the past 20 years for treatment delivery monitoring in charged particle radiotherapy. Based on previous work from our group on Monte Carlo simulations of real patient treatment plans, we report on the design and characterization of detector modules and readout electronics of a dedicated dual head PET system for treatment verification in proton therapy. The detector modules are made of LYSO:Ce matrices optically coupled to silicon photomultiplier matrices, with custom-designed compact readout electronics based on the 16-channel HRFlexToT ASIC, developed by our group in collaboration with the University of Barcelona. Each 64-channel detector module is read out by 4 ASICs, followed by an FPGA which records the time stamp of each event with respect to a global clock using a coarse and a fine counter, and digitizes the time-over-threshold output of each individual channel. Results from the characterization of the detector modules under irradiation of a 22Na source have shown that all channels are readout independently, with energy spectra for all channels yielding an average 16% energy resolution at the 511 keV photopeak. The time stamp of each event was calculated at FPGA level down to 87 ps binning, allowing time-of-flight PET image reconstruction. Moreover, individual readout of all channels allows the identification of intercrystal scatter events. We are currently developing a small prototype PET system based on these detector models and readout electronics to evaluate different treatment verification methods in proton therapy.

Abstract

Prompt gamma imaging is a promising technique for in vivo range verification in particle therapy. Compared with protons, Helium-ion (He²⁺) beams may produce prompt gamma (PG) emissions different from proton beams, affecting energy-window (EW) selection and range-positioning performance. This study investigates the PG energy spectra, EW-dependent yields, and spatial distribution features of He²⁺ irradiation in a homogeneous PMMA phantom.

Monte Carlo simulations were performed using GATE for He²⁺ pencil beams from 70 to 230 MeV/u. Secondary gamma rays generated within the phantom were recorded using a PhaseSpaceActor, including their production positions, kinetic energies, direction vectors, and particle information. The PG spectra were analyzed and compared with proton-induced spectra. Several energy windows were selected according to characteristic spectral regions, and their PG yields, longitudinal distributions, and range-positioning performance were evaluated using Falloff50%, Falloff80%, and Steepest Falloff metrics.

The simulations showed He²⁺ irradiation produced substantially higher PG yields than proton irradiation over the investigated energy range, while the main spectral peaks appeared at similar energies. For He²⁺, PG yield increased with incident energy in all selected windows. The broad 1.9–8 MeV window provided the highest yield and best statistical stability, whereas narrower characteristic windows showed stronger energy specificity but lower counting statistics. EW selection also affected the distal fall-off shape and the derived positioning value. Overall, the 1.9–8 MeV window combined with the Steepest Falloff metric provided the most robust PG range-positioning performance, supporting the development of helium-ion PGI range verification strategies.

Abstract

Time-integrated activity (TIA) estimation is essential for patient-specific Lu-177 SPECT dosimetry, but voxel-wise time–activity curve fitting remains computationally demanding because nonlinear kinetic models must be solved independently across many voxels. This study developed a GPU-accelerated voxel-wise kinetic modeling framework for rapid Lu-177 SPECT TIA estimation. Four serial post-therapy SPECT/CT scans from anonymized Lu-177 DOTATATE datasets were analyzed. The framework used a hierarchical TAC model library, including one mono-exponential model and three bi-exponential variants, to characterize different uptake and clearance patterns. Model selection was performed using the Akaike information criterion, and voxels with insufficient fitting quality were processed using trapezoidal integration with tail extrapolation. GPU acceleration was implemented through batched voxel-wise optimization using CuPy-based array operations and compared with a CPU-based Levenberg–Marquardt implementation. Evaluation included fitting quality, model-selection distribution, agreement with dataset-provided reference TIA maps, organ/lesion-level relative error, and computation time. The GPU implementation preserved the CPU-derived model-selection distribution, with 60.7% mono-exponential, 31.2% bi-exponential, and 8.1% trapezoidal fallback voxels. The median regional voxel-wise fitting R² was 0.848. Relative differences from reference TIA maps were 4.2% for kidneys, 3.5% for spleen, 2.7% for liver, and 8.5% for tumor lesions. Computation time decreased from 552.1 s to 18.5 s per patient, corresponding to approximately 30-fold acceleration. These preliminary results support the technical feasibility of GPU-accelerated voxel-wise TAC fitting for practical Lu-177 SPECT dosimetry, although validation in larger multi-patient datasets remains necessary.

Abstract

Boron Neutron Capture Therapy (BNCT) requires accurate in-vivo dose verification, yet current clinical methods rely on indirect blood-based boron concentration estimates that introduce substantial inter-patient variability. Imaging the 478 keV prompt gamma ray emitted during boron-10 neutron capture offers a direct route to real-time dose mapping. However, this requires a detector capable of efficient reconstruction at high gamma energies in the presence of significant background.

Building on our previously reported scintillator-based Compton camera concept [1], we present the full computational characterisation of a three-stage system comprising scintillator cylinders arranged in a three-dimensional spherical array. The non-planar geometry offers multi-directional cone orientations, improving global spatial resolution compared to conventional planar designs. Geant4 Monte Carlo simulations were used to model a realistic BNCT irradiation scenario [2], incorporating a modified Snyder head phantom with clinically representative boron concentrations (25 ppm brain / 65 ppm tumour) [3].

Reconstruction performance is evaluated across scenarios of increasing complexity: an idealised point source, a distributed gamma-ray source, a treatment-like gamma source, and a full phantom simulation. This progressive approach illustrates how reconstruction quality is affected by increasing complexity and the high neutron-induced background inherent to BNCT environments, providing insight into the current limitations of the design and potential strategies to address them. These results establish a validated simulation framework for a scintillator-based Compton camera tailored to BNCT prompt gamma imaging, providing the foundation for forthcoming experimental characterization and clinical translation.

Abstract

Selective internal radiation therapy using ⁹⁰Y microspheres provides targeted treatment of liver malignancies, with post-therapy PET enabling treatment verification and voxel-based dosimetry. Quantitative ⁹⁰Y PET imaging is challenged by a low signal-to-background ratio. The internal pair production branching ratio of ⁹⁰Y is low, yielding a weak 511 keV annihilation photon signal amid background contributions from bremsstrahlung following β^- decay and intrinsic radiation from ¹⁷⁶Lu in PET systems constructed with LYSO scintillation crystals. These background events increase random coincidences and degrade spatial localization and dosimetric accuracy. This study evaluated the effect of PET energy window selection on ⁹⁰Y radioembolization dosimetry using standard (425‑650 keV) and narrow (480‑542 keV) energy windows. Six patients underwent post-therapy PET/MR imaging. Images were reconstructed using TOF‑OSEM, and absorbed dose distributions were calculated using local deposition modelling. Randoms fraction, apparent activity outside the perfused volume, dose-volume histograms, and dose-volume metrics were compared between protocols. The narrow-window protocol reduced the randoms fraction from 91 ± 3% to 60 ± 10% and apparent activity outside the perfused volume by 51 ± 16%, indicating reduced background contamination and improved confinement of the reconstructed activity distribution. Mean absorbed doses to the tumour and healthy liver were preserved, and healthy liver V30 decreased by 13 ± 8%. Tumour D70 changed by -15 ± 33%, and tumour dose redistribution effects varied substantially between patients without a consistent directional trend. These findings demonstrate that narrow-window PET acquisition improves activity localization in ⁹⁰Y radioembolization while preserving overall absorbed dose estimates and may influence patient-specific voxel-based dosimetry.

Abstract

Particle therapy exploits the Bragg peak for precise dose deposition, yet the high-LET characteristic demands accurate real-time range verification and dose distribution monitoring. Compton cameras offer a promising collimator-free solution for the prompt gamma imaging induced during treatment. In this work, a Compton camera is constructed by using Gd₃Al₂Ga₃O₁₂:Ce (GAGG) scintillator coupled with silicon photomultiplier (SiPM) arrays. The scatter and absorber layers are both constructed from high-light-yield, high-energy-resolution GAGG with optimized geometry. An energy resolution of ~8% FWHM is obtained for 511 keV gamma rays, while the position resolution is 1.5 mm (scatter layer) and 2.0 mm (absorber layer) in both x and y directions. The scatter layer retains the capability for dual-ended readout, with an expected DOI resolution of ~2.5 mm. Image reconstruction is performed using GPU-accelerated LM-MLEM. Laboratory measurements with a ²²Na source demonstrate that an angular resolution of ~7° at 511 keV and ~6° at 1274 keV was achieved. Future work will focus on proton/carbon ion beam experiments and scaling the effective detector area to 100 mm × 100 mm.

Abstract

This study presents a data-driven framework for optimizing dosimetry regimens in radiopharmaceutical therapy using in vivo data from [¹⁷⁷Lu]Lu-PSMA-617-treated murine models incorporating therapy resistance. Experimental data demonstrated improved long-term outcomes for cumulative (3 × 10 MBq) and de-escalation (20 MBq + 10 MBq) regimens compared with conventional single-dose (30 MBq) and escalation (10 MBq + 20 MBq) schedules. Building on these findings, additional treatment regimens were evaluated in silico to identify combinations of favorable therapeutic characteristics. Dose-rate kernels derived from dynamic SPECT/CT imaging were used to calibrate an integrated dose–response and tumor growth model. These kernels captured both the temporal evolution of dose rates following treatment and spatial intratumoral heterogeneity through region-specific dose quantiles. Therapy resistance across treatment cycles was incorporated using generalized dose-rate patterns derived from experimentally observed relationships between treatment response and previously administered activity ratios. The spatial distribution of dose-rate regions used for kernel generation was best described by an exponential decay model, yielding a lower Akaike Information Criterion (AIC: −139 ± 31) than alternative distributions. Simulations identified the 3–2–1 (15, 10, 5 MBq), 2–1–1 (15, 2 × 7.5 MBq), and 2–2–1 (2 × 12, 6 MBq) regimens as outperforming the experimentally tested schedules. These results demonstrate the potential of simulation-based optimization to support the design of biologically informed radiopharmaceutical therapy regimens and provide a framework for systematic in silico treatment planning.

Abstract

While recent deep learning techniques for accelerating Monte Carlo (MC) radiotherapy dose calculation have shown promising results, most state-of-the-art methods rely on a denoising paradigm. This approach aims to predict a high-sampling (HS) reference dose map from a fast low-sampling (LS) input. However, information loss and stochastic noise in LS inputs often limit the accurate recovery of fine structural details and dose gradients, and may lead to hallucinations. To address these limitations, we propose a novel learning framework, Energy-Shifting, which reconstructs a time-intensive polyenergetic HS dose distribution from a GPU-accelerated monoenergetic dose map. Our results demonstrate that this method significantly outperforms conventional denoising approaches in accuracy.

Abstract

Image-based dosimetry requires multiple-time-point (MTP) imaging to estimate the time-integrated activity (TIA), which can be challenging to implement in clinical practice. Moreover, Mouse-to-human extrapolation methods produces large errors. This study aimed to evaluate single-time-point (STP) dosimetry approach using extrapolated human effective half-life derived from preclinical activity scaling based on body weight and organ mass. Six patients underwent where administrated with ⁶⁴Cu-DOTA-rituximab and underwent MTP PET/CT imaging at 1, 24, and 48 h post-injection (p.i.), while four mice were imaged at 1, 24, 48, and 72 h p.i. Regions of interest were manually drawn for the kidney, liver, and spleen. Human MTP reference TIA was calculated using OLINDA with mono-exponential fitting. Mouse organ activity was converted to estimated human organ activity using body weight and organ mass scaling as: (%ID/organ)_human=(%ID/g)_mouse×(kg_WB )_mouse×(g_Organ/kg_WB )_human. Human TIA was then estimated from the extrapolated activity curve. In the proposed approach, the extrapolated human time–activity curve was used to derive the extrapolated human effective half-life. This effective half-life was then combined with the measured human organ activity from STP scan and TIA was calculated as following: TIA=(A(T_sc)/ln(2))×2^{Tsc/Teff(extrapolated)}× T_{eff(extrapolated).} . The proposed STP approach reduced the estimation error for kidney from 88.7% with the conventional extrapolation approach to 29.0%, 40.6%, and 66.7% at 1, 24, and 48 h p.i., respectively. The corresponding errors decreased from 406% to 26.8%, 35.1%, and 65.3% for the liver, and from 480% to 42.2%, 57.7%, and 76.2% for the spleen, respectively. These results demonstrate that incorporating measured human STP activity with extrapolated human effective half-life improves TIA estimation and provide a practical dosimetry approach when full MTP imaging is not feasible.

Abstract

Respiratory phase mismatches between PET and CT commonly cause artifacts in PET/CT, especially at lung–liver and lung–spleen interfaces. Since CT captures a single respiratory phase while PET reflects an average over breathing, this misalignment leads to attenuation correction errors and inaccurate lesion quantification, particularly near the liver dome. We propose a deep learning framework that uses PET-based pseudo-CT to guide deformable CT-to-PET alignment. A retrospective dataset of 10 18F-FDG exams with respiratory phase mismatch (banana) artifacts, scanned on a Total-Body BGO-based PET/CT scanner, was used for evaluation. A previously trained 3D residual neural network was used to generate pseudo-CT images from non-attenuation corrected (NAC) PET images. The pseudo-CT images, inherently aligned with the PET data, were then used as a reference for deformable registration of the CT images. The original and registered CT images were used for attenuation correction and reconstruction of the PET data using a BSREM algorithm. Regions of interested were placed in dome of the liver to quantitatively measure the SUV changes in those regions affected by the artifacts. Compared to baseline PET/CT without correction, the proposed method effectively reduced or eliminated respiratory-induced PET/CT misalignment artifacts were significantly reduced or eliminated. Quantitatively, the method achieved improvements in liver SUVmean towards the expected values. By leveraging the native alignment between PET and the generated pseudo-CT, the method avoids direct multimodal registration and instead enables accurate deformable alignment of CT to PET space. This approach improves attenuation correction and quantitative accuracy, with potential to enhance diagnostic confidence in routine clinical PET/CT imaging

Abstract

The extension of the axial field-of-view (AFOV) represents a transformative advancement in PET, offering superior sensitivity and enabling ultra-low-dose 4D imaging. However, the increased AFOV necessitates wider acceptance angles and correspondingly larger coincidence time windows, leading to a substantially higher proportion of random coincidence events compared to conventional scanners. While traditional random correction relies on noisy, sparse delayed coincidence sinograms, the singles-rate method has been approved as a superior alternative for long-axial FOV (LAFOV) systems. A critical yet underexplored challenge in LAFOV PET is the significant variation in the required time window across different ring differences. Although object-specific customized windows have shown promise in short-axial systems, their application to total-body PET remains largely unoptimized. In this study, we propose a novel reconstruction strategy utilizing a variable time window tailored to specific ring differences to minimize random fractions. Using GATE simulations of a 1.5-m total-body PET system (uMI Panorama GS), we designed and evaluated a variable time window set based on practical coincidence time resolution (CTR). This approach was benchmarked against a standard fixed window (6.0 ns) and a 3-step window (4.7, 6.2, and 8.3 ns).  Systematic analysis of simulated coincidence data reveals that the variable window design significantly reduces the random fraction. Comparative reconstructions of a uniform cylinder, NEMA, and torso phantoms demonstrate that this reduction directly translates to lowered image noise levels and enhanced signal-to-noise ratios within a consistent reconstruction framework. These findings establish that ring-difference-specific time windows provide a robust technical pathway for maximizing the quantitative performance of next-generation LAFOV PET systems.

Abstract

The HD-MetaPET project is developing a high-resolution positron emission tomography insert for simultaneous magnetic resonance imaging. The insert will consist of 120 Hyperion Visor Max detector tiles with an area of 56.5 x 56.5 mm², arranged in two semi-circles and coupled to LYSO:Ce crystal scintillators. A crucial optimisation parameter for the upcoming system performance is the crystal topology. This study evaluates Time-of-Flight and Depth-of-Interaction positioning performance across several crystal topologies: continuous slabs, pseudoslabs, and pixelated arrays with a light guide coupled to the face opposite the photosensitive detector. The study utilizes the Hyperion Visor Max and a custom development board, both featuring the TOFPET2 ASIC. Experimental measurements are performed using a ²²Na point source. The preliminary results with crystals of different materials show that pixelated arrays with light guides outperform continuous slabs. Specifically, pixelated designs achieved a Coincidence Time Resolution of ≈235 ps, compared to >400 ps for the slabs. The results indicate that the pixelated designs utilizing the light guide also demonstrated superior accuracy in the evaluation of the Depth-of-Interaction. The pixelated light-guide topology shows promising performance, circumventing the traditional compromise between spatial depth accuracy and timing resolution. Consequently, this configuration emerges as a good candidate for integration into the upcoming HD-MetaPET scanner. While these preliminary results are promising, the comprehensive evaluations utilizing final-scale crystals of the same material and the final scanner electronics will be presented at the conference.

Abstract

Dynamic PET imaging enables tracer kinetics characterization, offering complementary quantitative insights beyond static imaging. The advent of Long Axial Field-of-View (LAFOV) scanners, with high sensitivity, has renewed interest in kinetic modelling by enabling whole-body dynamic imaging in clinical settings. However, acquisitions can last up to 1 hour, limiting adoption due to reduced throughput. Clinically, reporting relies on late-phase images (1-hour post-injection), whereas kinetic modelling requires data from injection, creating a conflict between clinical requirements and feasibility. This work investigates whether dynamic protocols can be shortened to 20 minutes post-injection while preserving clinically relevant information by predicting late-phase time-activity curves (TACs) from early dynamic data. The cohort included 107 subjects imaged with ¹⁸F-FDG, yielding 20,903 TACs extracted from multiple anatomical structures. Four approaches were compared: a three-exponential model, two-tissue compartment model (2TCM), a gated recurrent unit encoder-decoder, and a novel Neural Ordinary Differential Equation framework (NODE). At 60 minutes, all methods showed near-zero median residuals. Over 20–60 min, NODE outperformed competing approaches with the lowest root mean squared error (0.096) and χ² (0.43), and the highest concordance correlation coefficient (0.40), with reduced variability across TACs. Overall, prediction of late-phase PET information from 0–20 min data appears feasible. NODE provided the most accurate and stable results, although 2TCM remained competitive at 60 min, suggesting benefits from incorporating physiological priors. Future work will extend the analysis to voxel-wise reconstruction, with preliminary results showing promising outcomes, and will examine kinetic parameter consistency between observed early dynamics (0–20 min) and full predicted dynamic data (0–60 min).

Abstract

Partial volume effects (PVE) degrade quantitative accuracy and blur boundaries in PET. MR-guided partial volume correction (PVC) can mitigate PVE, but it requires anatomical information from MR, which may be unavailable. This work aims to develop a 2-stage deep learning-based PET to MR synthesis framework, SADCA-ResDiff, to provide structural priors for MR-guided PVC in 18F-AV45 amyloid brain PET. A total of 135 pairs of 18F-AV45 PET and MR were selected from ADNI [1]. A coarse MR is first generated from PET using a dual cross-attention UNet, while a residual diffusion network with an edge-aware mechanism is applied to refine the coarse MR. Joint mutual information, diffusion, reconstruction, and structure-aware symmetric consistency losses derived from a pretrained MR segmentation model were used to optimize the 2-stage network. Synthesized MR was segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid masks using FSL [2] and applied to the Iterative-Yang (IY) PVC [3]. SADCA-ResDiff was compared with DeepPVC [4], using standard IY as reference. Structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) was measured on the whole brain while regional error was evaluated using relative mean absolute error (RMAE) for GM and WM. Wilcoxon signed-rank test was used for statistical comparison.
SADCA-ResDiff achieved the highest SSIM and PSNR among all methods (0.97±0.02 and 32.62±3.43dB), significantly outperforming DeepPVC (0.93±0.02 and 27.95±3.67dB) and origin PET (0.90±0.02 and 26.73±3.89dB) (all p<0.05). It also yielded the lowest RMAE in GM and WM (0.12±0.05 and 0.08±0.05), compared with DeepPVC (0.14±0.04 and 0.15±0.03) and origin PET (0.16±0.04 and 0.15±0.03).
SADCA-ResDiff provides effective anatomical guidance for PVC in amyloid PET, offering a practical solution for MRI-free PET quantification.

Abstract

A locally developed database of human brain studies from 113 subjects has been used to develop and train convolutional neural networks (CNNs) for attenuation correction in SPECT. All subjects underwent brain  perfusion SPECT imaging studies with ^99mTc-ECD, an X-ray CT scan, and MRI. Images were preprocessed for anonymization, removal of the table in CT, co-registration, and spatial and intensity normalization. Deep learning architectures encompassed U-Net, Residual U-Net, Attention U-Net and Residual Attention U-Net. Three methods were evaluated: (a) direct mapping of uncorrected SPECT to corrected SPECT, (b)  multisequence MRI scans to synthetic CT (sCT) mapping, followed by iterative reconstruction with sCT attenuation correction, and (c) CNN reconstruction using sinograms to directly generate attenuation-corrected  SPECT images. To evaluate the attenuation-corrected images, the mean square error (MSE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were calculated. Attenuation corrections were successfully applied using three different approaches, addressing various clinical scenarios found in clinical medical equipment that does not have CT for attenuation correction, such as dedicated SPECT or SPECT/MRI.

Abstract

Multiplexed positron emission tomography (mPET) remains limited by the challenge of separating overlapping tracer signals, as tracers produce indistinguishable same-energy photon pairs.
Existing deep learning-based separation methods are constrained by the limited real data availability, particularly in the context of total-body imaging.
In this work, we propose a framework for training total-body dual-tracer mPET separation models using synthetic paired datasets generated from independent single-tracer acquisitions, and investigate the effect of synthetic data diversity on separation performance.
Two score-based generative models were trained to generate synthetic [18F]FDG and [18F]PSMA PET images conditioned on attenuation maps.
Complementary synthetic tracer images were combined with real PET acquisitions to construct paired dual-tracer datasets for training a Swin UNETR-based separation model.
Quantitative evaluation on dual-tracer combinations demonstrated the feasibility of the proposed framework and suggested that increased diversity of synthetic samples may improve separation robustness.
These results highlight the potential of generative augmentation as a scalable strategy for pretraining mPET separation models under limited paired-data availability.

Abstract

Diffusion models have shown strong potential for cardiac-gated myocardial perfusion single-photon emission computed tomography (CG MP-SPECT) denoising. However, centralized or fully federated diffusion training requires each client to process a complete high-capacity model, which may be impractical for resource-limited nuclear medicine departments. This study presents a multi-client split federated learning diffusion framework (SFLDiff) for resource-efficient CG MP-SPECT denoising. SFLDiff vertically partitions the diffusion network into lightweight client-side encoders and a server-side reconstruction module, enabling local clients to perform shallow feature extraction while computationally intensive generative restoration is executed on the server. This architecture reduced client-side GPU memory consumption from over 4300 MiB to approximately 257 MiB with a training time of 2950.39 s while maintaining competitive denoising performance. Experiments on 8-/16- cardiac gated datasets showed that SFLDiff achieved image restoration accuracy comparable to centralized and federated diffusion baselines and superior to standard split learning, particularly under high-noise gated conditions. These results suggest that multi-client split federated diffusion provides a practical quality–efficiency trade-off for deploying advanced generative denoising models in resource-constrained CG MP-SPECT environments.

Abstract

Dual-time-point positron emission tomography (PET) provides complementary temporal uptake information for lesion characterization, but delayed imaging increases examination time and clinical workload. This study proposes UCD-Net, a weakly supervised cyclic disentanglement framework for synthesizing delayed-pass PET from first-pass PET. UCD-Net decompose first-pass PET images into a delayed-equivalent component, a signal-dependent residual component, and a signal-independent residual component, followed by cyclic re-decomposition to improve latent-component separation. Reconstruction, zero-mean, identity-consistency, and noise-statistical constraints are used to preserve image content and suppress residual interference. A three-dimensional projection-excitation module captures global spatial statistics along depth, height, and width to enhance anatomical consistency. Real delayed PET images are incorporated as weak distribution-level guidance through a rank-based consistency loss, allowing the model to better recover delayed uptake patterns without direct image-to-image supervision. This retrospective study included 90 PET scans from 70 patients, including 86 ¹⁸F-FDG PET scans and 4 ⁶⁸Ga-PSMA PET scans, divided into training, validation, and test sets. The weakly supervised UCD-Net was compared with its unsupervised backbone and four baseline models, including EDCNN, CGAN, U-Net, and CVF-SID. Across three test scenarios, the weakly supervised UCD-Net achieved the best performance, with RMSE values of 0.0333, 0.0365, and 0.0236. Compared with unsupervised UCD-Net, weak delayed-image guidance reduced RMSE by 17.2%, 20.8%, and 30.0%. These results suggest that UCD-Net can synthesize delayed-equivalent PET images from first-pass PET and may reduce dependence on additional delayed imaging.

Abstract

This research aims to detect and reconstruct object changes using computed tomography (CT). We focus on stationary CT with linearly distributed sources (SLCT), which enables rapid projection data acquisition without mechanical rotation. However, due to the lack of sufficient angular coverage, SLCT inherently suffers from a severely limited-angle imaging problem, where direct reconstruction leads to strong anisotropic artifacts. To address this issue, we introduce the pre-change object as prior knowledge and propose two Neural Radiance Field (NeRF)-based change modeling strategies. The first method, Full-Object Nerf (FO-NeRF), continues optimization from a pre-trained NeRF representing the pre-change object, refining it into the post-change object. The second method, Change-Filed NeRF (CF-NeRF), explicitly models a change field and directly learns the residual component between the pre- and post-change objects. We validate the proposed methods on 2D simulated objects composed of simple geometric shapes. Experimental results demonstrate that, with prior constraints, both methods can successfully reconstruct object changes while effectively mitigating limited-angle artifacts. These results indicate that, by leveraging object priors within a NeRF-based representation, the highly ill-posed limited-angle reconstruction problem can be transformed into a more stable and tractable optimization problem, enabling reliable change detection and reconstruction under severely incomplete projection data.

Abstract

A partial-detector-ring Positron Emission Tomography (PET) system may be an option to reduce cost and/or enable integration within another modality for multimodality imaging. In some use cases, the PET system may not be able to rotate continuously to collect all lines of response needed to form an image, yet a rapidly available image is still required for image quality control or organ-motion detection when the detectors are stationary. In this setting, missing angular coverage yields incomplete projection measurements and degraded reconstructions. This work describes a learning-based method to generate results equivalent to full-ring acquisition from partial-ring data This work proposes a reconstruction-aware sinogram completion network that predicts a full-ring-equivalent sinogram while enforcing consistency after tomographic reconstruction. This method combines supervision in projection space with an additional image-space objective computed after reconstructing the predicted sinogram using standard PET reconstruction operators, encouraging outputs that both match full-ring sinograms and yield accurate reconstructed images. By incorporating reconstruction consistency, the reconstruction-aware approach is intended to enable partial-ring systems to produce full-ring-equivalent sinograms that can be reconstructed with existing pipelines without modification, supporting more practical imaging workflows when mechanical rotation is constrained.

Abstract

Automated bone metastasis detection in whole-body bone scintigraphy (WBBS) is often hindered by subtle contrast, scale variation, and physiological interference. We propose an enhanced YOLOv11-based framework incorporating a symmetry-aware representation—utilizing ECC (Enhanced Correlation Coefficient)-registered mirrored images and difference maps—to model pathological asymmetry. To handle multi-scale lesions, we introduce a scale-sensitive architecture featuring a P2 detection branch and Efficient Multi-scale Attention (EMA). Furthermore, an extremity-specific refinement branch and an anatomical cross-view fusion strategy (integrating anterior/posterior projections) are employed to improve sensitivity in limbs and reduce single-view occlusion. Evaluated on 1,800 patient studies, our method achieved 0.959 recall, 0.820 precision and 0.785 mAP50, outperforming YOLOv8 and Faster R-CNN, particularly for small, low-contrast lesions. The exceptionally high recall demonstrates superior sensitivity to small and low-contrast lesions, highlighting the method's potential as a reliable screening tool to minimize missed diagnoses in clinical practice.

Abstract

Accurate Aβ PET/CT classification is critical for Alzheimer's diagnosis and anti-amyloid treatment eligibility, yet existing methods struggle to model higher-order inter-regionbrain relationships and handle class imbalance and per-class learning hardness. We propose HyperAβ, which constructs a 30 node brain-region hypergraph over SynthSeg-parcellated multimodal PET/CT features with 10 anatomically-defined hyperedges,and employs hypergraph convolution for higher-order multiregion interaction modeling. To address class imbalance and per-class learning hardness, we introduce GCA-D loss combining static per-class confidence calibration with EMA-based dynamic hardness tracking. Under 5-fold cross-validation, HyperAβ achieves ACC of 0.8784 and AUC of 0.9341, outperforming state of-the-art methods.

Abstract

Targeted alpha therapy using ²²⁵Ac-DOTATATE has shown promise for patients with metastatic or inoperable neuroendocrine tumors. Image-based dosimetry is important for treatment optimization and toxicity assessment; however, quantitative ²²⁵Ac SPECT imaging is challenging due to low injected activity and limited gamma-ray count statistics. This study aimed to investigate the feasibility of generating ²²⁵Ac-DOTATATE SPECT images from planar and CT images using a deep learning-based image generation framework. Five patients were administrated administered with 205 µCi of ²²⁵Ac-DOTATATE were included. Planar, SPECT and CT imaging was performed 24 h after administration using a dual-head Symbia T2 scanner with energy windows centered at 78, 218, and 440 keV. Fusion images were generated from planar and CT images using the conjugate view method described in MIRD Pamphlet No. 16. Planar images, fusion images, and CT images were used as multimodal inputs to a 3D U-Net encoder–decoder network to generate quantitative SPECT images. All images were resized to 128 × 128 × 128 voxels and normalized. The model was optimized using the Adam optimizer with a learning rate of 2 × 10⁻⁵, batch size of 1, and 500 epochs, using L1 loss. The generated SPECT images showed uptake patterns comparable to acquired SPECT images. Quantitative organ-level analysis showed similar uptake distributions across major organs and tumor regions, with relative uptake error below 15% in most organs. These findings demonstrate the feasibility of planar-to-SPECT image generation using deep learning and suggest its potential to support ²²⁵Ac-DOTATATE image-based dosimetry by providing three-dimensional radiopharmaceutical distribution from clinically practical planar and CT images.

Abstract

Reducing positron emission tomography (PET) tracer dose lowers patient radiation exposure but degrades image
quality and the quantitative accuracy of standardized uptake values (SUVs). Paired same-subject low/full-dose acquisitions
are impractical to obtain at scale, motivating unpaired image translation. Applying Contrastive Unpaired Translation (CUT)
to ultra-low-dose PET, we identify a quantitative failure mode in which the trained generator improves image appearance but
introduces high-uptake SUV underestimation that can exceed the bias of the unprocessed low-dose input itself. We interpret this
behavior as a mismatch between the fixed residual weighting of the intensity-space ℓ1 identity loss and PET’s uptake-dependent
Poisson-derived noise. To address this, we propose PA-CUT (Poisson-Aware CUT), which replaces the ℓ1 identity loss with
one evaluated under the classical Anscombe transform, a Poisson-motivated variance-stabilizing reweighting of the identity term. A
multi-dose training scheme yields a single model across multiple dose reduction factors (DRFs). On the UDPET 2024 challenge
dataset, PA-CUT substantially reduces high-uptake SUV bias compared with CUT, with negligible changes in pixel-level and
perceptual quality.

Abstract

Quantitative accuracy in PET relies heavily on effective scatter correction (SC), especially for long axial field-of-view (LAFOV) systems such as the uMI Panorama GS where low-count scenarios are increasingly prevalent. Conventional SC requires an emission map generated through iterative OSEM reconstruction, which is inherently sensitive to noise and degrades at low doses. Our prior work demonstrated that a deep learning (DL) approach using physics-corrected histo-images achieves high quantitative accuracy for SC under standard-dose conditions [2]. 

In this work, we extend this framework to low-dose scenarios on the uMI Panorama GS PET/CT system. We collected 220 static ¹⁸F-FDG cases and 21 dynamic cases. Low-dose conditions were simulated by randomly downsampling events at seven dose levels (0.1%–100%). A U-Net was trained to predict emission maps from physics-corrected histo-images, with full-dose reconstructions as targets. 

At standard dose, both the proposed and conventional methods achieved relative differences below 2% compared with the ground truth. At the 0.1% dose level, the conventional method showed organ VOI deviations approaching 10%, while the proposed method remained below 2% across all dose levels. In dynamic studies, the proposed method preserved time-activity curve accuracy at 0.5% dose and produced reliable Patlak Ki images at 1% dose. Combined with regularized reconstruction, the proposed method enables ultra-low-dose PET data to yield clinically usable images with high quantitative accuracy, demonstrating its potential for low-dose PET applications.

Abstract

This study proposes Quantum Barren-Plateau Correction via Information-theoretic Regularization (QBPCIR), a hybrid quantum-classical architecture for CT radiomics-based pancreatic lesion classification. A parameter-shift-rule diagonal quantum Fisher information matrix (QFIM) estimator quantifies per-class output curvature and adaptively amplifies minority-class gradient contributions to mitigate barren plateau-induced gradient suppression. Evaluated on 1206 training and 71 test portal-venous CT cases with 294 LASSO-selected PyRadiomics features, QBPCIR-VQC achieved AUC 0.916 and recall 0.859, outperforming a classical MLP (AUC 0.846) and a standard VQC (AUC 0.880), and approaching the recall of an expert radiologist reference (0.830). The BCE ablation confirmed that the inter-class Fisher trace asymmetry is a genuine quantum geometric effect, validating the information-theoretic rationale for the proposed correction

Abstract

Accurate organ segmentation in positron emission tomography (PET) is essential for quantitative analysis, treatment personalisation, and clinical decision-making. In current practice, organ masks are commonly derived from CT using tools such as TotalSegmentator (TS) and transferred to PET directly; however, PET–CT misalignment systematically degrades accuracy and biases downstream functional measurements. PET-native supervised approaches mitigate this misalignment but require costly expert annotation and lack generalisability to unseen organs or acquisition protocols. We propose a zero-shot vision-language refinement framework that adapts CT-derived anatomical priors to PET without any PET-specific training. TS masks are converted into multi-modal prompts for SAM3 — comprising organ bounding boxes, interior positive points, boundary negative points, and organ-specific radiological text descriptions — which are applied independently across axial, coronal, and sagittal views. A Prior-Anchored Refinement Gate (PARG) then constrains each SAM3 prediction by enforcing spatial overlap, area consistency, and eroded-core preservation with respect to the TS prior, suppressing anatomically implausible outputs. Evaluated on 12 held-out oncological PET-CT studies across liver, spleen, and bladder, our zero-shot refinement consistently outperforms TS, improving Dice from 0.744 to 0.835 for bladder, 0.936 to 0.941 for liver, and 0.870 to 0.879 for spleen. Our SAM3 features are further incorporated as auxiliary inputs to supervised training of nnUNet with very limited true labels. The resulting PET+TS+SAM3 configuration achieves Dice of 0.929 (bladder), 0.945 (liver), and 0.886 (spleen) with only eight annotated cases. These results demonstrate that our method effectively corrects PET–CT mismatch and offers a label-efficient, annotation-free pathway towards generalizable PET-CT multi-organ segmentation.

Abstract

Convolutional neural networks and vision transformers are increasingly applied to medical image classification, yet their comparative decision-making behavior under equivalent aggregate performance remains poorly understood. We trained a convolutional neural network equipped with convolutional block attention modules and a vision transformer from scratch to predict major adverse cardiac events from nitrogen-13 ammonia positron emission tomography myocardial perfusion polar maps in a cohort of 1,079 patients. Across 100 independent training runs, both architectures achieved statistically indistinguishable area under the receiver operating characteristic curve (both 0.90; Wilcoxon signed-rank p = 0.897), yet discordance analysis revealed significantly different binary predictions on specific patients (McNemar p = 0.0013). Using architecture-appropriate explainability methods — gradient-weighted class activation mapping (the improved variant) for the convolutional network and attention rollout for the vision transformer — we propose a mechanistic hypothesis that the transformer's global self-attention integrates large-territory perfusion defects that the convolutional network's local spatial attention fails to aggregate. These findings demonstrate that equivalence in area under the curve does not imply decision equivalence, and motivate consideration of decision-level discordance when comparing architectures for clinical deployment.

Abstract

Accurately quantifying radiotracer uptake and metabolic activity in a subject using positron emission tomography (PET) requires correcting for photon attenuation. An attenuation map, typically produced using a second imaging modality, is required for attenuation correction but can increase radiation exposure, operational costs, and reconstruction complexity. To mitigate these drawbacks, methods for generating attenuation maps from the non-attenuation-corrected (NAC) PET image using deep neural networks were proposed. Preclinical PET images of small animals present additional challenges for artificial attenuation map generation because the attenuation is low and the subjects are not rigid bodies, making anatomical structures less consistent and more prone to artifacts. A conditional Denoising Diffusion Probabilistic Model (cDDPM) was implemented to generate attenuation maps from NAC PET images. The model was preliminarily trained on approximately 60 paired NAC PET and CT images acquired on the LabPET8 scanner. The dataset consisted of mouse whole-body and body-without-head scans obtained with two radiotracers: ¹⁸F-FDG, and ¹¹C-acetate. Preliminary slice-level evaluation showed that the generated pseudo-CT images were visually consistent with the reference CT images and preserved the main anatomical structures in the evaluated regions. The generated pseudo-CT images achieved a structural similarity index measure (SSIM) of 0.991, a mean squared error (MSE) of 4.47×10^-5 and a mean absolute error (MAE) of 8.51. These preliminary results suggest the feasibility of cDDPM-based pseudo-CT generation for preclinical PET attenuation correction, although full-volume validation and PET reconstruction studies using the generated attenuation maps are still required. Full results for more tracers and scanners will be presented at the conference.

Abstract

Mechanical collimators enforce a strict resolution-sensitivity trade-off, necessitate hardware exchanges for wide-energy imaging, and suffer from poor high-energy performance due to septal penetration. We propose Joint Self-Collimation Compton (JSCC) SPECT as a wide-energy solution using a multi-layer stochastic mosaic detector. JSCC integrates geometric constraints with Compton kinematics to screen pathways and suppress artifacts. Cylindrical-coordinate-based fast modeling accelerates volumetric reconstruction. Simulations using Hot-rod, Contrast, and Hoffman phantoms demonstrate that JSCC achieves a significantly superior signal-to-noise ratio compared to parallel-hole SPECT. At 511 keV, JSCC resolves 16-mm and 10-mm rods in 2-minute and 20-minute scans, outperforming the baseline (22-mm and 16-mm). At 140 keV, it resolves 9-mm rods in 30 seconds without hardware modification. These results demonstrate the potential of JSCC as a wide-energy, high-SNR SPECT imaging paradigm for next-generation molecular imaging and theranostic applications.

Abstract

The growing integration of robotics in medical imaging applications [1], the maturation of high-resolution room temperature semiconductor detectors, and the push towards safe, cost-effective and easy-to-use imaging platforms [2], motivates the development of a robotic SPECT system designed to overcome the inherent limitations of traditional platforms. The high flexibility of the robot enables acquisition trajectories not attainable with standard gantry-based systems, practically implementing the concept of adaptive SPECT [3].

We propose a system based on a collaborative 7-degrees-of-freedom robot, a CZT detector panel, a custom-designed multi-pinhole tungsten collimator, a LiDAR sensor for environment awareness, and a 3D structured light camera for object localization and occupancy.

Inverse kinematics and Bidirectional Rapidly exploring Random Trees are used for determining the robot configurations and collision-free movement. Energy calibration of the detector shows a FWHM of 4.94 keV at 122 keV, and an energy range up to 380 keV, suitable for imaging of most clinical and theranostics radiotracers. Integration of a LiDAR sensor and a 3D structured light camera shows high precision in object detection (AP_0.5>0.90) and point cloud stitching via Eye-in-Hand calibration.

Noisy Jaszczak phantom simulations show that robot-enabled spherical acquisition outperforms standard cylindrical scanning, yielding better metrics at the minimum NRMSE iteration: NRMSE improved by 18%, the peak-to-valley ratio for the 6.4 mm rods increased of 55.3%, and the average contrast-to-noise ratio and contrast recovery coefficient increased by 5.1% and 5.3%, respectively, with maximum gains over 13% for the smallest sphere.

Abstract

For patients undergoing cardiotoxic cancer therapies, baseline cardiac function is assessed using echocardiography to calculate left ventricle ejection fractions (LVEFs). At many sites, this requires additional patient visits and could delay treatment as use of these therapies increases. We investigate whether LVEF can be automatically derived from routine ¹⁸F-fluorodeoxyglucose (FDG) PET/CT scans acquired on a high-sensitivity Omni Legend scanner. Our retrospective study of FDG PET lymphoma scans (n=437) was organised into three complementary cohorts addressing gating validity, preliminary clinical comparison, and real-world applicability. In 5 electrocardiogram(ECG)-gated PET/CT scans, algorithm-generated triggers showed strong agreement with ECG triggers, with standard deviation trigger shift < 0.03. Of patients with ground truth echocardiograms (n=5), algorithm-derived LVEFs showed good agreement with echocardiography (mean absolute error = 5.1%), within the clinically accepted range for echocardiographic reporting variability (10%). Across the full dataset, sufficient myocardial uptake for heart localization and edge detection was observed in 153/437 cases (35.0%), while 207/437 (47.4%) lacked sufficient uptake despite successful algorithm execution. This cohort data suggests the potential for automatic LVEF estimation for one third of patients. Future work will investigate pipeline improvements such as refining the basal plane localisation to enable a LVEF estimation for a greater proportion of patients. Overall, these preliminary results indicate the potential for FDG PET derived LVEF with good ECG gating agreement and acceptable early agreement against echocardiographic reference values. Ongoing analysis of ground truth EFs will further evaluate classification of abnormal LVEF (< 50%) and establish myocardial uptake thresholds to guide clinical applicability.

Abstract

Compton scatter of gamma-rays within detector elements, denoted as inter-crystal scatter events, has been considered a resource for sensitivity improvement in PET imaging. The obstacle for their exploitation has been the resolution and contrast degradation due to inclusion of these events, related to the lack of knowledge of the first interaction position. We studied inter-crystal scatter events experimentally using a modular system based on single-end readout modules, also able to perform Compton polarimetry. Two modules comprising 16 x 16 GAGG:Ce scintillator matrices read out by silicon photomultipliers were mounted on a rotating gantry with 430 mm diameter. Two types of events were selected: (i) with a single pixel firing in one module and a two-pixel hit with forward Compton kinematics in the opposite module, and (ii) two-pixel Compton events in both modules with a pronounced polarization correlation. We imaged Ge-68 rod sources, and a NEMA NU-4 phantom with a Ga-68 solution. They results show that the statistics increases by 18 % using (i) and by an extra 8 % when adding (ii), compared to pure single-pixel events. At the same time, the inclusion of the selected inter inter-crystal scatter events has a marginal effect on image resolution (central field-of-view), shifting it from 2.5 +/- 0.1 mm (FWHM) for the single pixel events, to 2.7 +/- 0.2 mm (FWHM) for the combined sample, while the signal-to-noise and contrast-to-noise values are moderately increased. This validates the use of the inter-crystal scatter events for improved sensitivity and image quality in PET.

Abstract

Minimally-invasive radio-guided surgery requires simultaneous functional nuclear guidance and anatomical visualization to enable precise intraoperative localization of radiolabeled targets [1]. We propose a dual-modality imaging drop-in capsule designed for deployment through a standard 15-mm laparoscopic trocar. It consists of a 2-mm thick semiconductor detector with a sub-millimeter pixel pitch [2,3], a custom lofthole collimator embedded within a 3D-printed tungsten capsule, and two miniaturized stereoscopic micro-cameras.

Preliminary results show: real-time 3D depth maps validated against an ultrasonic distance sensor; a seamless 360° panoramic view stitched from sequential frames using cylindrical warping and multi-band Laplacian blending [4,5]; spectral localization of multiple isotopes ⁵⁷Co and ²⁴¹Am within a scattering biological medium; and a preliminary 2D co-registration overlaying the functional dataset onto the anatomical scene. We will present the system design and components, and discuss the steps towards a fully integrated prototype for laparoscopic radio-guided surgery.

Abstract

Lutetium-177 (¹⁷⁷Lu)  is one of the most representative radionuclides in integrated nuclear medicine theranostics. To image ¹⁷⁷Lu, we have developed a full-ring, high-sensitivity cascade gamma photon coincidence imaging system. This work presents the system construction and preliminary performance evaluation.

The detector ring consists of 40 continuous LaBr₃ crystal modules arranged in 10 radial × 4 axial configuration (each 50 × 50 × 6 mm³). The inner diameter of the ring is 110 mm, with axial module spacing of 60 mm. Each crystal is coupled to four MPPC arrays and read out via PETsys FEM256 electronics, with three FEB/D boards and one DAQ for data acquisition. Both the ASICs and MPPCs are cooled by a water-cooling system, maintaining water at 8 °C and ASIC temperature below 20 °C.

The collimator assembly comprises 10 tungsten plates (26 × 128 × 8 mm³), each covering four axial detector modules. Plates are arranged alternately with pinholes and slits (1 mm), placed at a vertical distance of 36.5 mm from the ring center, creating an imaging field of Φ30 × 30 mm. The prototype hardware has been fully assembled and is capable of normal data acquisition.

Simulation results indicate a coincidence sensitivity of 2.97 × 10⁻⁵ at the central field of view, spatial resolution of ~1.7 mm, and hot cylindrical object resolution up to 1.2 mm. Preliminary experimental tests have validated the feasibility of the system. Ongoing optimization includes crystal position calibration, energy resolution improvement, and collimator geometry alignment. The full system performance evaluation is expected to be completed before the upcoming conference.

Abstract

Positron emission tomography and single photon computed emission tomography are widely used nuclear medicine modalities. These modalities provide information on physiological and metabolic functions by visualizing radiotracer distributions. However, they cannot obtain local information, such as intermolecular interactions, pH, and other microenvironmental factors. We have proposed and demonstrated a novel nuclear medicine imaging and quantum sensing method utilizing the characteristics of cascade radionuclides. For imaging, the coincidence detection of cascade gamma rays enables localization of the source position using directional detectors. For sensing, changes in angular correlation between cascade gamma rays provide local information, such as pH [1]. In our previous studies, angular distribution information was used as an indicator of angular correlation changes because the time resolution of measurement system of approximately 50 ns was insufficient for time-differential perturbed angular correlation analysis. In this study, we developed the improved time resolution of the measurement system using GFAG pixel scintillator detectors, originally designed 64-channel slew rate limited time-over-threshold integrated application specific intesgrated circuits, [2] and 62.5 ps sampling data acquisition systems [3]. The time-dependent perturbation factor exhibited pH-dependent decay behavior consistent with a previous study [4], with slower decay at low pH and faster decay at higher pH. In addition, ¹¹¹In-labeled OctreoScan showed different after-effect behavior from aqueous ¹¹¹InCl₃ solutions, suggesting that time-differential analysis provides information beyond integrated angular distribution measurements. We will incorporate the time-differential perturbed angular correlation analysis into double-photon coincidence imaging and investigate its potential for nuclear medicine application.

Abstract

In matter, a positron annihilates with an electron into two photons through several mechanisms, including direct annihilation and the formation of positronium, a metastable bound state that exists as either para-positronium or ortho-positronium. Each channel has a characteristic lifetime, allowing their contributions to be distinguished in time. The two photons are predicted to be entangled in their polarizations [1], and recent measurements indicate that the degree of entanglement depends on the annihilation channel [2]. In particular, the ortho-positronium decay is sensitive to the local concentration of molecular oxygen, and can serve as a probe of tissue hypoxia in PET imaging [3]. The Jagiellonian Positron Emission Tomograph is a scanner built from plastic scintillators, in which 511 keV photons interact predominantly through Compton scattering [4]. The Compton scattering plane of each photon is correlated with its polarization. We measure annihilations of positrons emitted by a sodium-22 source embedded in porous polymer. Sodium-22 emits a positron accompanied by a 1274 keV prompt photon, which provides a time reference for the positron lifetime. The polarization correlation distribution of the annihilation photon pairs is consistent with quantum entanglement, between the limits for maximally entangled and classically separable states [2]. Selecting events by positron lifetime reveals distinct polarization correlation distributions for different annihilation channels. A detailed comparison will be presented at the conference. These results motivate quantum-informed PET imaging, with tissue hypoxia as a clinical biomarker.

Abstract

This research presents the development and first in-vivo demonstration of Positronium Ratio Imaging, a novel functional imaging technique that uses the three-photon to two-photon decay ratio of positronium. While conventional Positron Emission Tomography primarily maps radiopharmaceutical distribution, Positronium Ratio Imaging provides insights into the tissue microenvironment by exploiting the sensitivity of ortho-positronium to local oxygen levels and morphology. We characterized the system's geometric stability using a four-rod phantom with varying dimensions and evaluated the results against Monte Carlo simulations using the Geant4 toolkit. The phantom study identified a size-dependent uncertainty baseline of approximately twenty percent, primarily arising from detector sensitivity and geometric factors. Furthermore, we achieved the first in-vivo ratio imaging using a tumor-bearing mouse model administered with fluorine-18 fluorodeoxyglucose. Quantitative analysis revealed that the three-photon to two-photon ratio in brown adipose tissue was approximately two and a half times higher than in the heart and bladder. This significant contrast is consistent with theoretical trends derived from tissue-specific positronium lifetimes and far exceeds the established hardware-induced uncertainty. These findings validate the feasibility of using standard positron emitters for functional three-photon imaging without the need for specialized prompt-gamma emitters. Future work will focus on refining quantitative calculation methods, evaluating signal-to-noise ratios, and implementing hardware-specific calibrations to enhance absolute quantification. This technology lays a foundation for clinical applications such as the non-invasive visualization of tumor hypoxia and metabolic characterization of the tissue microenvironment, offering a breakthrough in functional molecular imaging.

Abstract

Positronium annihilation lifetime spectroscopy (PALS) has recently attracted considerable interest as a potentially sensitive tool for tissue characterization. In standard PALS experimental setups, at least two detectors are needed to provide start and stop signals for positron emission and annihilation creation. However, due to the isotropic nature of prompt-gamma emission, in a very significant fraction of events, both prompt and annihilation photons are detected in the same detector, resulting in a pileup event. In this work, we propose using these pile-up events to simplify o-Ps lifetime estimation workflows,
enabling estimation with a single detector. Using a 22Na source sample, we successfully measured the o-Ps lifetime using both standard coincidence PALS and the proposed approach with excellent agreement (1.88±0.04 and 1.90±0.07 ns). Measurements of o-Ps lifetime in an isopropanol-based liquid using 44Sc, as well as in ex-vivo biological samples from mice (liver and plasma) after a PET acquisition using 52Mn, were also obtained. These results demonstrate the feasibility of single-detector o-Ps measurement and open new lines of research, by facilitating the characterization of samples from PET studies.

Abstract

We present quantitative 4D PET imaging of photosynthate carbon transport and rhizosphere exudation in intact Phaseolus vulgaris plant–soil systems using the R-PET (Rhizo-PET) scanner with the ¹¹C radiotracer (t₁/₂ = 20.3 min). Across N = 6 plants, time–activity curves (TACs) extracted from 225 cross-pattern ROIs per dataset—spanning 3 roots × 3 axial depths × ±3 mm radial extent—reveal a reproducible source-to-sink carbon allocation hierarchy: upper root zones receive ≈50% of total plant carbon, while root 2 (primary taproot) dominates at 20.8 ± 6.3%. Time-to-peak (TTP) increases from 31.5±4.4 min (upper) to 48.0±13.1 min (lower root 2), corresponding to a phloem transport velocity of ≈2.4 mm min⁻¹. Rhizosphere carbon exudation follows an exponential decay AUC(r) = A e⁻ᵏʳ with k = 0.10–0.20 mm⁻¹, placing the active deposition zone within 4–14 mm of the root surface—the first spatially resolved 3D+t measurement of this process in intact soil.

Abstract

Positron-labeled tracers are powerful tools for studying diseases and developing new therapeutics, yet conventional radiochemistry platforms are complex, costly, and low-throughput, making PET tracer development slow and resource-intensive. To address this bottleneck, we are constructing a high-throughput radiochemistry system that will enable multiple radiosyntheses to be performed in parallel. A key challenge is accurately quantifying radioactivity in each microliter-scale droplet throughout the radiosynthesis process. While plastic scintillators and an overhead CMOS camera offer a compact detection approach, the scintillation signal is confounded by variations in droplet volume and geometry. Geometry correction is therefore critical for reliable activity quantification.

We investigated an optical approach to infer droplet geometry by leveraging its optical reflective and refractive properties. Droplets were illuminated with a ring light mounted around the camera lens, and we sought to infer their geometries from the resulting optical features. Experimental images were acquired from evaporating water droplets on a Teflon-coated silicon chip, and ray-tracing simulations were performed on droplet models reconstructed from side-profile images. Both revealed two distinct features from independent optical processes: a mirror reflection effect and a lensing magnification effect. These varied systematically with droplet volume, suggesting they encode geometry information and may provide complementary information for inference.

These preliminary findings support the feasibility of this optical method for geometry estimation and radioactivity quantification correction. The approach adds minimal complexity to the existing system and is well-suited for integration into a compact radiochemistry robot. Future work will focus on refining droplet shape models, extending the method to off-axis droplets, and evaluating image-analysis approaches for robust geometry inference.