Online program Overview Session: MIC-05

Abstract

A Variable Angle Slant-Hole (VASH) collimator has been developed for integration with the SmartBreast’s Eve ClearScan e750 Molecular Breast Imaging (MBI) system, formerly known as the GE Discovery NM750b. This effort builds upon prior success implementing VASH technology in the Dilon 6800 camera platform, where slanted collimator geometries demonstrated improved resolution and lesion detectability in breast imaging applications. The SmartBreast MBI system incor-porates high-resolution cadmium zinc telluride (CZT) detectors, replacing earlier NaI(Tl)-based designs, and thus necessitates a redesigned collimator geometry to fully leverage the intrinsic resolution and detector compactness. The new VASH design incorporates a reduced hole pitch of 1.84 mm, fixed septal thickness of 0.4 mm, and a shortened collimator depth of 13 mm, in contrast to the original lead-based parallel-hole collimator (2.46 mm pitch, 0.4 mm septa, 21 mm depth). These parameters were selected to balance sensitivity and spatial resolution. Ana-lytical calculations indicate that the redesigned VASH achieves comparable or improved resolution and sensitivity performance at clinically relevant object-to-collimator distances up to 60 mm. The first prototype is under construction, with mechanical inte-gration, phantom imaging, and angular performance evaluation scheduled for upcoming testing. The VASH system offers imaging flexibility in clinical MBI and has potential implications for improving detection of lesions at variable depths in dense breast tissue.

Abstract

The performance of gamma detectors for time-of-flight positron emission tomography (TOF-PET) has improved significantly over the last decade. With commercial systems surpassing the 200~ps coincidence time resolution (CTR) barrier, the avenues for improving TOF-PET detector performance become progressively more limited. We report on CHLOE (Cherenkov Light Module), a PET detector readout with high channel density composed of dual ended readout bismuth germanate (BGO) crystals coupled one-to-one to silicon photomultipliers (SiPMs) with the goal of providing a trade-off between 3D detector segmentation, energy resolution, and scalability in an intrinsic radiation-free crystal. The electronics were based on the FastIC+ chip and the complete readout chain, from the front-end to the back-end was customized for this concept.

This work presents the first electronic and detector characterization of the CHLOE readout approach. The noise and dark counts contributions of the system are presented, including measurements of channel dependencies in the response. Energy spectra were obtained for $^{22}$Na and $^{133}$Ba sources. The work presented in this contribution represents the very first steps towards a full functional characterization of the CHLOE readout.

Abstract

Multi-tracer preclinical single-photon emission computed tomography (SPECT) enables simultaneous tracking of multiple physiological processes inside organs, requiring detectors with high spatial and energy resolution over a broad energy range. In this study, we developed and evaluated a high-resolution detector comprising a 28 × 28 gadolinium aluminum gallium garnet (GAGG) array (0.32 mm pitch, 2 mm thickness) coupled to a linearly graded silicon photomultiplier (LG-SiPM) using radiation sources with different energies: 55, 122, 202, 307, 511 keV. Flood histograms demonstrated clear separation of crystal elements over a wide energy range from 55 keV to 511 keV. Furthermore, the detector achieved energy resolutions of 18% at 122 keV and 12% at 511 keV, along with excellent energy linearity across 55–511 keV (R² = 0.996). These results demonstrate that the developed detector is a promising candidate for high-resolution small-animal SPECT systems.

Abstract

This study presents a sub-millimeter resolution small-animal PET detector based on a dual-ended readout configuration with highly multiplexed signal processing, targeting applications in compact and MRI-compatible PET insert systems. The detector employs a LYSO crystal array coupled to silicon photomultiplier (SiPM) arrays on both ends. A row-column summation multiplexing scheme is adopted to compress multi-channel SiPM signals into a limited number of position-encoding outputs, while a dedicated multi-channel ASIC (MPT2321) is utilized for signal amplification, timing extraction, and digitization. The multiplexed signals are directly processed by the ASIC without requiring individual channel readout, significantly reducing the number of electronic channels and system complexity, which is particularly beneficial for integration in high-field MRI environments. Experimental results demonstrate that all crystal elements in the 26 × 26 LYSO array can be clearly resolved in the flood histogram. The detector achieves an average energy resolution of 15.3 ± 0.7%, a DOI resolution of 2.56 ± 0.46 mm, and a detector timing resolution (DTR) of 2.40 ± 0.20 ns. The proposed detector design demonstrates that highly multiplexed readout combined with ASIC-based signal processing can effectively reduce system complexity and cost while maintaining good spatial, energy, and DOI performance, making it suitable for high-resolution small-animal PET imaging and compact PET/MRI insert applications.

Abstract

We evaluate the spatial decoding capability and energy resolution of a position sensitive SiPM (PS-SiPM, model PSS 11-6060-S, Capital Photonics Technology, China) with an active area of 6.24 mm × 6.24 mm. The PS-SiPM is optically coupled to a checkerboard-pattern scintillator array, which is composed of alternating LYSO and BGO crystals with a 0.38 mm crystal size and a 0.46 mm pitch, arranged in a 28 × 28 array. To compensate for the lower scintillation light yield of BGO, a custom amplifier board with a nominal gain of 16 was incorporated into the readout chain.

Delayed charge integration (DCI) technique was applied for distinguishing LYSO and BGO events based on pulse shape. The average rise and decay time constants were measured 31 ns and 313 ns for BGO, and 22 ns and 60 ns for LYSO, respectively. Using an Anger-logic positioning algorithm, the resulting flood image successfully resolved a 13 × 13 sub-region of the crystal array, with minimal edge decoding difficulty. It highlights the advantage of the PS-SiPM in identifying edge crystals.

Energy resolution was evaluated within a central 9 × 10 crystal pixels to mitigate intrinsic geometric constraints. The 45 LYSO crystals in this region achieved an average energy resolution of 18.52% ± 2.39%, and the 45 interleaved BGO crystals achieved 25.54% ± 2.48%. These results confirm the feasibility of using a PS-SiPM for decoding sub-millimeter crystal array, and provide a performance reference for the implementation of PS-SiPMs in ultra-high-resolution nuclear imaging detectors.

Abstract

 We present a physics-constrained joint deconvolution framework for reconstructing photon arrival dynamics. In scintillator-based detectors, measured waveforms arise from the convolutional superposition of stochastic photon detection events, while finite analog signal shaping limits the observable temporal resolution. Mathematical deconvolution approaches implicitly assume continuous photon arrival recovery; however, such formulations are fundamentally ill-posed under band-limited measurements. In this work, we instead represent the latent photon arrival process on a discretized temporal grid as a sampled intensity. Dual-channel timing and energy waveforms are modeled as complementary observations of a shared latent process and jointly inverted using sparsity-constrained optimization. The proposed framework stabilized reconstruction in the presence of bipolar timing responses and temporal smearing, yielding physically consistent photon-train estimates. Joint modeling mitigates ambiguities inherent to single-channel reconstruction and improves robustness in timing extraction under degraded signal conditions. These results establish photon-train reconstruction as a physically grounded representation of detector response and a principled framework for interpreting temporally overlapping signals.

Abstract

Over the past years, significant advances have been achieved in Time-of-Flight Positron Emission Tomography (ToF-PET) systems, reaching coincidence time resolution (CTR) values below 100 ps. However, this performance has been demonstrated in small-scale detectors using short scintillation crystals, which favor timing performance at the expense of detection efficiency, and high-power electronics. In contrast, realistic system designs require scintillator crystals with thicknesses of >=10 mm to ensure high detection efficiency. Pushing the CTR limits in large crystals remains challenging due to optical photon transport effects and depth-of-interaction (DOI) uncertainties. 

For this purpose, we investigate a dual-ended read out detector based on a 3x3x20 mm³ LYSO:Ce,Ca crystal against a 2x2x3 mm³ Fast-LGSO crystal used as a timing reference. The read out was performed with the FastIC+, a custom application-specific integrated circuit (ASIC) multi-channel front-end with on-chip digitization. This ASIC is specifically design for fast-timing applications and optimized for the read out of high-gain photodetectors such as silicon photomultipliers (SiPMs).

In this work, we apply an energy-based DOI correction method to improve the CTR performance of thick-crystal detectors. Using a dual-ended detector based on a 3x3x20 mm³ LYSO:Ce,Ca crystal coupled to 4x4 mm² SiPMs and measured against a 2x2x3 mm³ Fast-LGSO reference detector coupled to a 3x3 mm² SiPM, we achieve a CTR of (135 +/- 1) ps without DOI correction. After applying the correction, the CTR improves down to (110 +/- 1) ps. These results obtained for thick crystals opens a clear path toward sub-100 ps performance in realistic scanner configurations.

Abstract

Depth-of-interaction (DOI) estimation is a bottleneck in next-generation PET, contributing to parallax error and degraded timing resolution. Existing solutions using double-sided readout, phoswich, monolithic, or layered crystals require substantial modifications to the readout or detector stack. As part of the development of the SymPET waveform-digitizing readout ASIC, we investigate whether photon arrival-time waveforms alone carry sufficient information to recover DOI without auxiliary sensing. Using DETECT2000 Monte Carlo simulations of 2×2 mm LYSO rods coupled one-to-one to SiPMs, we compare polished and unpolished surface treatments at 12 and 20 mm lengths and train a hybrid 1D CNN-Transformer regressor on 50 ps-binned photon arrival-time histograms. Unpolished surfaces yield lower per-Z MAE and residual std than polished. Bundled high-SNR events reach 0.18 mm worst-per-Z MAE, while single-event MAE is 2.53 mm. Showing the information is present in the waveform, closing this SNR gap is the open problem. (Preliminary results.)

Abstract

This study is aimed at the development of a dynamic micro imaging technique based on nuclear medicine methods. Single-cell resolution imaging is strongly required to study the properties and mechanisms of cells for improved efficiency of drug development. Conventional nuclear medicine imaging techniques are not suitable for this purpose due to their limited spatial resolution. Autoradiography works in terms of spatial resolution, but it cannot obtain dynamic images. Therefore, we are developing a micro radioisotope (RI) imaging system to realize dynamic cell observation. This system consists of a CCD camera and a scintillator plate. Instead of the eutectic scintillator that we introduced in the last IEEE MIC conference, which had non-uniformity and insensitive region issues for this purpose, in this study we use an optical-guiding crystal scintillator (OCS) which has a 10 μm fiber structure.

Abstract

The S-PET project is developing a high-resolution positron emission tomography detector module intended for simultaneous brain positron emission tomography / magnetic resonance imaging. The detector needs to feature good position resolution including depth-of-interaction capabilities as well as good timing resolution. A core challenge is the limited space for electronics. The detector integrates 12×12 or even 16×12 SiPMs with one 64-channel TOFPET2 ASIC, necessitating efficient channel compression schemes without additional elements that leverage light-sharing for event positioning. Crucial optimization parameters for detector performance include the crystal topology and a matching compression scheme. This work evaluates Time-of-Flight and Depth-of-Interaction performance across multiple compression schemes and crystal topologies: continuous slabs, pseudoslabs, and pixelated arrays with a light guide coupled to the face opposite the photosensitive detector. This study utilizes modular development boards, that allow for quick swapping of the compression scheme without reattaching the crystal array, and the first full-scale sensor tile with one of these compression schemes, the Hyperion Visor HD. Preliminary results with a 48×48 mm² BrainPET 7T crystal array on the latter show planar resolutions with Mean Absolute Error below 1.5 mm. First tests without compression schemes on a development board with 16×16 mm² crystal arrays (2 mm and 4 mm pixel size) with light guide on top show good timing (Coincidence Time Resolution of 235 ps) and Depth-of-Interaction resolution (Mean Absolute Error <1.5 mm). While these preliminary results are promising, the comprehensive evaluations utilizing several crystal topologies combined with the different compression schemes will be presented at the conference.

Abstract

Radioembolization is widely used for treating cancer, e.g. in the liver. The procedure currently consists of two sessions, one for imaging and one for treatment, typically several weeks apart. Cost-effectiveness, accuracy of treatment, and patient experience could all be improved if both steps could be combined into a single interventional session. A multimodal detector that integrates an X-ray flat panel with a γ-camera would make this possible, by capturing anatomical information and dose distribution at the same time. During X-ray imaging, however, a portion of the radiation passes through the flat panel and blinds the γ-camera. In principle, γ-photons could be detected in between X-ray pulses, but the high afterglow of the currently used NaI(Tl) scintillator interferes with the γ-photon signals. To overcome this, we explore the use of a low-afterglow scintillator in a pixelated γ-camera design. Reducing afterglow facilitates γ-detection shortly after X-ray exposure. Pixelation further increases the γ-photon signal intensity relative to the afterglow background. The pixels are coupled one-to-one to silicon photomultipliers (SiPM). Here, we characterize a prototype module based on NaI(Tl,Sr,Y), which we previously showed to have a low afterglow of 500ppm at 100ms. The module is hermetically sealed to protect the hygroscopic scintillator and comprises 6 x 6 crystals of 3.75 x 3.75mm x 9mm, separated by 0.25mm thick Teflon reflectors. The central 4 x 4 pixels are coupled to a 4 x 4 SiPM array. The 20 inactive outer pixels ensure equal inter-crystal scatter conditions for all active pixels. Event positioning is based on a winner-takes-all algorithm. The performance of the prototype detector is evaluated by measuring Am-241, Ce-141 and Co-57 pulse-height spectra of all active pixels, with and without simultaneous exposure to a pulsed X-ray source.

Abstract

The dominant cost driver in total-body positron emission tomography (TBPET) systems is the scintillator material. Current commercial scanners rely on LYSO crystals, which are costly and depend on geographically concentrated rare-earth supply chains. CRYSP proposes an alternative TBPET design based on large monolithic pure cesium iodide (CsI) crystals operated at cryogenic temperatures (~100 K). At these temperatures, CsI achieves a light yield of ~10⁵ photons/MeV, yielding an energy resolution below 7% at 511 keV. The monolithic geometry provides inherent depth-of-interaction (DOI) capability, enabling millimeter-scale spatial resolution in all three dimensions and minimizing parallax error. These properties compensate for the lower sensitivity and absence of time-of-flight information, allowing image quality competitive with state-of-the-art TBPET scanners at substantially reduced cost. The slow decay time of cryogenic CsI (~800 ns) renders conventional readout electronics unsuitable; a front-end based on the CAEN CITIROC ASIC, providing pulse-height analysis well suited for slow scintillators, was adopted instead for the prototype. Event position reconstruction employs both analytic methods and a deep-learning algorithm trained on Geant4 simulations, achieving millimeter-scale resolution including DOI in Monte Carlo studies. As a first proof-of-concept, a small-scale PET scanner comprising 18 monolithic CsI(Tl) crystals operating at room temperature is under development, providing hands-on experience with the full acquisition and reconstruction chain before advancing to a cryogenic prototype.

Abstract

The precision to which the arrival time of 511 keV photons are able to be known directly affects the signal-to-noise-ratio (SNR) for reconstructed images in time-of-flight positron emission tomography (TOF-PET). Currently, significant research is ongoing to push the full-width-half-maximum (FWHM) coincidence time resolution (CTR) for TOF-PET systems well below 100 ps towards the few-10’s-of-ps regime. In this work, we designed an application specific integrated circuit (ASIC) in a commercial 65nm CMOS process to translate ultra-fast timing performance currently demonstrated in state-of-the-art bench-top measurements into a fully integrated scalable topology for large scale time-of-flight positron emission tomography (TOF- PET) systems. We targeted a scalable, low-noise, ultra-fast silicon photomultiplier (SiPM) analog front-end that maintains excellent SNR for single photon response while achieving ≤35 ps FWHM electronic noise jitter for single photon detection with 4x4 mm² Broadcom AFBR-S4N44P014M Near-ultraviolet metal-trench SiPMs. This “analog” chip uses low-voltage differential signaling (LVDS) transmitters to drive discriminator data to an external time-to-digital converter (TDC). We present evaluation of achievable single photon time resolution and CTR with the first prototype test chip.

Abstract

We present a high-resolution PET detector module with fine-pitch crystals and dual-ended readout for small-animal imaging. The detector employs a 28 × 28 LYSO crystal array with a pixel size of 0.75 mm × 0.75 mm × 15 mm, separated by 0.08 mm BaSO₄ reflector layers and read out at both ends by SiPM arrays. Experimental results demonstrated clear crystal separation with a corrected depth-of-interaction (DOI) resolution of 1.59 mm (FWHM) and an energy resolution of 14.9% in the selected 1/4 detector region. These results indicate that dual-ended readout can effectively improve DOI resolution in fine-pitch crystal arrays while preserving crystal separation, demonstrating the feasibility of fine-pitch dual-ended PET detectors for high-resolution small-animal PET imaging.

Abstract

Mechanical collimators in SPECT introduce a fundamental sensitivity-resolution trade-off that becomes critical in targeted radionuclide therapy (TRT) imaging, where administered activities are low. We explore a collimator-free detector concept based on an 18x18 pixelated GAGG scintillator array in which collimation is achieved through staggering, that is, height offsets between neighboring pixels that produce characteristic shadow patterns in the resulting flood map. Three configurations are evaluated: a flat array (stagger-0), a single-pixel checkerboard (stagger-1), and a 3x3-block checkerboard (stagger-3). For each geometry, OpenGATE 10 Monte Carlo simulations of a 140 keV point source were performed at 400 positions on a fixed-height plane. The resulting flood maps were used to train a 2D convolutional neural network that regresses the source coordinates. Evaluated on 40 test positions, the 3D root-mean-square error of the neural network decreases with increasing staggering degree, from 0.44 mm (stagger-0) to 0.40 mm (stagger-1) and 0.34 mm (stagger-3), a 23% reduction that supports staggering as a promising strategy for collimator-free source localization.

Abstract

A fine-pitch CZT detector-head prototype was assembled and experimentally characterized for SPECT imaging. The head consists of a 5 × 10 tiled array of CZT modules, each containing 16 × 16 anodes with 1.45 mm pixels and a 1.6 mm pitch. The prototype achieved a global energy resolution of 4.8% at 140 keV. The intrinsic integral and differential uniformities were 0.56% and 0.44%, respectively. To exploit the fine intrinsic spatial sampling of the CZT detector, a fast ray-tracing Monte Carlo framework was used to explore dual-layer 1H/4HMPD matched-pixel collimator candidates and to compare them with the 1HL30.5 reference design, defined as a one-hole-per-pixel collimator with a 30.5 mm hole length, under prototype-constrained geometric and mechanical conditions. A representative dual-layer candidate showed an alternative main-lobe sensitivity–resolution trade-off, but GATE point-source simulations revealed more pronounced non-Gaussian tails than the 1HMPD reference. Therefore, the 1HL30.5 configuration was retained as the near-term practical collimator design. GATE-simulated Derenzo phantom reconstruction of this configuration showed that 5 mm hot rods were visually distinguishable at 10M and 5M total projection counts. These results support the feasibility of the fine-pitch CZT detector-head prototype and justify retaining 1HL30.5 for subsequent experimental phantom validation.

Abstract

This study investigates the influence of magnetic fields on F-18 and Ga-68 positrons and how they affect image quality metrics in tissue-equivalent phantoms relevant to PET/MR imaging. To this end, we performed Monte Carlo simulations using the PENELOPE code. Since the main PENELOPE routine penmain.f does not account for the influence of electromagnetic fields on radiation transport in matter, a modified version of the penfield.f module was used instead. Simulations were carried out in four representative media (lung, soft tissue, water, and cortical bone) and in NEMA-like image quality phantoms to assess the effect of uniform magnetic fields of 0, 3, and 9 T. We simulated positron point sources at the center of a 10 cm-radius spherical phantom of homogeneous materials and in a set of three image-quality phantoms to evaluate specific image-quality parameters: spill-over ratio, recovery coefficient, and lesion detectability, as in a previous experimental study. The results indicate that positron range significantly affects recovery coefficients and spill-over ratios, particularly for high-energy positron emitters in low-density tissues. In the recovery-coefficient phantom, when the cylinder separation distance is smaller than the positron range, a slight overestimation of the recovery coefficients is observed due to contributions from adjacent cylinders. The magnetic field reduces these effects in the direction perpendicular to the magnetic field lines, leading to improved spatial resolution and quantitative accuracy. These findings represent the maximum image quality that can be achieved under ideal conditions in a PET system in the presence of uniform magnetic fields and provide a physical basis for the development of positron-range correction methods in PET/MR imaging.

Abstract

To improve the detectability of small lesions and early metastases, we are developing a breast PET insert that operates in conjunction with a whole-body (WB) PET/MRI scanner. We implemented a joint reconstruction framework with attenuation correction that explicitly accounts for the different coincidence types from the insert and the WB-PET, aiming to improve spatial resolution inside the insert field of view (iFOV), as well as outside. Performance was evaluated using Monte Carlo simulations of an anthropomorphic phantom with lesions of 4, 8, and 12 mm diameter and contrast ratios of 4:1 and 8:1 relative to the breast background. Lesions were placed at clinically relevant locations, including within the breast, as well as in axillary and parasternal lymph nodes. Compared to WB-PET alone, the inclusion of all coincidence types improved lesion conspicuity, particularly for small (4–8 mm) and low-contrast lesions. Within the iFOV, breast lesions show increases in contrast recovery coefficent (CRC) of up to 16.4% (4 mm, 8:1) and 16.9% (8 mm, 4:1), with 4-mm lesions becoming resolvable. Similar trends were observed for axillary lymph nodes, with CRC increases of up to 11.9% (8 mm, 4:1). Future work will extend this study beyond the idealized 511 keV case to more realistic energy windows with finite energy resolution and higher scatter contribution, as well as incorporate spatially variant point spread function modeling and scatter correction into the reconstruction framework.

Abstract

Dedicated cardiac semiconductor SPECT systems enable short acquisition times owing to their high sensitivity; however, artifacts may occur depending on patient body habitus, and quantitative reconstruction images cannot be obtained because attenuation correction using X-ray CT images (μ-maps) is not available. In this study, we propose a novel SPECT system in which two gamma cameras equipped with NaI scintillators are arranged in an L-shaped configuration, allowing data acquisition while keeping the detectors stationary, and achieving attenuation, scatter, statistical noise and collimator aperture correction using a convolutional neural network (CNN). To improve sensitivity, 44 multi-pinhole collimators were employed, and projection images were acquired in an overlapped manner and accurately separated using the CNN. A U-Net++ architecture was adopted to construct a unified framework for overlap separation of projection data, statistical noise reduction, aperture correction and attenuation/scatter correction without requiring μ-maps. Image reconstruction was performed using the ML-EM algorithm, and image quality was evaluated using the normalized mean squared error (NMSE) of the polar map. As a result, approximately uniform image reconstruction was achieved without using μ-maps, and the NMSE was 0.0402, demonstrating the effectiveness of the proposed method.

Abstract

Accurate physics-based simulations are essential for assessing and validating scanner geometries and reconstruction algorithms. EduPET is an NPSS-supported educational PET technology demonstrator. This work introduces the Educational PET SIMulation system (EduPETSIM), an integrated web-based framework designed for educational purposes to streamline the entire PET workflow. The platform unites an OpenGATE-based simulation engine, a custom coincidence sorter, and an image reconstruction module within a unified environment. A primary novel contribution of this framework is the integration of a Conditional Wasserstein Generative Adversarial Network (cWGAN) to facilitate rapid source modeling, which significantly reduces the computational overhead associated with traditional particle transport simulation. Current developments focus on fine-tuning the cWGAN model for a ²²Na source within the two-detector EduPET system, with the completed model available for the upcoming IEEE NPSS schools on medical imaging.

Abstract

Modeling light scattering in biological media requires balancing the accuracy of Mie theory with the speed of simplified phase functions. We propose a new phase function that replaces the standard cos(θ) term with a quadratic θ² relationship, significantly improving large-angle fidelity compared to the Henyey-Greenstein (HG) model. A key advantage is the function's analytical invertibility, enabling direct Inverse Transform Sampling that accelerates Monte Carlo simulations by bypassing transcendental computations. Mathematical consistency is validated against the stochastic Rejection Method for various tissues. Transport simulations using the PhoSim package demonstrate enhanced spatial accuracy and closer alignment with the Mie ground truth, offering a high-speed framework for real-time IR tomography and Time-Resolved Optical Tomography.

Abstract

We present a GPU-accelerated system matrix computation method for multi-pinhole SPECT based on spatial modeling of position-dependent point spread functions (PSFs). Sparse Monte Carlo-generated PSFs are fitted with an anisotropic Gaussian model, and the resulting parameters are expressed as geometry-dependent functions to enable estimation of voxel-dependent responses at unmeasured locations. Implemented in CUDA for forward and backprojection, the method supports on-the-fly iterative reconstruction without storage of a fully precomputed system matrix. This approach was evaluated using Derenzo phantoms and NEMA image quality phantoms filled with isotopes whose emitted photon energies range from 140 to 440 keV, demonstrating its potential for efficient multi-isotope and theranostic SPECT imaging.

Abstract

This work presents a Monte Carlo based performance evaluation of a dual-panel PET&SPECT trimodal imaging system, in terms of spatial resolution, system sensitivity, and count rate performance following the NEMA NU 2-2018 standard. Spatial resolution measurements yield an average value of 1.01 mm for PET and 11.1 mm for SPECT. The SPECT sensitivity with a 700 mm length line source is 101.5 cps/MBq at the center of the field of view (axial offset z = 0 cm) and 103.0 cps/MBq at z = 10 cm. When using a 350 mm line source, SPECT sensitivity increases to 152.1 cps/MBq at z = 0 cm and 129.1 cps/MBq at z = 10 cm. The PET sensitivity measured with a 700 mm length line source achieves 43.4 cps/kBq at z = 0 cm and 43.5 cps/kBq at z = 10 cm. The PET peak noise equivalent count rate reaches 131.5 kcps at 193 MBq in standalone acquisition mode, and decreases under simultaneous PET&SPECT acquisition due to elevated dead time induced by SPECT detections. These results validate the overall system design, quantify the interference between PET and SPECT detections, and provide a foundation for further optimization toward clinical trimodal imaging applications.

Abstract

Advanced detector designs employing multi-layer crystal configurations and multiplexed readouts introduce complex light transport and inter-crystal scatter effects. These result in position-dependent differences in energy response and crystal identification that are difficult to reproduce using idealized simulations. We present an experimentally calibrated GATE simulation framework for a dual-layer depth-of-interaction capable detector developed for a dedicated breast PET/MRI insert. The detector block consists of a 10×10 bottom LSO crystal array (1.5×1.5×9.81 mm³) and a 9×9 top array (1.5×1.5×5.19 mm³). Simulations were performed using GATE v9.4.1 for two geometries and compared against experimental measurements. A reference configuration of 18 detector modules in a 100 mm ring was used to derive crystal-specific correction factors, then applied to the same 18 modules in a 178 mm ring representing the breast insert geometry. Measurements were acquired using a centered Na-22 point source for three energy windows (250–760 keV, 511±50 keV, and 430–610 keV) with experimentally informed coincidence timing resolutions. The uncorrected simulations systematically overestimated experimental coincidence count rates, with the largest deviations at edge and corner crystals of the bottom layer. Prior to calibration, mean absolute deviations for the 178 mm geometry were 37.1%, 38.0%, and 43.0%; following calibration, deviations were reduced to 9.3%, 8.4%, and 8.8%. Experimental detector blocks additionally showed substantially larger crystal-to-crystal variations in energy resolution (top: 12.5±1.1%, bottom: 11.0±1.2%) compared with the fixed simulated values, indicating complex optical transport and detector-specific effects are not yet fully reproduced in the current model.

Abstract

Accurate quantification in preclinical PET remains a fundamental challenge due to the spatial resolution constraints inherent to small-animal imaging, where partial volume effects compromise biodistribution and dosimetry estimations in monoclonal antibody-based radiopharmaceuticals. Here, we present the first integrated digital-twin framework for immuno-PET quantification in preclinical imaging developed around a validated Monte Carlo model of the Bruker (BioSpec 3T) PET/MRI scanner implemented using the SimPET platform — which integrates the SimSET Monte Carlo engine and STIR reconstruction software — and validated following the NEMA NU 4-2008 standard. This validated model was then combined with physical and digital rodent phantoms to assess a highly accurate immuno-PET quantification pipeline. Scanner validation showed a good agreement between simulated and experimental data, with underestimation below 5%.  On the other hand, quantification of a representative immuno-PET biodistribution using the Bio-Rat phantom revealed underestimations below 20% across most organs, consistent with partial volume effects at this spatial resolution. 
This framework establishes a reproducible strategy to characterize and correct quantification biases in preclinical PET, with direct impact on the translational development of radiopharmaceuticals.

Abstract

The development of portable, low-field MRI systems is currently limited by the computational complexity of optimizing high-density permanent magnet arrays. While traditional Halbach cylinders offer a baseline for transverse fields, achieving medical-grade homogeneity in non-standard, compact geometries remains prohibitive using conventional loop-based physics. This work introduces an ultra-vectorized "Gen-Tensor" engine that revolutionizes the magnet design process by replacing iterative loops with 4D broadcasting and tensor evolution. By parameterizing 300 discrete magnets into a multiple of 2,100-variable generative DNA block (encompassing 3D position, 3-axis rotation, and activation masks), the framework enables the autonomous synthesis of inherently shimmed, multi-layered topologies.

A significant breakthrough of this study is the discovery of a non-standard Cassinian oval architecture that achieves a field magnitude nearly double that of traditional circular cylindrical benchmarks while maintaining high directional purity of the field. This 100% increase in field strength is achieved without an increase in the magnet budget, utilizing the algorithm’s ability to exploit non-linear flux-path efficiencies and dynamic compaction. The optimization employs a dual-stage differential evolution protocol, transitioning from a high-entropy exploration phase to a low-mutation precision shimming phase.

Abstract

Last year, we developed an ultrasensitive preclinical positron emission tomography (PET) scanner with 4-layer depth-of-interaction (DOI) detectors, thereby allowing sub-second dynamic imaging of a rat. However, the crystal pitch and DOI bins were 2.8 mm and 7.5 mm, which hampered mouse imaging. On the other hand, assembling smaller crystals remains technically challenging. In this study, we designed a novel preclinical PET scanner that uses a sub-surface laser engraving (SSLE) technique to fabricate the finely pixelized crystal elements, and we call it the SSLE-PET. The SSLE-PET consists of 12 DOI detectors, each of which has tapered LYSO crystals with a thickness of 20 mm. The LYSO crystals were pixelized by SSLE, providing fine crystal pitches (top=1.0 mm, bottom=1.6 mm) and 2.5 mm DOI bins. We used GATE simulation to investigate the effects of inner diameter (i.e. detector gap) and axial coverage on the performance of the SSLE-PET. To evaluate spatial resolution and sensitivity, a NEMA ²²Na source was used. For imaging performance evaluation, a resolution phantom was used. The GATE output was reconstructed by using a filtered back projection algorithm without any gap filling method. An energy window of 400-600 keV was used. The spatial resolution (1.27 mm) and resolvable rod diameter (1.0 mm) were not affected by the inner diameter. However, gap artifacts appeared for an inner diameter of 70 mm. Hence, an inner diameter of 65 mm would be the best choice to avoid any mechanical collision between detectors, while maintaining the image quality. With an axial coverage of 88 mm, sub-second PET imaging was demonstrated. As a next step, we plan to develop a prototype SSLE-PET scanner.

Abstract

Real-time monitoring in Boron Neutron Capture Therapy (BNCT) exploiting the 478 keV gamma rays emitted by the neutron capture reaction, remains a major challenge due to the intense background present in the irradiation room during treatment. To date, no clinically established system is available, and the choice between SPECT and Compton camera approaches remains an open and highly relevant question.
Here, we present a Monte Carlo (MC) comparison study between a previously developed BNCT-SPECT detector and a LaBr3-based Compton camera for 478 keV prompt-gamma detection.
The goal is to assess the capability of Compton cameras to exploit electronic collimation and directionality for background rejection, thereby improving the signal-to-background ratio (SBR) with respect
to SPECT.  The results confirm that, in the absence of shielding, SPECT detection is severely dominated by room-induced background. For a vial containing 1000 ppm of ¹⁰B irradiated with a thermal neutron
flux of 3 × 10⁶ n cm⁻² s⁻¹, the SPECT configuration yielded an SBR of only 0.009, while the Compton camera achieved an SBR of 3.2. These findings suggest that a Compton camera-based system may represent a promising strategy to address one of the main limitations of BNCT-SPECT systems towards clinical applicability. However, the choice between the two approaches remains challenging, since properly shielded SPECT configurations can still provide high SBR values, while clinically relevant neutron fluences may increase the single-interaction rate in Compton cameras, making true-coincidence identification more demanding and possibly requiring dedicated shielding as well.

Abstract

Inter-crystal scatter (ICS) in PET detectors causes multiple interactions within a detector block, and misidentification of the first interaction crystal can lead to incorrect LOR assignment and image degradation. Rejecting all ICS events avoids this error but reduces system sensitivity. This study proposes a stochastic conditional-probability-based ICS positioning method using ordering information from deposited energy and DOI. The method estimates first-interaction events by applying conditional probabilities defined over binned observables, where the conditioning variable is either the energy or DOI difference between two crystals involved in an ICS event. These probabilities were obtained statistically from GATE simulations of an 8×8 LYSO crystal array with 3.2 mm pitch. Point sources were placed at 5, 10, 20, and 25 cm, and two-crystal ICS events were selected for validation. The effects of bin count, binning strategy, and detector resolution were evaluated using 2, 4, and 8 bins with equiquantile and equal-width binning. Accuracy exceeded 85% using only 3% of the validation data, corresponding to approximately 20,000 ICS counts per detector block. DOI-based conditioning achieved slightly higher accuracy than energy-based method, suggesting that DOI ordering provides useful geometric information for first-crystal identification. However, the performance difference was small, indicating that either DOI or energy ordering can be used depending on detector readout capability. Accuracy was insensitive to bin count, binning strategy, and detector resolution. These results demonstrate that reliable ICS positioning can be achieved using predetermined conditional probabilities matched to detector characteristics, supporting practical recovery of ICS events without requiring high detector resolution or excessive readout complexity.

Abstract

Compton cameras are electronically collimated imaging devices that reconstruct the origin of gamma rays using Compton kinematics, making them particularly suited for targeted radionuclide therapy (TRT), where conventional gamma cameras struggle with high-energy and low-activity emitters. The European project AIDER is developing a high-performance scintillator-based Compton camera to improve imaging and enable dosimetry in TRT, including targeted alpha therapy. Accurate Monte Carlo modeling is essential in this context for detector design and image reconstruction optimization. We present the validation of Compton camera simulations using GATE version 10, a major evolution of the Geant4-based Monte Carlo toolkit for medical physics with a Python based user interface. To this end, data acquired with MACACO1, a two-layer Compton camera prototype based on  monolithic LaBr₃ scintillation crystals, developed at IFIC-Valencia, was employed. A Monte Carlo model of this system  was previously validated  using a dedicated Compton camera module in an older version of GATE based on legacy static macro-based files. In this work, the model was developed using GATE 10. The detector response was modeled by a digitizer chain that includes the measured  energy and spatial resolution of the prototype. Time coincidences were obtained offline using a time coincidence window of 50 ns.
 Simulated and experimental coincidence energy spectra  showed good agreement despite small discrepancies at low energies. The angular resolution measure at 1275~keV was in good agreement between both simulated and experimental data with a relative difference of 0.7%. The validated framework for Compton camera simulations with  the new GATE 10 modular version, provides a flexible platform for detector  design  and image reconstruction optimization.

Abstract

Whole gamma imaging (WGI) is a hybrid nuclear medicine technique under development at QST that integrates PET and Compton imaging by placing a scatterer detector ring inside a PET detector ring. This configuration is expected to enable simultaneous PET and SPECT tracer imaging and to allow high‑sensitivity PET acquisition using radionuclides such as ⁸⁹Zr, which emit both positrons and prompt gamma rays. In this study, we investigated the feasibility of extending a previously optimized brain‑dedicated WGI design to a whole‑body configuration by scaling the scatterer‑detector ring diameter. Our proposed system expands the scatterer detector ring diameter in the whole-body mode, and fills the missing data due to the detector gaps by a mechanical rotation. Monte Carlo simulations using Geant4 were performed to evaluate the detection performance of the proposed system for ⁸⁹Zr imaging in the whole-body mode, with particular emphasis on the effects of inter‑detector gaps introduced by geometric scaling and the necessity of mechanical rotation in the scatterer‑detector system. Point‑source simulations were conducted to assess Compton sensitivity and angular resolution measure (ARM), and sparse detector configurations were reproduced by masking half of the scatterer detectors. The results showed that Compton sensitivity was approximately doubled when all scatterer modules were installed, indicating that sensitivity is primarily determined by the number of available detectors rather than the presence of gaps. In contrast, ARM exhibited minimal variation across configurations, suggesting that mechanical rotation has little influence on angular‑resolution performance. These findings provide essential design guidelines for whole‑body WGI systems and will be further validated through imaging simulations to quantify the impact of detector gaps and rotation on reconstructed image quality.

Abstract

We have achieved 1 mm rod separation performance in a brain-dedicated PET system, CLS-7, which employs our original crosshair light-sharing (CLS) detector with both time-of-flight (TOF) and depth-of-interaction (DOI) capabilities. From that work, our next question to explore is how CLS technology can improve whole-body PET imaging. PET is widely used in cancer diagnosis, and improving spatial resolution with low image noise is important for detecting small lesions at earlier stages. In this study using Monte Carlo simulation, we investigated the imaging performance of a whole-body PET system using CLS detectors. A whole-body CLS-PET system was modeled using the Geant4 with a ring diameter of 750 mm and an axial field-of-view of 228 mm. The detector was modeled as a 14 × 14 LGSO crystal array with 1.45 × 1.45 × 15 mm³ crystals divided into three depth segments for DOI modeling. The system had 92 detector blocks per ring and seven rings. Spatial resolution and image quality were evaluated by simulating point-source and image quality phantom measurements based on the NEMA NU 2 standard and compared to cases with and without DOI and TOF information. As a result, we found the use of DOI information improved spatial resolution by 40% compared with the no-DOI information condition, while TOF information substantially reduced image noise in the image-quality evaluation. These results suggest that whole-body PET using CLS detectors can suppress peripheral spatial-resolution degradation and improve contrast recovery for small lesions.

Abstract

Reducing radiation dose in positron emission tomography / computed tomography (PET/CT) is possible, but ultra-low-dose (ULD)-CT increases image noise which may affect quantitative reliability and accuracy due to attenuation correction. Photon-counting CT (PCCT) offers improved noise characteristics compared with conventional energy-integrating detector CT (EID-CT), particularly at low dose levels. This study compares PCCT and EID-CT under low-dose (LD) and ULD-CT conditions using a cadaver, an anthropomorphic phantom (PBU-60), and a tissue characterization phantom (Gammex 467). Quantitative evaluation is performed using CT values in HU and corresponding linear attenuation coefficients (μ). Results show that ULD increases noise in both systems, with stronger degradation and reconstruction dependency in EID-CT. Compared with EID-CT, PCCT shows less changes in mean HU values between LD and ULD acquisitions, particularly in dense materials, although noise remains present. After HU-to-μ conversion, variability is substantially reduced for both systems, though not uniformly across all conditions. These findings indicate that attenuation coefficients are less sensitive to dose reduction than CT values and suggest that PCCT may provide more robust quantitative performance than EID-CT under ULD conditions. However, given non-matched acquisition and reconstruction settings results should be interpreted with caution, and require validation in further standardized and PET-based studies.

Abstract

Automated parcellation of the brain into correct regions using dual-modal PET and MRI is essential for quantitative neuroimaging. Existing high-performing fusion networks often rely on large model sizes, which limits their deployment in resource-constrained settings. This paper presents the DualModal Gated Residual Network (DMGR-Net), a lightweight architecture that integrates Bidirectional Cross-Modal Feature Gating (Bi-CMFG) and Residual Soft Deep Supervision (RSDS). Bi-CMFG performs spatial attention across modality-specific encoder branches to preserve fine structural details, while RSDS applies a boundary-weighted loss at multiple decoder scales to sharpen parcellation boundaries by explicitly penalizing errors along anatomical interfaces. Built upon a depthwise-separable 3D U-Net with only 0.121M parameters, DMGR-Net contains 0.206M parameters and 5.30 GFLOPs, yet achieves the lowest HD95 and ASSD among compared methods, surpassing substantially larger models on surface-distance metrics. Ablation studies confirm the complementary roles of each module, and qualitative comparisons illustrate progressively refined boundaries across configurations, demonstrating that a lightweight design can deliver precise surface delineation.

Abstract

Combined PET/MRI is a powerful tool for studying neurological diseases, but resolving cortical structures requires a spatial resolution of about one millimeter. Dedicated brain inserts can achieve this by bringing high-resolution detectors close to the head, provided that the detector electronics meet stringent constraints on integration, thinness, width, length, and MRI compatibility.

The S-PET project develops such a scalable PET detector platform for neuro PET/MRI. The detector module concept is built around a single mounting and cooling structure that spans the full module length and carries sensor tiles, mainboard, and power supply.

The Visor HD sensor tile (49 × 49 mm²) combines a 12 × 12 Broadcom 4 mm NUV-MT SiPMs with a PETsys TOFPET2 ASIC. A dedicated compression scheme maps the 144 SiPM outputs to its maximum 64 ASIC channels: 36 SiPMs are read out directly, while the remaining channels each combine four spatially separated SiPMs that can be disambiguated through their direct neighbors.

The SuperVisor mainboard (134 × 49 mm²) hosts six sensor-tile connectors, an AMD/Xilinx Kintex-7 FPGA, individual voltage and current regulation per tile, and fiber-optic transceivers on a second PCB layer to keep the board short. A separate, exchangeable power-supply board feeds the module.

Thin flex-PCB tile cables, produced in precise lengths up to 454 mm, connect the sensor tiles to the mainboard on the opposite side of the cooling structure.

A first measurement with a BrainPET 7T crystal array (three DOI layers of LSO at 2 mm pitch), scanned across a fan-beam-collimated ²²Na source, yields a mean absolute error below 1.4 mm in both planar directions, confirming that the Visor HD compression scheme works as intended. SuperVisor PCBs are in production and expected in time for the conference.

Abstract

Simultaneous positron emission tomography/magnetic resonance imaging (PET/MRI) can combine PET functional information with MRI soft-tissue contrast in a single imaging session, offering potential benefits for neurological applications such as neurodegenerative disease assessment and brain tumor localization. However, fully integrated PET/MRI systems remain limited, motivating the development of compact PET inserts that can operate inside or around existing MRI scanners. In this work, radiofrequency shielding was developed and evaluated as an enabling step toward an MRI-compatible Radialis brain PET insert using digital detector electronics. Shielding was implemented for the digital front-end modules, data-acquisition electronics, high-speed data-transmission cable, and auxiliary power cable of a Radialis PET detector array. Each detector module consisted of a 24 × 24 cerium-doped lutetium yttrium orthosilicate scintillator array coupled through a borosilicate light guide to an 8 × 8 silicon photomultiplier array. The shielded detector configuration was evaluated during simultaneous acquisition on a 3.0 T MRI system using a ²²Na point source in a water phantom positioned within the MRI brain-coil geometry. The redesigned shielding enabled stable PET data acquisition, coincidence processing, and point-source image reconstruction during MRI operation. PET energy resolution was maintained at approximately 22.5–22.6% during simultaneous acquisition, although reconstructed image signal-to-noise ratio decreased from 76.64 outside the MRI environment to 37.83 during PET/MRI acquisition, likely reflecting residual MRI-related effects, head-coil attenuation, and source-positioning differences. These results demonstrate that targeted shielding of Radialis digital PET electronics enables reliable simultaneous PET/MRI operation and supports further development of a high-sensitivity brain PET/MRI insert.

Abstract

Small-animal single-photon emission computed tomography (SPECT) serves as a cornerstone of functional nuclear imaging, with extensive applications in preclinical research, drug discovery, and life sciences. Building upon our previously developed full-ring cascade gamma coincidence imaging system with continuous LaBr3 crystals, this work redesigns a multi-pinhole collimator to enable full-ring SPECT imaging for ^99mTc. To optimize computational efficiency, the system transfer matrix is implemented using a sparse storage strategy. Furthermore, GPU acceleration is employed for key processes, including system matrix calculation, forward and backward projections, and image updates, significantly enhancing the performance of the iterative reconstruction algorithm. Experimental results demonstrate that the storage requirement for the system transfer matrix is reduced from 10.06 GB to 77.7 MB. Concurrently, the computational efficiency is significantly enhanced, achieving a single iteration time of 0.94 ms in binning mode. Moreover, the full-ring SPECT system achieves a 1.0×10^-3 detection efficiency and successfully resolves 1.2 mm diameter hot rods. In conclusion, the proposed full-ring SPECT system achieves a high-performance balance between detection sensitivity and spatial resolution.

Abstract

At the Biomedical Imaging Laboratory of the Institute of Physics, UNAM, a dual-panel positron emission mammography (PEM) prototype is under development. This work evaluated its spatial resolution through the line spread function, and image quality via recovery coefficients. A custom variable-geometry phantom was developed using 3D-printed supports for four ¹⁸F line sources with a 0.69 mm internal diameter. Data acquisitions were performed at 11 in-plane positions and at different cross-planes heights to map the 17.4 cm × 17.4 cm × 6.0 cm effective field-of-view. Experimental setup included a panel separation distance of 6.0 cm, 8 ns coincidence window and a 360 – 757 keV energy window. The interaction position of each event is estimated using a raise-to-a-power center-of-gravity algorithm. Images are reconstructed using a list-mode ordered subsets expectation-maximization algorithm (3 iterations, 4 subsets) using a 1 mm isotropic voxel. Image quality was evaluated with a NEMA-like phantom containing 4 spherical lesions (diameters of 4, 6, 8 and 10 mm) filled with ¹⁸F, placed at different field-of-view positions and phantom orientation. The effect of restricting oblique lines of response was also studied. In addition, the sensitivity was analyzed using a point-like ²²Na source in different field-of-view positions. The best spatial resolution was 2.1 ± 0.2 mm at the center of the PEM field-of-view and decreasing to 4.5 ± 0.3 mm at the PEM edges. FWHM values closest to the real sphere diameters were obtained at the detector center. Limiting oblique lines-of-response to reduce depth-of-interaction effects resulted in lower spatial resolution. Overall, the prototype showed good performance, with adequate image quality regardless of spherical-lesion position or orientation.

Abstract

Monolithic scintillators are a promising alternative to standard pixelated crystals in positron emission tomography, as they can be more cost-effective, reduce dead space, and capture an unobstructed light distribution, which can be used for depth of interaction estimation. We focus on the improvements seen by using an array of integrated corner reflectors (ICR) machined directly into the entrance face of the monolith, which we previously showed via simulation to improve light collection and narrow the light distribution. In this study, we built a physical prototype of the device using a CNC machining process. We compared this prototype with other monolithic block designs. We show that the ICR monolith with the CNN positioning algorithm has potential for high-resolution imaging with ~1 mm resolution at 12 mm thickness. Using GATE simulations, we constructed a full scanner with the performance of the monolithic blocks and tested the capabilities compared to existing systems. We show the improved spatial resolution and feasibility for a non-cylindrical design for axillary lymph node scanning.

Abstract

Pixelated crystal arrays are the standard in commercial PET scanners. However, these detectors usually lack DOI information, essential to achieve a homogenous spatial resolution along the entire FOV. Yet, studies have been already performed with pixelated detectors with DOI capabilities but at the cost of using extra hardware material and/or crystal assembling. Inspired by semi‑monolithic crystal geometries, pseudo-slabs were proposed as an alternative. This geometry consists of an N×N pixel array optically isolated only in one direction, allowing light crosstalk in the orthogonal one. The N pixels sharing light in this direction form a pseudo-slab. The light sharing within a pseudo-slab is enough to enable DOI retrieval, while the physical separation between pixels confines the light more than in a pure slab geometry, yielding better time resolution.

In this work we evaluated the DOI resolution of a BGO block formed by 8×8 pixels, each measuring around 3×3×15 mm³. Two different readout configurations were employed and compared. First, all the 64 channels corresponding to the TOFPET2 ASIC channels were digitized. Then, a multiplexing circuit was utilized, reducing the signals to 8+8 by summing rows and columns. The DOI was obtained using neural networks trained on experimental data.

Abstract

The spatial resolution of PET imaging is fundamentally limited by the positron range, whose extent depends on both the emission energy of the radionuclide and the electron density of the surrounding medium. In integrated PET/MRI systems, the static magnetic field exerts a Lorentz force on the positrons, confining their trajectories in the transverse plane to the main magnetic field, potentially improving image resolution. Two complementary experiments were performed using a dedicated monolithic-crystal based PET insert operated at 0T and inside a 9.4T MRI: capillary line sources embedded in three tissue-equivalent phantoms of increasing density (EPF, PLA, and Rigid10k resin), and a microDerenzo phantom. Both tests were carried out with three positron emitters (¹⁸F, ⁸⁹Zr, and ⁶⁸Ga) eventually use in pre-clinical imaging. The results demonstrate that the magnetic field significantly enhances transverse spatial resolution for high-energy positron emitters in medium- and high-density materials, with FWTM reductions of ~60% for ⁶⁸Ga and ~18% for ⁸⁹Zr, while the axial direction remains largely unaffected. Rod resolvability improved at 9.4T from 1.5 mm to ~1.0 mm and from 1.2 mm to 0.9 mm for ⁶⁸Ga and  ⁸⁹Zr, respectively. . These findings highlight the relevance of both positron energy and tissue-equivalent density in determining the magnitude of magnetic-field-induced resolution improvements in preclinical PET/MRI.

Abstract

The aim of this study was to optimize detector parameters of crosshair light-sharing (CLS) detectors to achieve a better resolution-sensitivity balance in the CLS-7, a brain-dedicated PET prototype we developed in 2025. In particular, we focused on position lookup tables (LUTs), which are generated from flood histograms calculated using Anger logic. In the 2025 model, restricted position LUTs fitted to crystal clusters were used to reduce inter-crystal scattering (ICS), thereby improving spatial and time-of-flight (TOF) resolution at the cost of decreased system sensitivity. In this study, we employed expanded position LUTs and compared the performance of restricted and expanded LUTs in the CLS-7. The detectors consisted of 15-mm-long fast-LGSO scintillators arranged with a crystal pitch of 1.6 mm. Sensitivity, noise-equivalent count rate (NECR), and TOF resolution at the peak NECR were evaluated for both settings. In addition, image quality was evaluated using a rod phantom, an IQ phantom, and a Hoffman phantom. Compared with the restricted LUTs, the expanded LUTs significantly improved sensitivity and the peak NECR by factors of 1.7 and 1.9, respectively. The loss in TOF resolution was only 12%, and no significant difference was observed in the rod phantom results. In both the IQ phantom and the Hoffman phantom, images obtained using the expanded LUTs had lower noise levels than those obtained using the restricted LUTs. These results demonstrate that expanded position LUTs can improve overall image quality despite some degradation in TOF resolution in the CLS-7 system.

Abstract

This study presents a performance evaluation of N-PET, a high-resolution TOF-DOI PET system designed for human brain imaging. The system has a compact ring geometry comprising 18 detector modules, with a ring diameter of 302 mm, a transaxial FOV of 210 mm, and an axial FOV of 52 mm. Each module employs dual-layer, multi-resolution LYSO detectors that enable both high spatial resolution and DOI measurement.

Abstract

We have developed a high resolution small-animal in-beam PET scanner as part of the SIRMIO project for accurate image-guided proton irradiation at clinical facilities. The scanner consists of 56 three-layer depth-of-interaction (DOI) LYSO detectors with 1.0 mm-pitch scintillator pixels arranged in a compact spherical geometry with a with an entrance radius of 72 mm. Using different irradiation platform, we successfully demonstrated small-animal in-beam PET imaging. Since the number of positron emitters generated during irradiation is limited by the prescribed dose, intrinsic Lu background can contribute to background noise. In this study, we investigated the optimal energy window settings for different activity levels using Monte Carlo simulations based on coincidence count statistics in the SIRMIO PET scanner. The three-layer DOI detector structure enables independent energy window settings for each DOI layer. Since 511 keV gamma rays are more likely to interact near the shallow side of the detector, wider energy windows were selectively applied to shallow DOI layers, while narrower windows were used for deeper DOI layers to suppress Lu-176 background contributions. Geant4 simulations were performed for source activities ranging from 0.05 to 50 kBq. Net significance was evaluated for combinations of lower energy thresholds of 250 - 450 keV with a fixed upper threshold of 600 keV. For low-activity cases such as 0.05 kBq, the optimal lower energy thresholds were 400, 450, and 450 keV for the first, second, and third DOI layers, respectively. In contrast, wider energy windows became more advantageous at higher activity levels. These results indicate that layer-dependent energy window optimization is an effective method for improving low-count in-beam PET imaging performance.

Abstract

Background

Depth-of-interaction (DOI) is one of the remaining challenges for high resolution PET. The resolution degradation stems from incorrect LOR endpoint estimation when photons impinge on the detector at oblique angles. This problem increases with distance from the symmetry axis and poses a challenge for LAFOV PET, where more LORs are acquired at oblique angles. Terapet SA manufactures radiation detectors capable of localising gamma ray interactions in 3D within the detector volume, potentially mitigating this effect. The aim of this project was to characterise the performance of a prototype PET scanner using Terapet’s proprietary detectors, particularly the potential of DOI artefacts reduction.

Method

The NEMA NU-2 performance tests were conducted at the Department of Nuclear Medicine at Inselspital, Bern. Additionally, so far, two patient scans have been performed.

Results

The prototype achieved a FWHM of 2.56, 2.60 and 2.96 mm, averaged over all axes, at 10, 100 and 200 mm radial offset, respectively. The loss of resolution with increasing radial offset was lower in the prototype than in other clinical scanners, particularly along the radial axis.

The prototype had a low detection sensitivity of 0.99 cps/kBq. This is explained by the restrictive selection requirements on single- and coincidence events, optimised for high-quality data rather than high sensitivity.

The image-quality phantom and two patient images have been acquired. Results will be presented at the conference. The expectation is improved contrast recovery due to the improved spatial resolution, particularly for the smallest sphere.

Conclusion

The prototype demonstrated improved spatial resolution with the proprietary detectors. In particular, the spatial resolution is maintained with increasing radial offset. Future work will be aimed at improving the image reconstruction algorithms and correction methods. Construction of a LAFOV scanner using the same proprietary detectors is ongoing.

Abstract

High-sensitivity, high-resolution brain-dedicated SPECT systems are critically important for advancing diagnosis and surgical planning in complex neurological disorders such as epilepsy, Parkinson's disease, and Alzheimer's disease. In this paper, we present SurfaceCam, a close-proximity multi-pinhole SPECT architecture with a radially adaptable rotating gantry, designed specifically for clinical brain imaging. The system comprises 15 detector-collimator panels distributed across three detector heads in a cylindrical arrangement, operating at a 1:8 minification geometry with a 28 cm bore. Each panel integrates 16 high-performance CZT sensors in a compact 2×8 configuration. The three detector heads translate radially to accommodate varying patient head sizes while maintaining close-proximity geometry, and the rotating gantry enables micro-rotation between acquisitions to facilitate angular super-sampling strategies. Pinholes are organized into groups of four with a twisted geometry, each subtending partial coverage of the 20 cm spherical FOV, while eliminating inter-pinhole projection multiplexing on the detector surface. Analytical comparison against a conventional 1:11 minification compound-eye reference design demonstrates approximately 1.4-fold higher geometric sensitivity and approximately 27% improvement in effective pinhole spatial resolution. The front-end electronics and data acquisition systems have been fully developed and validated, with Monte Carlo simulation studies and experimental phantom validation currently underway.

Abstract

α-emitting radiopharmaceutical therapies (αRPT) show great potential to treat solid and diffuse tumors. However, following α-decay, the recoil energy and chemistry of daughter isotopes separate them from the targeting vector, exposing healthy tissue to an elevated radioactive dose. These effects must be studied in small animal experiments before new αRPTs make it to clinical trials. However, since extremely low activities are used, current single-photon emission computed tomography is insufficient due to its low sensitivity, especially for the 440 keV gamma ray emitted by ²¹³Bi (a daughter of one of the most promising radionuclides for αRPT, ²²⁵Ac). In previous work, we used a high-purity Ge Compton camera as a high-sensitivity solution for pre-clinical imaging of αRPT. We demonstrated quantitative Compton imaging of ²¹³Bi using total-variation maximum-a posteriori expectation-maximization and a post-reconstruction global activity calibration. In this work, we improve upon this by implementing non-region-of-interest (nROI) reconstruction and a Gaussian Process Prior (GPP). We image a mouse phantom at relevant activity concentrations and measurement times to demonstrate that nROI reconstruction mitigates artifacts by redistributing events outside of the field-of-view across the image. We also show that, by using a GPP that correlates the activity in proximate voxels, we improve activity concentration quantification in all the volumes of the mouse phantom with respect to ML-EM. We modify both methods from previous iterations for local activity and uncertainty quantification, which is not straightforward in conventional ML-EM, and compare uncertainties obtained with the Laplace Approximation and preconditioned Crank-Nicholson Markov-Chain Monte Carlo to explain discrepancies from the ground truth. Finally, we experimentally measure a spatial resolution of ~3mm FWHM with a mini-Derenzo phantom.

Abstract

Grating-based X-ray interferometry dark-field (DF) imaging provides exceptional sensitivity for detecting microstructures below the spatial resolution of standard detectors. Traditionally, this technique requires either large-scale synchrotron facilities or the integration of a source grating into a conventional X-ray tube setup. This study investigates a compact, simplified Talbot interferometer configuration utilizing a microfocus X-ray tube paired with a pixelated, energy-resolved photon-counting detector. In this setup, the microfocus source ensures the necessary spatial coherence, while the detector’s energy-resolving capability enables the selection of specific energy bins—directly correlated to wavelength—to optimize reconstruction at the pixel level. Validation was conducted by imaging a metallic screw (0.5 mm pitch) using the phase-stepping method. Results demonstrate that while the sharp thread lines were heavily blurred in attenuation and phase-contrast modalities, they were clearly resolved in the dark-field images. Two optimization methods were evaluated for the full-energy dataset. Method 1, based on fitting effect correction through pixel filtering and interpolation, increased the Contrast-to-Noise Ratio (CNR). Method 2, based on the probability distribution approach, further improved the dark-field image quality compared to Method 1. This method resulted in a brighter signal along the threadline and significantly fewer rejected pixels, ensuring a more complete and better-defined representation of the object. These results demonstrate that high-resolution dark-field imaging is achievable through a compact laboratory setup combined with targeted data-processing optimizations.

Abstract

Virtual X-ray imaging experiments are a crucial research technique in the development of X-ray technology. In recent years, advanced imaging detectors, such as photon-counting detectors, have been developed and are increasingly being used in clinical applications. It further increases the importance of virtual X-ray imaging experiments. To improve the reproducibility of patients in virtual experiments, the development of digital phantoms having diseases is desired. The purpose of this study is to develop a procedure for creating digital phantoms that can reproduce diseases. Virtual monoenergetic images (VMIs) at 40 keV and 130 keV were generated using a clinical photon-counting CT (PC-CT) system, and effective atomic number (Z_eff) and effective physical density (ρ_eff) were calculated using a newly developed algorithm. Z_eff was calculated by using the correlation between the ratio of measured linear attenuation coefficient (μ) values and atomic number (Z). ρ_eff was calculated from the fitting coefficient obtained by fitting the reaction cross-section (σ) of each Z material to the measured μ values. Then, 10 patient data examined using PC-CT were analyzed, and digital phantoms were created. We investigated the reproducibility of Z_eff and ρ_eff values of the disease tissues. Our procedure could reproduce the disease tissues by continuous Z_eff and ρ_eff distribution. In conclusion, we developed a methodology for constructing digital phantoms of patients having diseases from VMIs acquired with PC-CT.

Abstract

CBCT image quality is strongly degraded by scattered radiation. Conventional anti-scatter grids (ASGs) are commonly used to mitigate this effect; however, they can only remove a portion of the scattered photons. In this study, we propose a novel computational scatter estimation method based on a half anti-scatter grid (h-ASG), which covers only half of the flat-panel detector. The full scatter distribution is estimated by combining two reconstructed CBCT images: one containing the full scatter component and the other containing a reduced scatter component. Experimental results using CatPhan and anthropomorphic head phantoms demonstrate that the proposed method effectively suppresses scatter-induced artifacts and significantly improves CBCT image quality.

Abstract

X-ray computed tomography (CT) is essential nondestructive internal imaging. Photon-counting computed tomography (PC-CT) acquires X-ray energy information to enable material identification and quantitative density evaluation. Combining PC-CT energy information with magnetic resonance imaging (MRI) data is expected to provide a novel diagnostic approach to visualize inflammation through tissue water fraction estimation. We developed a PC-CT system combining high-speed scintillators (YGAG:Ce) and silicon photomultipliers (SiPMs), specifically utilizing multi-pixel photon counters, MPPCs, which are optical semiconductor devices. This study evaluated healthy and fatty liver rat models (fed a high-fat diet for two or four weeks). Liver fat fractions, estimated from energy-dependent linear attenuation coefficients, showed good agreement with MRI and pathological results. Furthermore, we performed a detailed pathological evaluation focusing on inflammation induced changes in tissue water content. By integrating high reliability MRI fat fraction data with PC-CT data, we improved the estimation accuracy of water content and quantitatively visualized tissue ”water-richness” (edema). Increased water proportions in fatty liver groups suggested inflammation progression. These findings demonstrate that PC-CT is highly valuable for quantitative evaluation of liver disease within a multi-modality imaging approach.

Abstract

Many image reconstruction methods for x-ray computed tomography (CT) use optimization approaches with cost functions that consist of a data fidelity term and a prior term. Extensive efforts have been devoted to construct better priors, some using machine learning, for high resolution, low dose, and spectral imaging. There seems to be a belief that a good prior can overcome any problems in data acquisition, downplaying the data fidelity term. It would be desirable if we could quantify the sufficiency of the problem setting (or adequacy of the data acquisition) toward a set of desired imaging properties. We propose “Truth index” or “T-index” to quantify data range and data sampling issues in this work. When fully developed, the T-index will show, independent of specific image reconstruction methods used, how much the solution can be trusted or has a risk of hallucination.

Abstract

Photon-counting CT (PC-CT) is a powerful spectral imaging technique that identifies materials by counting individual X-ray photons. While direct-conversion systems are entering clinical practice, we are developing a cost-effective, large-area scintillator-based PC-CT system using YGAG:Ce coupled with silicon photomultipliers (SiPMs). PC-CT excels at identifying high-atomic-number elements like gold (Au) by utilizing its K-edge (80.7 keV). Gold nanoparticles (AuNPs) are promising for drug delivery, as their biodistribution can be controlled by particle size, enabling selective tumor accumulation via the enhanced permeability and retention (EPR) effect.

In this study, we evaluated the in vivo biodistribution of AuNPs in tumor-bearing mice using our SiPM-based PC-CT. Two sizes of AuNPs (1.9 nm and 15 nm) were intravenously administered, and their accumulation was quantitatively assessed using multi-energy CT concentration estimation. Results showed distinct size-dependent patterns: 1.9 nm AuNPs exhibited rapid renal clearance and bladder accumulation, while 15 nm AuNPs showed prolonged circulation and significant hepatic accumulation. Tumor accumulation was confirmed for both sizes, demonstrating the potential of our SiPM-based PC-CT for quantitative in vivo nanoparticle imaging.This presentation reports on the
experimental setup, the analysis methods, and the results obtained.

Abstract

Cone-beam computed tomography (CBCT) is a vital imaging modality rooted in nuclear science. It uses ionizing radiation and information processing to provide patients' internal anatomical details. When patients with metal implants undergo CT scans, projection data is contaminated. Consequently, reconstructions produce severe metal artifacts, compromising diagnostic accuracy. Thus, developing high-performance metal artifact reduction (MAR) methods for CBCT is essential for accurate clinical diagnosis. Some existing MAR methods ignore physical phenomena, like X-ray scattering, during imaging. Others use scatter correction as a preprocessing step; however, this non-end-to-end approach increases complexity. Here, we propose an end-to-end MAR method incorporating scatter correction and metal segmentation within the projection domain into a single network, followed by morphological post-processing and artifact suppression. Qualitative and quantitative analyses on simulated and real datasets demonstrate the superiority of this end-to-end method.

Abstract

In radiography imaging, dual-layer flat-panel detectors (DFDs) can be used for two purposes: acquiring high-quality x-ray images and performing material decomposition.
When using cone beam projection in DFD, the x-ray beam is scattered inside the object, increasing and distorting the intensity. In this paper, we experimentally observe x-ray scattering in acquiring images from DFD. The image acquired from the lower detector layer shows more scatter to prime ratios (SPRs) than the upper layer case due to the scattering inside the upper and intermediate layers. We perform parameter modeling for experimental SPR curves and simulations regarding the measurement of bone mineral density (BMD) estimates, including scattering intensity. We can observe that using the 2nd-order polynomial BMD function or fully connected neural network can successfully estimate BMD functions without scatter correction.

Abstract

The growing demand for accessible, high-resolution imaging in pre-clinical and materials research has driven interest in low-cost micro-computed tomography (micro-CT) platforms. In this work, a bench-top micro-CT system was designed, assembled, and experimentally characterized, built around a microfocus X-ray tube, a CMOS flat-panel detector, low-cost open-source electronics based on an STM32 microcontroller, and the open-source TIGRE toolbox for volumetric reconstruction. A step-and-shoot acquisition strategy --- in which the sample rotates incrementally, and the detector captures frames only during active X-ray exposure via a hot duration trigger mode --- was employed for synchronized image acquisition. Three reconstruction approaches were implemented and comparatively evaluated: filtered back projection (FBP), Feldkamp-Davis-Kress (FDK) algorithm as the primary volumetric reconstruction method, and iterative techniques for noise suppression. System performance was assessed through established phantom protocols evaluating image uniformity, noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and a QRM micro-CT Bar-pattern phantom for spatial resolution characterization. While a persistent precession artefact attributed to rotation stage mechanical instability was identified as the primary performance limitation, the system demonstrated tomographic imaging capability and sub-millimeter spatial resolution, establishing a promising foundation for the development of accessible and cost-effective micro-CT platforms for laboratory research.

Abstract

A large-scale X-ray image dataset acquired under diverse specimen and imaging conditions is essential for developing robust AI-based container-structure inspection models. However, single-view transmission imaging often fails to resolve structural ambiguities in complex container environments, where X-ray attenuation and scattering characteristics vary significantly with specimen geometry, material composition, and overlap conditions. In this study, a robot-arm-based automated transmission and backscatter X-ray image acquisition system was developed to enhance the diversity, reproducibility, and traceability of AI training data. The system employs a 225 kV, 13 mA X-ray source to sequentially acquire vertical and horizontal transmission images and horizontal backscatter images. The transmission unit consists of 1,408 detection channels based on eleven 1×128-pixel line-detector modules, whereas the backscatter unit incorporates four 600×500 mm plastic scintillation detectors. By precisely controlling specimen orientation using a robot arm, the system stores acquired images together with preprocessing results, quality-assessment data, and imaging-condition metadata for subsequent AI model training, validation, and evaluation.

Abstract

We quantify the impact of scatter on the performance of bone suppression in virtual non‑calcium chest radiographs (VNCa‑CXR) generated with (i) a multi‑layer flat‑panel detector (ML‑FPD, three 200-um thick CsI layers) and (ii) with a kV‑switching protocol (60/140 kVp). We show that in the setting of log-subtraction, the decomposition weight (DW) can be adjusted for scatter-to-primary ratio (SPR) such that the scatter-induced bias in VNCa is locally corrected (scatter-compensating weight, DW-scat). However, DW-scat is spatially varying. If a constant DW is used in log-subtraction, the magnitude of bone residual in VNCa radiographs will depend on the degree of spatial inhomogeneity in DW-scat. We investigate this effect is simulation studies in anthropomorphic chest phantoms (20 – 24 cm thickness) with and without an anti-scatter grid (ASG).  Without a scatter‑compensating DW, scatter‑induced bone residual increased by >100% is most configurations except for ML-FPD with a grid. Applying DW-scat averaged over central lung region reduced the residual by up to 50% for ML-FPD without an ASG and by ~40% for kV-switching with an ASG. Overall, our results reveal that DE VNCa imaging is relatively robust to scatter in scenarios with a grid, where kV-switching appears to benefit from spatially-averaged scatter-compensating DW more than ML-FPD, likely due to differences in scatter content of the two energy channels.  Without a grid, both configurations may require a spatially-dependent scatter compensation for optimal bone suppression.  

Abstract

X-ray time of flight information can be used in computed tomography to characterize photon interactions and discriminate scattered from unscattered events. Specifically, photons undergoing Compton and Rayleigh scattering are subject to trajectory deviations that increase their path length relative to primary photons. Therefore, with adequate temporal resolution, scattered photons are detected with larger time delay with respect to primary photons, providing a physically grounded mechanism to reject scatter. Recent technological developments in ultra-fast X-ray sources (e.g: 50ps) and high-temporal-resolution photon-counting detectors may allow for this imaging modality.

In this work, Monte Carlo (MC) simulations of a water cylinder were performed assuming Gaussian source pulse shape and detector timing jitter, with combined full-width-half-maximum as low as 50ps. A threshold-based scatter rejection method was applied, sweeping multiple thresholds to find the optimal balance between scatter rejection efficiency (η) and primary retention (ε). Filtered-backprojection reconstructions were performed and evaluated in terms of root mean square error and bias. Rejection performances were compared with an ideal, MC-based scatter subtraction technique, that leverages template fitting to estimate the average scatter contribution from time-delay histograms. A combined rejection-subtraction approach was also evaluated.

Abstract

Spectral CT enables material-specific imaging but suffers from noise amplification in low-concentration iodine decomposition. This work investigates advanced material decomposition algorithms for improving iodine detectability in a simulated four-channel photon-counting CT system. Diffusion Posterior Sampling (DPS) with hybrid priors was compared with conventional image-domain and projection-domain decomposition methods. Results demonstrate improved noise suppression and low-concentration vessel detectability while highlighting the noise-bias tradeoff in diffusion-based material decomposition.

Abstract

Low-dose computed tomography (LDCT) is essential for mitigating radiation risks to patients, and its effectiveness is significantly enhanced by automatic tube current modulation (TCM) that optimizes photon allocation across different projection angles to maintain image quality at reduced doses. However, most traditional methods offer fixed modulation strategies based on typical human phantoms, thereby ignoring individual anatomical variations and failing to achieve personalized dose allocation. This paper proposes a personalized TCM method driven by an end-to-end neural network (TCMnet). Specifically, we construct a differentiable optimization chain that integrates the modulation strategy, forward projection, and image reconstruction into a unified framework, allowing the system to learn optimal, patient-specific current profiles by directly minimizing reconstruction errors under specific dose constraints. Experimental results demonstrate that, under identical photon count (radiation dose) constraints, the proposed method yields significantly higher reconstruction quality compared to conventional approaches. By analyzing the performance of these optimized profiles across varying dose levels, this study reveals an important physical phenomenon: under extreme low-photon-count conditions where photon resources are severely limited, the optimal modulation strategy shifts from seeking a globally balanced noise distribution to concentrating photons on fewer key projection angles in high-attenuation regions. This finding not only validates the optimization capabilities of the end-to-end framework under complex constraints but also provides new insights and design concepts for TCM mechanisms for ultra-low-dose CT.

Abstract

Depth-of-interaction (DOI) encoding in positron emission tomography (PET) systems effectively mitigates parallax errors and improves spatial resolution, however, it introduces substantial complexities for accurate and efficient image reconstruction. Specifically, the dramatic increase in valid lines-of-response (LORs) and the inherent limitations of conventional correction methods—such as component-based normalization (CBN) and single scatter simulation (SSS)—often compromise quantitative accuracy in DOI enabled systems. To address these challenges, we propose a fully GPU-accelerated reconstruction framework specifically designed for DOI-PET with complete correction techniques. The proposed framework extends standard CBN and SSS methods through sub-layer-based geometric modeling, enabling layer-specific normalization and scatter correction. A maximum-likelihood-based strategy is introduced to estimate DOI-dependent scatter scaling factors, incorporating inter-layer attenuation effects and axial response variations. The framework was evaluated using Monte Carlo simulations of the NeuroEXPLORER scanner. Our results demonstrate that the reconstructed uniform cylinder phantom exhibits high spatial homogeneity. Furthermore, the spatial resolution quantified via the Derenzo phantom and the quantitative accuracy achieved in a digital brain phantom are consistent with previously established benchmarks. These findings suggest that the proposed framework provides a robust and computationally efficient solution for high-resolution, quantitative reconstruction in DOI-PET systems.

Abstract

Micro X-ray computed tomography (micro-CT) enables non-destructive visualization of fine internal structures. Resolving finer details requires higher geometric magnification, typically achieved by reducing the source-to-rotation-center distance (SRD). However, short-SRD acquisition can suffer from a limited angular range due to physical interference with the X-ray source and from projection truncation when the object extends beyond the field-of-view (FOV), causing streak artifacts and structural distortions. Supervised learning-based reconstruction methods have been investigated to mitigate such artifacts but applying them to non-destructive high-magnification micro-CT is difficult because fully sampled, non-truncated reference data cannot be readily acquired without destructive sample preparation. Moreover, many existing methods are designed for 2D fan-beam geometry rather than 3D cone-beam geometry commonly used in micro-CT. We propose an unsupervised reconstruction method that simultaneously reduces limited-angle and projection-truncation artifacts by incorporating long-SRD measurements. Long-SRD acquisition provides full-view, non-truncated projections, albeit at lower spatial resolution, complementing the high-resolution but incomplete short-SRD acquisition. By jointly using these complementary measurements, the proposed method preserves fine structural details while recovering the global object structure without requiring ground-truth images. Numerical experiments demonstrated that the proposed method reduced streak artifacts and recovered structural information outside the short-SRD FOV. To the best of our knowledge, this is the first unsupervised 3D cone-beam reconstruction method to simultaneously address both artifacts in high-magnification micro-CT.

Abstract

Time-of-flight PET (TOF-PET) enables CT-free attenuation correction (AC) through simultaneous activity and attenuation (SAA) reconstruction from emission alone. In this study, we propose a self-supervised SAA framework based on implicit neural representation. The proposed framework employs parallel hash encoders and multilayer perceptrons to learn the mapping from spatial coordinate to image space, thereby jointly reconstructing activity and attenuation maps. To further accelerate reconstruction and reduce GPU memory consumption, an ordered-subset training strategy is introduced, in which projection angles used for supervision are progressively increased during optimization. The proposed method was evaluated using analytical simulations based on a BrainWeb brain phantom with ¹⁸F-FDG uptake. Compared with non-attenuation-corrected MLEM, maximum likelihood reconstruction of attenuation correction factors (MLACF) and maximum-likelihood activity and attenuation (MLAA), OSIR demonstrated improved visual consistency and quantitative accuracy. For activity reconstruction, OSIR achieved the highest PSNR of 24.62 dB and the lowest NRMSE of 5.87%, with SSIM comparable to MLAA. For attenuation reconstruction, OSIR outperformed MLAA across all evaluation metrics. In addition, OSIR reduced the reconstruction time from 60 min to approximately 40 min. These results demonstrate that OSIR provides an efficient and robust solution for CT-free attenuation-corrected TOF-PET reconstruction, enabling accurate joint estimation of activity and attenuation maps without requiring CT images or large-scale training datasets.

Abstract

Full-ring multi-pinhole SPECT offers high sensitivity and compact detector coverage, but eFOV reconstruction remains challenging due to projection truncation, incomplete angular sampling, and reduced peripheral sensitivity, especially for large-body or off-center activity distributions.

We propose a staged deep-learning-assisted OSEM framework for eFOV recovery. The framework combines a physics-based forward model, two-stage sinogram completion, and image-domain refinement. Truncated projections are first reconstructed by OSEM as the baseline. Two residual U-Nets then sequentially restore missing detector-width information and angular views in the sinogram domain. Finally, an image-domain network suppresses residual artifacts and improves structural continuity in the eFOV region.

Simulation studies used FDG activity distributions with randomized body scaling, translation, sparse sampling, and peripheral lesion insertion. Evaluation metrics included recovery coefficient, MAE, RMSE, SSIM, and lesion-level recovery. Compared with truncated-data OSEM, the proposed method improved peripheral lesion recovery and reduced reconstruction errors. In representative cases, recovery coefficient increased from 25.9% to 96.6% at 550 mm and from 10.1% to 82.5% at 650 mm. Preliminary 3D validation also showed lower errors and higher structural similarity in axial and sagittal views. These results suggest that combining physics-based modeling with dual-domain deep learning is feasible for eFOV reconstruction in full-ring multi-pinhole SPECT. Future work will validate the framework using physical phantom and clinical data.

Abstract

Motion during scanning results in motion artifacts that distort anatomical structures in computed tomography (CT) images. In radiation therapy, planning CT scan duration are high-speed, whereas linear accelerator gantry mounted CT takes more than 15 seconds, making them more susceptible to motion artifacts. Recently, CT image reconstruction methods using supervised learning have been shown to suppress motion artifacts. However, obtaining sufficient training data is difficult because actual organ motion cannot be directly visualized. In this study, we propose an unsupervised CT image reconstruction method incorporating deep image prior and VoxelMorph (DIP-VM). The proposed method iteratively alternates between DIP-based image reconstruction and VM-based estimation of the motion vector field, in which planning CT images are leveraged as reference images. The main advantage of the proposed DIP-VM is its ability to correct patient motion using only patient-specific data, without requiring large training datasets. The proposed DIP-VM was evaluated on an XCAT phantom with cardiac and respiratory motion. Experimental results demonstrate that, using only patient-specific projection data, the proposed DIP-VM method reduces motion artifacts more effectively than the comparison methods. This study provides an effective solution for suppressing motion artifacts using only patient-specific projection data, without requiring large training datasets.

Abstract

Recent advances in diffusion models for natural image processing [1] have led to the development of several diffusion prior-based methods for medical image reconstruction [2,3,4,5,6]. For these diffusion model-based methods, balancing fidelity to the observed data (e.g., projection data) and the diffusion prior is essential. In diffusion models, low noise levels correspond to fine-grained details and therefore better preserve consistency with the observed data, whereas high noise levels correspond to coarse-grained information and rely more heavily on the diffusion prior. In our previous study, we proposed an ensemble method using multiple diffusion priors [7]. This method successfully optimized the balance between observed-data fidelity and diffusion priors in moderately sparse-view settings. However, we found that the advantage of the ensemble method decreases in extremely sparse cases with fewer than 25 views. Under such extremely sparse conditions, a diffusion prior with a large time step alone performed better than the ensemble method.
In this study, we propose a boosting diffusion model for sparse-view CT reconstruction, particularly targeting cases with severe information loss in the observed data. In the proposed method, image reconstruction is first performed using a large time step, resulting in a coarse-grained reconstruction. Subsequently, reconstruction is refined using smaller time steps, leading to fine-grained image reconstruction. We compared the proposed method with existing methods, including iterative reconstruction (IR), DPS, and SDPS, in terms of SSIM and PSNR. The results demonstrated that the proposed method outperforms the existing methods in extremely sparse-view cases with fewer than 25 views.

Abstract

Single photon emission computed tomography (SPECT) remains the standard modality for uptake verification and patient-specific dosimetry in molecular radiotherapy (MRT), however it has limited spatial resolution and noise characteristics compared to pre-therapy positron emission tomography (PET). Existing approaches that incorporate PET priors into SPECT reconstruction are often constrained by hand-crafted feature representations and suffer from inter-modality mismatch and registration-induced artifacts. In this work, we employ a conditional generative framework to enhance SPECT reconstruction toward PET-level image quality. A denoising diffusion probabilistic model (DDPM) is trained to learn the distribution of high-quality PET images. During inference, the reverse diffusion process is conditioned on measured SPECT data and CT images, enabling anatomically consistent, high-fidelity image synthesis. The method is initially assessed using 2D Zubal phantom simulations. Performance is compared with conventional maximum likelihood expectation maximization (MLEM) and kernel based methods. Both quantitative metrics and qualitative analysis demonstrate that the diffusion-based approach achieves enhanced edge preservation and superior noise suppression. The reconstructed images show closer agreement with reference data compared to MLEM and kernel based reconstructions. By avoiding explicit PET–SPECT registration, the framework mitigates spatial misalignment and associated artifacts. These results indicate that conditional diffusion sampling provides a principled and flexible approach to improve resolution and noise limitations in SPECT, with potential to improve quantitative accuracy in MRT applications.

Abstract

Diffusion models (DMs) have recently demonstrated remarkable performance in positron emission tomography (PET) image reconstruction. However, in diffusion posterior sampling, data fidelity and diffusion prior are coupled within each sampling step, which can make the balance between them sensitive to parameter selection. In this study, we propose a plug-and-play diffusion-based PET image reconstruction method (PnP-DM) that enables explicit incorporation of a DM as a prior within a penalized reconstruction. The proposed method is formulated based on Bayes’ theorem and employs a half-quadratic splitting algorithm to decouple the likelihood and prior terms into penalized reconstruction and diffusion sampling. The proposed method alternates between the likelihood and prior steps, where the likelihood step penalizes deviation from the diffusion result, while the prior step is guided by the reconstructed image from the likelihood step, thereby maintaining data consistency throughout the optimization process. We evaluated the proposed method using simulated [¹⁸F]FDG BrainWeb phantom datasets. The proposed method suppressed noise while preserving structural details and achieved higher contrast-to-noise ratio and contrast recovery coefficient than maximum likelihood-expectation maximization (EM) and maximum a posteriori-EM. These results suggest that the proposed PnP method provides diffusion-based reconstruction that enables more stable optimization and flexible balancing between data fidelity and diffusion prior.

Abstract

We present a GPU-accelerated list-mode reconstruction framework for gamma-positron brain imaging that models PET coincidences, prompt-gamma Compton events, and triple events, in which one decay produces both detected event classes. The goal is to reconstruct the radionuclide decay distribution, rather than only the positron annihilation location, by combining the sensitivity of 511-keV coincidences with the geometric constraint from a coincident prompt gamma. The framework uses UAMonteCarlo emission and detector simulations to generate sorted list-mode events and voxelized sensitivity maps for an edge-on crossed-strip CZT geometry. The test configuration contains 690 detectors on a 30-cm-diameter ring surrounding a 24-cm water sphere and models 44Sc, including emission and transport of a 1474 keV-max positron and a 1157 keV prompt gamma. Event likelihoods are evaluated for PET-only events, two-interaction prompt-gamma events, and triple events. PET likelihoods include annihilation-photon attenuation and positron range effects; prompt-gamma likelihoods include Compton kinematics, energy deposition, detector interaction type, and attenuation; triple-event likelihoods combine PET line-of-response and prompt Compton-cone constraints. The reconstruction code computes event probabilities on the GPU and applies a list-mode maximum-likelihood expectation-maximization update. Initial consistency studies with simulated 44Sc point-source data show that PET-only and prompt-only reconstructions recover point-source locations after three iterations using 2000 events per mode, while sensitivity maps and likelihood images show the tradeoff between high-sensitivity PET/prompt events and lower-sensitivity, higher-weight triple events. These results establish machinery for combined gamma-positron reconstructions and motivate tests with triple events, mixed event-type reconstructions, and further phantom studies.

Abstract

Attenuation correction (AC) is essential for quantitative PET imaging and is conventionally performed using CT-derived attenuation maps. In this work, we present an initial clinical evaluation of a CT-less AC approach based on detector backward inter-crystal scattered (BKS) coincidences on the NeuroEXPLORER (NX) PET/CT system. Two patient datasets were acquired with a full-energy-window protocol that preserved standard emission coincidences, BKS events, and (^{176})Lu background events within the same list-mode dataset. For each patient, three reconstruction pipelines were compared: CT-AC+OSEM, Lu-AC+OSEM, and BKS-AC+OSEM. Results showed that, in the higher-activity case, Lu-background data were more strongly affected by random events, leading to a noisier attenuation map, whereas the BKS-derived attenuation map showed improved quality, particularly in the skull region, and yielded PET images in closer agreement with the CT-AC reference. In the lower-activity case, BKS- and Lu-derived attenuation maps were generally comparable. A short-acquisition study further showed that BKS-derived attenuation maps remained reasonably consistent for acquisition times of 5 min or longer. These results indicate that BKS-based AC is a promising direction for CT-less PET attenuation correction and warrants further evaluation in larger studies.

Abstract

In this work, we propose an unrolled Block Sequential Regularized Expectation Maximization (BSREM) reconstruction framework in which the gradient of the regularization term is learned using a convolutional neural network. Building on our previous work, we adopt Jacobian-Free Unrolling (JFU) to enable memory-efficient training on 3D synthetic data. This training strategy backpropagates only through the final iteration of the unrolled optimization, significantly reducing memory consumption while improving training stability. We investigate the influence of relaxation strategies within the unrolled BSREM scheme and analyze their impact on convergence behavior, stability, and reconstruction quality. Experiments conducted on realistic simulated SPECT datasets and clinical  ¹⁷⁷Lu-DOTATATE patient data show that JFU-based unrolling consistently outperforms classical unrolled training for an equivalent number of parameters and memory usage. Unlike classical unrolling, which requires strong relaxation to ensure stable training, JFU achieves optimal performance with weak or no relaxation. Excessive relaxation is shown to be suboptimal, leading to reduced quantitative accuracy. Quantitative evaluations demonstrate improvements in both global image quality and organ-level activity accuracy compared to OSEM and conventional BSREM with a Relative Difference Penalty prior. As observed in our previous study using Green’s one-step-late algorithm, the proposed approach may introduce over-separation artifacts between regions of low and high activity. This behavior appears intrinsic to learning the regularization gradient rather than to the unrolling strategy itself. Overall, this work highlights the advantages of Jacobian-Free Unrolling for scalable and stable learning-based SPECT reconstruction and provides insights into the role of relaxation in unrolled optimization schemes.

Abstract

The Multiple Emission Tomography (MET) detector, a single-layer ring of LYSO:Ce scintillators, offers a versatile imaging platform combining Positron Emission Tomography (PET) with collimator-free Compton-mode Single-Photon Emission Computed Tomography (SPECT). However, Compton reconstruction for single-layer cylindrical geometries remains a major challenge due to high data dimensionality and a dense system matrix. In this work, we present a GPU-accelerated reconstruction framework based on List-Mode Ordered Subset Expectation Maximization (LM-OSEM). Building on the open-source CoReSi library, our framework introduces several contributions. i) High-Performance Architecture: we implement a vectorized PyTorch pipeline with fused broadcasting and event batching. This approach achieves O(batch size · voxels) peak GPU memory occupancy, enabling the processing of massive datasets. ii) Advanced Physics Modeling: we introduce a system matrix that incorporates a per-event Angular Resolution Measure (ARM). This model integrates Doppler broadening, energy resolution, and detector spatial resolution, allowing for optimal event weighting and noise suppression. iii) Geometry-Specific Sensitivity: we propose a novel sensitivity model tuned for the MET cylindrical geometry, accounting for solid-angle coverage and LYSO self-absorption.The framework is validated through Monte Carlo simulations of the MET prototype. We evaluate spatial resolution and noise on reconstructed images from 511 keV to 1 MeV. Spatial resolutions between 5 mm and 10 mm are achievable, confirming MET's potential for imaging high-energy exotic nuclei inaccessible to conventional gamma cameras. This work provides a scalable, fast, and physically accurate foundation for next-generation multi-modal molecular imaging.

Abstract

Total-body positron emission tomography (TB-PET) provides significantly improved sensitivity, enabling ultra-short time framing for dynamic imaging. However, the resulting low count statistics lead to increased noise and unstable time activity curves (TACs), limiting quantitative accuracy. We investigate a kernel-based reconstruction approach within a list-mode (LM)-derived synthetic framework, where a vendor reconstruction is forward projected through an event-specific LM system matrix to generate synthetic data, implicitly includes attenuation, normalisation, scatter, and randoms effects while enabling flexible algorithm design.
Spatial, temporal, and spatio-temporal kernels are evaluated for short-frame FDG data acquired on a Biograph Vision Quadra system, with features constructed from multiple dynamic intervals. Reconstructions are compared with a 10x higher count vendor reference images using spatial and temporal PRMSE. Spatially, the spatio-temporal kernel achieves the lowest error. In addition, we demonstrated that kernel methods benefit from larger neighbourhoods resulting in improved visual quality compared to synthetic PET. This was computationally feasible on consumer hardware by using region-of-interest specific reconstructions.
These results demonstrate that LM-derived synthetic kernel reconstruction provides a flexible framework for incorporating spatial and temporal priors in TB-PET. While performance is currently limited by kernel parameter selection, the observed improvements in spatial and temporal behaviour indicate strong potential for optimisation through tuning the kernel strength and feature vector design.

Abstract

Positron emission tomography provides functional images from two-photon coincidences. In human tissue, approximately 40% of annihilations are preceded by positronium formation, of which the ortho-positronium component partially decays into three photons [1]. This three-photon channel carries additional information about the local micro-environment [1], including the three-to-two-photon yield ratio as a potential diagnostic marker [3,2], and its exploitation does not require a prompt photon, making it compatible with standard radionuclides such as FDG. We propose the Three-Photon Bayesian Imaging of Ortho-Positronium algorithm, a novel event-by-event reconstruction method formulated as a Bayesian maximum a posteriori inference problem, which unifies time-based trilateration [4], energy-based reconstruction [3], and a physics-informed prior derived from the quantum electrodynamics description of ortho-positronium decay [5-6] within a single probabilistic framework. Monte Carlo simulations modelled after the Siemens Biograph Quadra scanner [7] demonstrate an improvement of a factor of two over time-based trilateration and by an order of magnitude over energy-based reconstruction. The proposed approach is compatible with existing scanners capable of registering three-photon coincidences, broadening its applicability beyond positronium lifetime imaging to standard clinical radionuclides.

Abstract

⁹⁰Y bremsstrahlung SPECT imaging remains challenging due to its broad continuous energy spectrum. In this study, we proposed and evaluated a novel event-by-event energy-dependent list-mode reconstruction algorithm for quantitative ⁹⁰Y bremsstrahlung SPECT. To enable real-time and reasonably accurate energy-dependent collimator-detector response modeling, we proposed and incorporated a True Half-Width method into the list-mode reconstruction framework. Monte Carlo simulations were performed using three energy windows: 90-160 keV, 105-285 keV, and 90-400 keV. Quantitative evaluation was performed using contrast recovery coefficient and background variability. Compared with projection-based reconstruction, list-mode reconstruction achieved the best quantitative performance using the 90-400 keV energy window, owing to accurate event-level energy-dependent system matrix modeling over the broad continuous energy spectrum. These results demonstrate that event-by-event energy-dependent list-mode reconstruction is essential for improving the quantitative accuracy of ⁹⁰Y bremsstrahlung SPECT.

Abstract

xSPECT CG (Conjugate Gradient) SPECT reconstruction uses the Mighell Chi-Square (χ²) statistic as its optimization cost function [3]. While effective under standard count conditions, the Mighell statistic is known to be suboptimal when Poisson statistics are poorly approximated at low counts. The Cash C statistic offers an alternative formulation derived from a modified chi-square for Poisson data, better suited to low-count regimes[2]. This work evaluates Cash C as a direct replacement for Mighell χ² in xSPECT CG, assessing resolution-noise trade-offs across count levels using a standard phantom. Resolution was quantified using a metric based on the approach described in [1], further developed here to handle arbitrary source-to-background contrasts. Cash C consistently achieved better noise properties at equivalent resolution compared to Mighell χ², with both algorithms converging to similar minimum resolution values 16.4 mm FWHM) at high update steps for the 8 mL sphere, confirming no loss of performance at high-count conditions. These characteristics are particularly relevant for clinical applications with limited count statistics, including cardiac SPECT [4] and emerging theranostic radiopharmaceuticals.

Abstract

In SPECT imaging, accurate modeling of the collimator–detector response (CDR) function is essential for improving image resolution and quantitative accuracy. The CDR profile depends on several physical factors whose modeling can be complex and computationally demanding, particularly for multi-photopeak datasets, since the CDR, photon attenuation, and Compton scattering are energy dependent. In this work, we investigate the modeling of the CDR function using profiles generated through Monte Carlo simulations and its incorporation into the reconstruction process using a multi-line projector approach.
Monte-Carlo simulations were processed in order to generate CDRs of a SPECT intevo system with MELP collimator using 177Lu point sources located at different distances of the imaging system. These data were combined to generate CDR reponses for both 208kev and 113kev photopeaks. A multi-lines projector was implemented to integrate these profiles and processes the corrections specific to each photopeak. Simulations of NEMA IQP phantom was conducted to assess the proposed implementation, as well as phantom data acquired with the same system and variable source to collimator distance for each projections.
The proposed model improves the CRC-versus-noise trade-off as well as the overall image quality. Although increasing the number of projection lines enhances the signal-to-noise ratio, the evaluated metrics show limited improvement beyond a certain number of lines. These results suggest that for the studied acquisition and reconstruction setup, a reduced number of lines per detector channel may be sufficient to accurately model the CDR while limiting the computational cost of the reconstruction process.
Ongoing work includes the evaluation of the proposed implementation with different collimator geometries exhibiting various septal penetration and scattering characteristics. Future investigations will focus on simulations involving high-energy photon-emitting radionuclides.

Abstract

Newly emerging radiation imaging modalities—such as Compton camera imaging of alpha-emitting radiopharmaceuticals, positronium lifetime imaging, and dynamic PET imaging—frequently operate under count-starved, low signal-to-noise ratio (SNR) conditions, where the conventional maximum likelihood expectation maximization (ML-EM) algorithm is prone to severe overfitting. While recent neural network-based reconstruction methods have achieved impressive results, they require large amounts of task-specific training data, often neglect physics constraints, and largely lack principled uncertainty quantification frameworks.
In this work, we propose a Bayesian image reconstruction algorithm incorporating a Gaussian process (GP) prior with a neural network-based kernel, combining the robustness and principled uncertainty quantification of the GP framework with the representational flexibility of neural networks. The proposed method is highly robust against measurement noise, naturally incorporates physics constraints through the forward model, supports anatomically guided reconstruction through co-registered structural imaging data, and provides a full posterior covariance characterization via variational inference. We demonstrate significant improvement in reconstruction quality over the conventional ML-EM algorithm alongside meaningful uncertainty estimates.

Abstract

Quantitative positron emission tomography (PET) requires standardized and reproducible image quality assessment to ensure comparability across scanners, protocols, and institutions. However, current workflows for NEMA-based evaluation often rely on manual region-of-interest (ROI) definition and fragmented analysis tools, limiting reproducibility and scalability. In this work, we present ChameleonIQ, an open-source, configuration-driven framework for automated PET image quality assessment. The platform supports both NEMA NU-2 2018 and NEMA NU-4 2008 protocols within a unified pipeline, providing automated ROI definition, deterministic metric computation, and structured reporting. Validation was performed using standardized IEC Body and NU-4 IQ phantoms. For the NEMA NU-2 protocol, contrast recovery ranged from 29.5% (10 mm sphere) to 55.1% (37 mm sphere), with background variability maintained between 6.1% and 7.0%. For the NEMA NU-4 phantom, recovery coefficients exceeded 0.8 for rods larger than 3 mm. The framework demonstrated high quantitative fidelity, achieving an average error in hot rod recovery of only 3.9% compared to reference manual evaluations. By providing a semi-automated pipeline that maintains this high precision while removing operator bias, ChameleonIQ facilitates efficient protocol optimization and multi-center benchmarking.

Abstract

The reconstructed spatial resolution of a PET image is determined by both intrinsic scanner properties and the chosen reconstruction algorithm. Standard NEMA measurements use point sources reconstructed with FBP and  therefore mainly characterize the scanner.  They are thus suboptimal for predicting the resolution of images reconstructed with  MLEM and point-spread-function modeling, as is common in clinical practice.

We propose a 3D-printed resolution cube with an off-center inner compartment which, combined with a mid-acquisition phantom flip and a single reconstruction, produces sharp edges at finite (2:1) contrast along three orthogonal directions. Position-, direction- and contrast-dependent edge and line spread functions (ESFs, LSFs) are then extracted using smoothing-spline fits.  As a proof of concept, the method is applied to the NeuroExplorer brain PET system at the isocenter and 10 cm off-center, with and without depth-of-interaction (DOI) modeling.

The LSFs are non-Gaussian with pronounced negative side lobes.  With DOI, the resolution is approximately 2.5 mm FWHM in all directions at both positions; without DOI it degrades radially to about 4 mm at 10 cm off-center.  Notably, differences between finite and infinite contrast are small in most configurations.

The method provides a position-, direction- and contrast-dependent characterization  of PET resolution from clinically reconstructed images and complements NEMA point-source values for image quality comparison, harmonization and partial volume correction.

Abstract

Accurate motion correction is critical for dynamic brain PET, particularly on the high-resolution NeuroEXPLORER (NX), where residual head motion can blur focal targets and bias time-activity curves (TACs). We evaluated multiple motion correction (MC) strategies using 17 multi-tracer human NX PET scans, including no motion correction (NMC), post-reconstruction registration (PRR), registration before attenuation correction (RBA), UMT-based frame-averaged correction (AVG), and UMT-based event-by-event correction (EBE). Performance was assessed using qualitative image comparison and ROI-based residual TAC variability (%SD_res) from 20-70 min. All MC methods reduced %SD_res compared with NMC. Across all tracers, subjects, and ROIs, EBE achieved the lowest residual variability (1.40 ± 1.27%), followed by PRR (1.79 ± 1.49%), RBA (1.94 ± 1.64%), AVG (2.01 ± 1.63%), and NMC (4.68 ± 4.26%). Poisson noise-floor analysis further showed that residual motion effects were tracer- and count-statistics-dependent. Overall, UMT-based EBE provided the most effective reduction of residual motion effects in dynamic NX brain PET.

Abstract

Monte Carlo scatter correction is well established as a highly accurate approach for scatter estimation in PET/CT imaging. It demonstrates superior robustness compared with model based scatter estimation methods, particularly under extreme variations in activity distribution and in the presence of patient motion, conditions under which model based approaches often produce artifacts. Despite these advantages, Monte Carlo scatter correction is computationally intensive, and does not inherently account for out-of-field scatter. In this work, we address both limitations and propose an implementation suitable for integration into routine clinical workflows. A multi-GPU implementation, as well as parallel processing with efficient data streaming significantly reduce the processing time. To handle out-of-field scatter, we separately handle regions where information about out-of-field distribution is known, and can be used for a highly precise estimation. In regions where out-of-field information is inaccessible, the effect on scatter is approximated by the measured data, while ensuring a smooth transition in which the high-accuracy region governs the final estimate. Our approach is scalable and robust across system designs ranging from a standard axial field-of-view to a long axial field-of-view, each presenting distinct considerations for scatter estimation. We demonstrate the effectiveness of our method using phantom and clinical data.

Abstract

In simultaneous PET&SPECT imaging, PET-induced down-scatter contaminates SPECT data, degrading image quality and quantitative accuracy. This study presents a correction method that estimates PET down-scatter based on PET source modeling and corrects contaminated SPECT data to suppress PET interference.
A PET&SPECT simultaneous imaging system is modeled using GATE Monte Carlo simulation. The system adopts a dual-head detector rotating around the field of view. An image quality phantom consisting of background region, cold region, and multiple PET and SPECT hot regions is simulated to evaluate cross-modality interference. PET and SPECT activities are modeled using 511 keV annihilation photons and 140 keV gamma photons. A uniform PET activity distribution is simulated to establish a voxel-wise PET down-scatter contribution model. The actual PET activity distribution is obtained from reconstructed PET images, and the corresponding down-scatter components are estimated from the model and subtracted from the SPECT projections.
Experimental results demonstrate that PET-induced down-scatter significantly degrades SPECT projections and reconstructed image quality. The auxiliary energy window (AEW) method partially mitigates this degradation, while the proposed method achieves more effective suppression, yielding images closer to a pure SPECT reference. Quantitative evaluation shows that the proposed method outperforms AEW in terms of spillover ratio and contrast, confirming its effectiveness in simultaneous PET&SPECT imaging.

Abstract

A dedicated brain positron emission tomography (PET) system utilizing multi-resolution depth-of-interaction (DOI) detectors is under development at Shenzhen Bay Laboratory. These detectors employ variable-pitch, multi-layer scintillators to optimize intrinsic spatial resolution and coincidence timing resolution while minimize parallax errors. However, the multi-resolution architecture introduces structural and detection heterogeneities that complicate quantitative image reconstruction. This study proposes an extended quantitative correction framework to address these challenges. We adapted standard component-based normalization (CBN) and single scatter simulation (SSS) via sub-layer-based geometric modeling to achieve layer-specific normalization and scatter correction. Specifically, a 3D fan-sum method was implemented to calculate crystal efficiencies by treating lines of response (LORs) within shared DOI layers as independent subsystems. Furthermore, a likelihood-based optimization framework was formulated to derive layer-specific scatter multipliers, compensating for detection biases caused by inter-layer shielding and axial sensitivity fluctuations. The framework was validated using GATE simulations. Ablation studies on uniform cylinder data confirmed the necessity of the integrated correction suite for spatial uniformity. Quantitative evaluation of a custom IQ phantom yielded contrast recovery coefficients (CRC) of 84.53%–65.13% and background variability (BV) of 1.69%–3.07% for spheres ranging from 37.0 to 13.0 mm. Reconstructed Derenzo phantom images resolved 1.0 mm rods at both central and off-center positions, validating its high spatial resolution. Finally, digital brain phantom studies indicate that the system’s imaging performance is visually comparable to the NeuroEXPLORER benchmark.

Abstract

Patient motion remains a challenge for dynamic positron emission tomography (PET) imaging of the brain, due to long acquisition times. Data-driven registration-based motion estimation for static PET using sum of squared differences or normalized cross correlation as alignment metric (while popular for their simplicity and generalization potential) are not suitable for dynamic PET, as they cannot handle the changing tracer distribution throughout the scan. Alternative methods for dynamic PET motion estimation rely on the availability of a good kinetic model or suitable training data for deep learning, reducing their generalization potential. We therefore propose a registration method based on mutual information that makes it suitable for dynamic PET, and included the motion estimates in a listmode reconstruction using event-by-event motion correction. We tested this approach for 3 tracers on 2 patient scans each.

Abstract

Head motion during PET data acquisition is a common source of artifact that degrades image quality and biases tracer uptake quantification. We hypothesize that motion correction improves the performance of a machine learning (ML) classifier for Alzheimer’s disease versus cognitively normal subjects using brain PET images alone. A ResNet10 feature extractor combined with a Support Vector Machine classifier was applied to two cohorts encompassing a total of 301 ¹¹C-UCB-J and ¹⁸F-FDG PET. On motion-corrected images, the classifier achieved AUROC of 0.96 (¹¹C-UCB-J) and 0.94 (¹⁸F-FDG), with accuracy increases of 5% and 10%, respectively, when motion correction was applied. Our results highlight the critical need for efficient motion correction solutions to enable effective ML analysis of PET AD imaging.

Abstract

This work develops and evaluates CT-less data correction methods for quantitative cardiac PET by combining deep learning-based attenuation correction (DLAC) with energy-based scatter (EBS) estimation. Robustness was assessed in three human FDG studies spanning a wide BMI range (24-50 kg/m²). DLAC showed stable performance across BMI groups, with mean myocardial SUV biases below 5% relative to CT-based AC (with single scatter simulation). For a low BMI case, the DLAC+EBS method resulted in low myocardial SUV bias (mean: 1.7%, P₅-P₉₅: [–2.1%, 10.8%]) relative to standard CT-based reconstruction. To assess generalization, the DLAC model (originally trained on supine non-cardiac human data) was extended to cardiac ⁸²Rb studies in pigs using transfer learning. These pig studies served as a surrogate for evaluating whether the network could be adapted to other imaging conditions (cross-tracer, cross-species, and other patient positionings, e.g., seated cardiac imaging) with limited re-training. With limited additional pig data (n=8), relative biases in mean myocardial SUV decreased to within 2%. Overall, these findings support CT-less data correction for cardiac PET systems.

Abstract

Existing software solutions for PET quantification in neurological disorders remain fragmented and lack comprehensive end-to-end pipelines suitable for multi-site studies involving heterogeneous tracers, and scanners, including Total-Body PET systems. We present AmyPET, an open-source toolkit for automated end-to-end preprocessing and quantification of static and dynamic PET data directly from raw DICOM images. The framework supports processing in both native and MNI image spaces and integrates multiple brain atlases, motion correction, partial volume correction, and harmonised analysis within a single modular environment.

AmyPET comprises four integrated processing engines for static and dynamic PET analysis: (1) an alignment engine for moderate motion correction based on a novel displacement metric, (2) spatial normalisation and tissue segmentation, (3) regional sampling and image extraction, and (4) a post-processing engine enabling kinetic modelling, Centiloid quantification, and additional quantitative analyses, including dual-time-window acquisitions. Automated quality-control outputs are generated for all processing stages. The framework has been calibrated against the GAAIN reference datasets for amyloid PET and is extensible to tau and TSPO tracers.

Validation was performed using multiple amyloid PET cohorts, including test–retest studies, an amnestic mild cognitive impairment cohort, and the AMYPAD Prognostic and Natural History Study comprising more than 1500 scans. AmyPET demonstrated high repeatability (within-subject coefficient of variation, wCV=3.5%; ICC>0.98) and strong agreement with established clinical and research pipelines (R²>0.96), supporting its use as a robust and scalable framework for quantitative neuroPET imaging.

Abstract

Recent increases in computational power as well as new tools being more readily available have led to increased interest in Monte Carlo scatter estimation. However, Monte Carlo scatter simulation is still computationally expensive. This is exacerbated by the fact that model-based scatter estimation requires multiple iterations of scatter simulation and image reconstruction. We propose to combine the current clinical standard (analytic single scatter simulation and tail-fitting) with Monte Carlo scatter simulation.

Abstract

The Alpha-SPECT Mini system integrates SPECT imaging with CT-based attenuation correction, providing a platform for quantifying in vivo distributions of alpha-emitting radiopharmaceuticals and daughter radionuclides. In this study, three Ac-225 filled syringes with different volumes, 0.5, 4.2, and 18.0 mL, respectively, were scanned with Alpha-SPECT Mini. The ground truth activities were[YD1] 26.12 µCi, 21.62 µCi, and 18.80 µCi from the smallest to the largest syringe, respectively. SPECT listmode acquisition was performed for 20 minutes for each syringe, together with an 80 kV CT scan for attenuation correction. SPECT images were reconstructed for the following photopeaks: 100 keV for Ac-225, 218 keV for Fr-221, 440 keV for Bi-213, and 117 keV for Tl-209. Activities of Ac-225 and these daughter radionuclides, were then estimated from SPECT images using CT defined ROIs and compared with ground truth.

The results showed that quantification accuracy varied with the radionuclides and syringe sizes. Fr-221 and Bi-213 showed relatively lower errors (<5%), while Ac-225 and Tl-209 exhibited larger deviations. The higher errors likely due to the low abundance of low energy (100 keV for Ac-225 and 117 keV for Tl-209) photopeaks and interference from downscatter and collimator penetration from high energy photons. Tl-209 also exhibited larger deviations with a size-dependent trend, suggesting sensitivity to source geometry and scatter correction. Overall, the Alpha-SPECT Mini system showed potential for quantitative imaging of Ac-225 and daughter radionuclides. Future work includes further calibration and implementation of scatter correction.

Abstract

PETcoil Gen II is a novel MRI-compatible time-of-flight (TOF) PET system for simultaneous PET-MRI brain imaging. Current calibration uses an annulus phantom to correct time skew on a per line-of-response (LOR) basis, leading to a large number of correction factors (order $n^2$). We propose a convex-optimization calibration method to calculate and store per-channel (order $n$) time skew offsets using rod phantoms. We evaluate the method using simulated annulus and rod phantoms with known applied skew offsets and point-source validation. The proposed per-channel approach achieves a mean absolute timing error of 19.9 ps (annulus-based) and 18.8 ps (rod-based) compared to 142 ps for the per-LOR method. When applied to point-source correction, the proposed approaches improve coincidence time resolution (CTR) performance by +18.7\% and 13.4\% using annulus and rod phantoms respectively, whereas the per-LOR approach resulted in a -70.0\% degradation relative to uncorrected data. Results demonstrate improved calibration accuracy and robustness in using rod-based phantoms. Future work will validate these methods experimentally and assess the quality of reconstructed images.

Abstract

Digital breast tomosynthesis (DBT) improves lesion detection compared with 2D digital mammography (DM), but its large number of reconstructed slices leads to increased reading time and workload for radiologists. As a potential solution, thick-slab image generation can be used to combine adjacent reconstructed slices into fewer representative slab images. However, conventional slab generation methods typically rely on simple slice-combination strategies, such as maximum intensity projection (MIP) or average intensity projection (AIP) of adjacent slices, which may reduce image sharpness and obscure diagnostically relevant details. To overcome these limitations, we propose a new thick-slab image generation method. For each slab, the proposed method adaptively blends a representative slice with the corresponding MIP image. The representative slice is selected from the DBT slices within the slab based on linear structure information obtained using the oriented bins method. The blending map is generated by applying a 1D high-pass filter to the MIP image along the X-ray source travel direction. Using a numerical phantom, the proposed method was compared with conventional MIP- and AIP-based slab generation methods and showed improved image sharpness and structural preservation. These preliminary results suggest that the proposed thick-slab image generation method has the potential to reduce DBT review time by decreasing the number of images to be reviewed without substantial loss of diagnostically relevant information.

Abstract

STIR is a widely used open-source framework for PET and SPECT image reconstruction; here we explored PyTorch as an alternative array backend for STIR. We replaced the STIR arrays with a tensor-based variant while preserving the existing reconstruction framework and backwards compatibility. Correctness was assessed using existing and newly developed unit tests. Performance was evaluated with dense arithmetic benchmarks, stir_timings, and stir_math. For image-based operations, the tensor backend accelerated in-memory processing, with the largest gains observed for larger volumes. In projection-data workflows, initial performance was limited by layout-conversion and storage overhead, but targeted optimization substantially reduced these costs. In projection-data stir_math tests, the Torch-backed implementation was faster than the native STIR Array backend in end-to-end execution, with gains of roughly 7% to 23% depending on the operation. Initial GPU benchmarking further showed that the tensor backend can exploit consumer GPU hardware for dense arithmetic, achieving an approximate 3.7x speedup over CPU in the tested workload. However, projector-style kernels based on sparse, irregular scalar access did not yet benefit consistently from direct backend substitution, indicating that these critical paths require further redesign.

Abstract

Accurate quantification of pulmonary PET activity is often limited by spill-in contamination from adjacent tissues. Although reconstruction-based partial volume correction techniques effectively reduce resolution-related blurring, residual spill-in can remain significant in pulmonary analysis due to respiratory motion and proximity to high-uptake extrapulmonary structures. Post-reconstruction correction methods often assume homogeneous activity within clearly defined anatomical compartments, assumptions not satisfied in inflammatory whole-lung volumes. As a result, manual refinement of the lung boundary is often necessary but is also time-intensive. Here, we compare three automatic post-reconstruction spill-in correction strategies in 18F-FDG PET scans of tuberculosis-infected macaques, using manually refined lung masks as a reference. Two methods directly adjust PET values within the lung: the first estimates spill-in from local background activity, while the second applies a more conservative correction using extrapulmonary hotspots. The third method automates conventional morphological mask refinement by adjusting the lung boundary. The value-based spill-in correction methods achieved higher agreement with manual refinement (mean log ratio 0.032 and 0.116) but showed greater sensitivity to boundary perturbation, whereas morphological correction showed lower boundary sensitivity but systematically underestimated uptake (mean log ratio -0.408). These findings suggest that localized spill-in correction methods may provide a practical alternative to manual mask refinement in 18F-FDG PET pulmonary imaging by reducing extrapulmonary contamination while preserving anatomical boundaries.

Abstract

Wearable brain PET enables neuroimaging during free movement, but head motion causes severe degradation. Compact scanners are particularly sensitive because their small detector rings (∼100 mm) amplify angular displacements at the event level. We compare five data-driven motion correction approaches on a wearable scanner using programmed Derenzo phantom motion, with all-data 3-DOF control groups isolating event retention from algorithmic accuracy. The hybrid phase-correlation with mutual-information 6-DOF method achieves the best resolution–noise trade-off—FWHM 4.68 mm vs. 5.42 mm for all-data 3-DOF and 15.15 mm for uncorrected—with SNR of 4,586. Results confirm that 6-DOF event-level correction is essential for compact-geometry wearable PET.

Abstract

Radiographic image quality is degraded by X-ray scatter, which reduces contrast and obscures fine anatomical structures, especially in thick body regions. This study proposes a calibration-free software scatter correction method for radiographic imaging based on pixelwise alpha blending of dark channel prior and bright channel prior. The observed radiograph was modeled as attenuated primary radiation plus additive scatter. Scatter intensity was estimated using self-guided filtering, and transmission maps were generated from dark and bright channel priors. Because dark channel prior can fail in bright anatomical regions such as bone, an attenuation-based alpha weight was used to adaptively combine the two transmission maps. The restored radiograph was obtained using the estimated scatter intensity and refined transmission map. The method was evaluated using in vivo animal radiographs acquired with a mobile radiographic imaging system at 100 kVp and 40 mA. Rabbit, beagle dog, and mixed-breed dog images were used to assess performance across different body sizes and scatter conditions. Image quality was compared among the observed image, dark-channel-prior-only correction, and the proposed alpha-blended correction. The proposed method improved anatomical visibility and reduced overcompensation artifacts in bright bone regions. The contrast-to-noise ratio increased from 0.85 in the observed image and 1.78 with dark-channel-prior-only correction to 2.18 with the proposed method.  These results demonstrate that pixelwise blending of dark and bright channel priors can improve radiographic contrast and fine-structure visibility without additional hardware or calibration.

Abstract

Timepix4 hybrid pixel detectors, originally optimized for single-particle tracking applications, are increasingly attractive for medical X-ray imaging owing to their large area (∼7 cm2), 4-side buttability and appreciable data-driven throughput (3.6 × 106 hits/mm2). Adapting them to imaging workflows, however, demands a substantial shift in data handling. A notable challenge is presented by coping with intensity-dependent data
rates, and efficiently aggregating hits and clusters into fixed-size structures under heavy loads. This contribution describes a collection of stackable software tools that adapt an established event-driven tracking framework for Timepix-based systems to X-ray imaging, organized over a modular processing pipeline. The presented platform includes energy-band splitting of hit and cluster streams and corrections for commonly seen effects:
internal fluorescence of CdTe sensors, per-pixel high-energy correction, cluster-volume energy loss, non-uniform flat field response and beam hardening. The composition order of these operations is treated as a user-driven parameter, permitting construction of dedicated processing workflows depending on the imaging task at hand. Fluorescence recovery, high-energy correction, and cluster volume correction are already operational
and have been tested experimentally; the remaining features are under active development. The full platform is expected to be applied to representative medical imaging tasks over the summer of 2026, with quantitative results to be presented at the conference.

Abstract

Phase-shifting sub-pixel sampling can improve the spatial granularity of compact gamma-ray detector modules without using analog charge sharing or Anger-logic interpolation.In this work, the phase-shifting concept is extended toward depthof-interaction (DOI) estimation by combining controlled detectordisplacement with double-sided MPPC readout. A monolithic scintillator is coupled to two readout sides, denoted as A and B, and the DOI-sensitive observable is defined by the signal ratio A/(A + B). The detector is translated laterally by 1 mm steps relative to a narrow slit so that the effective irradiation position changes within the scintillator thickness. Coincidence signals are then analyzed using A-B correlation plots. The measured slopes vary systematically from the B-side region to the A-side region,
indicating that interaction depth information is preserved in the double-sided signals. These results demonstrate the feasibility of integrating phase-shifting sampling with DOI estimation for compact SPECT or Compton camera detector modules.

Abstract

Micro-single photon emission computed tomography (µSPECT) is constrained by an inherent sensitivity-resolution trade-off imposed by the collimator. High-sensitivity collimators mitigate statistical noise but blur spatial details, whereas high-resolution collimators limit count sensitivity, thereby amplifying noise. To bridge this trade-off and overcome both low sensitivity and poor resolution, we propose and compare three frameworks for joint µSPECT denoising and super-resolution: a single Joint U-Net, a sequential Cascade U-Net, and a Dual-Encoder U-Net utilizing task-specific encoders and a shared decoder. Models were trained on paired mouse datasets to map noisy low-resolution acquisitions (10-50% of standard counts) to high-resolution targets, acquired with two different collimators. The Dual-Encoder architecture most effectively balanced the competing objectives of noise suppression and edge preservation by extracting distinct representations before merging. Compared to the input, the model achieved a 32% reduction in overall RMSE and an 8% improvement in SSIM. Furthermore, it demonstrated superior resolution recovery relative to the Joint and Cascade approaches, reducing edge map and spectral errors by 27% and 21% compared to the input, respectively. Visually, it was the only model capable of resolving rod peaks in a Derenzo phantom. This approach enables significant reductions in acquisition time or dose while maximizing precision for preclinical quantification.

Abstract

PET reconstruction aims to transform raw sinograms data into images using analytical, iterative or deep learning approaches. In addition to the ill-posedness of the problem, inherent Poisson noise in sinograms makes the problem more difficult, especially in low-dose contexts.
Using deep learning for image denoising and reconstruction has shown promising performances. However, supervised training requires large sets of paired data with corrupted images and their associated ground truth, which is hard or even impossible to acquire in the medical field, especially in emission tomography.
In this work, we adapted Noise2Noise, a self-supervised framework, to PET imaging in order to learn how to denoise images from noisy measurements only.

Abstract

Modern PET reconstruction methods increasingly rely on advanced optimization strategies and deep learning-based priors. However, integrating such approaches into established reconstruction toolkits remains challenging due to limited interoperability with high-level programming environments. In this work, we present a framework that enables Python integration within the CASToR reconstruction platform, allowing direct interaction between C++ reconstruction routines and Python-based modules. As a proof of concept, we implemented an end-to-end PET reconstruction approach incorporating the Deep Image Prior (DIP) and a forward projection model. This example illustrates how the proposed integration facilitates the use of modern Python libraries for deep learning within a physics-based reconstruction framework. This work opens the door to rapid prototyping and deployment of advanced reconstruction algorithms within CASToR using the Python ecosystem.

Abstract

High-resolution PET systems require accurate DOI calibration to reduce parallax errors. Conventional DOI calibration methods rely on lateral detector access and collimated irradiations, resulting in complex and time-consuming experimental setups. In this work, we investigated the use of an analytical DOI model and a multilayer perceptron (MLP) network for DOI calibration and DOI resolution estimation using only detector front-face irradiations. The methods were evaluated on a dual-ended-readout 4 × 4 detector panel. Crystal-wise DOI calibration estimates obtained using the analytical model and the MLP showed good agreement with ground-truth measurements, with median calibration RMSE values below 0.6 mm and 0.7 mm, respectively. However, the MLP yielded superior DOI resolution prediction, with consistent mean DOI resolution estimates of approximately 2.0 mm.

Abstract

Radiomics is used to extract high-dimensional features from medical images such as computed tomography. However, radiomic features are susceptible to acquisition protocols in a multi-center setting. This results in domain shift in data distribution even when the patient anatomy is the same. This can increase intraclass variability and lead to algorithms learning protocol-specific artifacts rather than clinically relevant imaging information. In this work, we developed a virtual imaging trial (VIT) of pancreatic ductal adenocarcinoma (PDAC) patients scanned across multiple scanners and protocols to quantify protocol-induced radiomic variability post harmonization. A virtual patient cohort of 50 patients was generated using the XCAT. PDAC tumors extracted from clinical cases were implanted in individual virtual patients. DukeSim was used to simulate Siemens clinical scanners. For each scanner, two clinically representative acquisition protocols were designed by varying tube current, pitch, and reconstruction kernel. Radiomic features were extracted from tumor regions and harmonized using scanners and protocols as batch effect. Scanner and protocol dependency of each feature was assessed using Wilcoxon and Kruskal–Wallis tests, respectively. Feature robustness was quantified using intraclass correlation coefficient. Protocol-related batch effects were prominent in the non-harmonized data, with 85.98% of radiomic features dependent on protocols. Protocol-based CovBat achieved the best overall performance, reducing protocol-dependent features to 19.63% and increasing reproducible features by 3.8% post harmonization. Therefore, the proposed VIT-based framework can be used to evaluate harmonization methods and assess radiomic feature robustness before developing clinical decision-support models.

Abstract

Prostate-specific membrane antigen (PSMA) PET/CT enables standardized quantitative tumor burden assessment, yet routine SPARC-aligned metrics, molecular tumor volume (MTV), total lesion activity (TLA), and total lesion fraction (TLF), show limited predictive value for early biochemical response. Allometric scaling laws (TLA ∝ MTV^β) govern metabolic activity superlinearly (β > 1) across multiple cancers, but their application to PSMA PET outcome prediction remains unexplored. We hypothesized that each patient's deviation from the cohort-level scaling line, the Distance from the Scaling Law (DSL), encodes biologically distinct information beyond raw burden metrics and predicts PSA50 biochemical response (PCWG3: ≥50% PSA decline). In 103 prostate cancer patients with PSMA PET/CT and paired PSA measurements from a real-world clinical cohort, we fitted the allometric law (TLA = α × MTV^β) and computed per-patient DSL residuals. The cohort exhibited a superlinear scaling law: TLA = 0.574 × MTV^1.157 (R² = 0.782, p < 0.0001) consistent with literature values (β ≈ 1.22–1.28). DSL ranked 4th among 12 candidate features in univariate discrimination (AUC = 0.668), outperforming standard SPARC burden metrics including TLF (0.589), TLA (0.588), and MTV (0.523). Non-responders showed lower DSL than responders (mean −0.243 vs +0.018; rank-biserial r = 0.336), indicating hypoactive PSMA uptake relative to tumor volume as a failure signature. Nine cost-sensitive classification strategies under 5-fold stratified cross-validation reached balanced accuracy of 0.57–0.64, constrained by extreme class imbalance (93.2% responders). These findings establish DSL as a physics-grounded, automatically computable biomarker that outperforms conventional burden metrics and is deployable within SPARC workflows.

Abstract

We propose a flocking‑based framework for image structure analysis inspired by the Cucker–Smale model, interpreting pixels as interacting agents whose alignment patterns reveal self-organized regions. Here, we model each pixel as coupling with its neighbors through distance‑weighted interactions, and identify structural boundaries via a normalized measure of local misalignment. This yields an edge‑aware representation that suppresses intra‑region variation while emphasizing transitions between self-organized flocks. We evaluate the method on diverse medical imaging data, including brain MRI, retinal fundus images, and microaneurysm patches. Across modalities, the approach robustly captures both large‑scale anatomical boundaries and fine pathological details, adapting to different contrast regimes without requiring reference images. Quantitatively, strong‑edge masks achieve edge/non‑edge contrast ratios of ~66 in MRI, ~11 in fundus images, and ~4.4 in microaneurysm patches, with edge fractions ranging from 0.06 to 0.15—demonstrating high structural selectivity and reduced sensitivity to noise.

Abstract

Multicenter SPECT/CT radiomics is highly sensitive to variations in imaging protocols, reconstruction settings, and scanner configurations, limiting reproducibility and masking biological patterns. To address this challenge, we propose a harmonized latent radiomics framework for robust patient-level modeling in theranostic imaging. A retrospective multicenter dataset of ¹⁷⁷Lu-DOTATATE SPECT/CT scans from three institutions was analyzed. A total of 93 IBSI-compliant lesion-based radiomic features were extracted and harmonized using ComBat to reduce inter-center variability. Principal component analysis was then applied to generate a compact latent representation, followed by unsupervised clustering for lesion phenotype discovery. Patient-level representations were constructed by aggregating lesion embeddings. In addition, a perturbation strategy was implemented to assess latent-space stability and responsiveness. ComBat substantially reduced center-related variability, with η² values decreasing from 0.75 to near-zero levels (0.000–0.006). The latent representation demonstrated improved biological interpretability, with stronger association to intensity-related characteristics (0.351 to 0.956) and reduced dependence on texture-related confounding (0.967 to 0.528). Clustering results showed that lesion grouping was primarily driven by intrinsic phenotypic patterns rather than acquisition center. Perturbation analysis further revealed low phenotype transition rates (<10%) with gradual centroid shifts, indicating stable yet responsive latent behavior. The proposed framework combines harmonization and latent representation learning to improve robustness in multicenter radiomics and provides a promising foundation for digital twin modeling in theranostic imaging.

Abstract

Accurate dosimetry for ultra-high dose rate (FLASH) carbon ion irradiation is challenging because ionization chambers can exhibit dose rate dependent recombination and charge-collection inefficiency. Alanine/ESR dosimetry is a passive technique for high dose verification, but its response under FLASH conditions remains unclear. This study evaluated dose rate dependent alanine/ESR response by comparing calibration curves under conventional and FLASH irradiation. Alanine pellets were irradiated with a 290 MeV/u carbon ion beam at 10–100 Gy. Measurements were performed at SOBP centers under conventional irradiation at 0.09 Gy/s and FLASH irradiation at 174–265 Gy/s. An Advanced Markus chamber provided the reference absorbed dose to water. Monte Carlo simulations examined possible geometry-related dose perturbations and beam-quality effects. Under conventional irradiation, the ESR response was highly linear, and reconstructed doses agreed with the reference dose within 2%. However, applying this calibration to FLASH-irradiated alanine produced dose dependent deviations, with errors up to 28.6% in the low dose region and approximately 3–4% at higher doses. Simulations showed that chamber related scatter and positional dose differences at the alanine locations were below 1%, indicating that the discrepancy was not mainly caused by the irradiation geometry. A FLASH-specific calibration restored agreement within 1.31%, with an approximately 7.4% lower slope than the conventional calibration. These results indicate that alanine/ESR response is modified under FLASH irradiation, likely due to combined FLASH delivery and high-LET beam-quality effects. Therefore, conventional carbon ion calibration shows limited direct transferability to FLASH conditions, requiring FLASH-specific calibration for accurate dose verification.

Abstract

Accurate dose delivery in proton therapy requires reliable control of the proton range to ensure that the beam deposits its energy at the planned treatment position. However, the proton range is highly sensitive to tissue heterogeneities, anatomical changes, and patient positioning errors, making in vivo range verification an important component of proton therapy quality assurance. Positron emission tomography (PET) provides a non-invasive approach for range assessment by detecting proton-induced positron-emitting nuclei generated in the irradiated material. In this work, we demonstrate preclinical proton range verification using a dedicated in-room All-Digital PET system installed in a clinical proton treatment room. The system was evaluated using polymethyl methacrylate (PMMA) phantoms irradiated with a 130 MeV proton pencil beam. Controlled range perturbations were introduced using intact, 1 mm air-gap, and 4 mm air-gap phantom configurations. PET data were acquired for 300 s following irradiation, and one-dimensional activity profiles were extracted. At 130 MeV, the PET-derived activity profiles exhibited distinct distal fall-off shifts that correlated with the introduced air gaps. These results demonstrate the capability of the system to resolve millimeter-scale range perturbations under the tested phantom conditions and provide experimental support for future preclinical animal studies and clinical in-room proton therapy verification.

Abstract

Proton therapy is recommended as the "new standard-of-care option for patients with head and neck oropharyngeal cancer". Still, the absence of a monitoring tool equivalent to electronic portal imaging hinders its upmost potential, and forces the application of safety margins and field-patching, thereby increasing the dose to normal tissue. Prompt gamma-rays produced along the proton track can cover this shortfall, since their production cross section is strongly correlated with the proton dose depth profile. We present the first experimental proof of principle of proton range verification with a single-channel prompt gamma-ray detector located behind the treated area, in coaxial geometry. Simple heterogeneous targets were irradiated with 1e8 protons at clinically realistic energies at beam currents of 1.6 nA. The mean statistical precision was 1 mm (sigma). The absolute experimental error with respect to the ground truth predicted by a Monte Carlo simulation was 2 mm on average. These findings encourage further investigations with anthropomorphic phantoms and realistic treatment plans, and could open up the door for proton range verification with a single scintillation detector, that may have better prospects of integration into rotating gantries and compact treatment rooms than collimated multi-channel systems.

Abstract

Positron emission tomography (PET) has been widely investigated for range verification in carbon-ion radiotherapy (CIRT). All-Digital PET, based on multi-voltage threshold (MVT) direct digital signal acquisition with high count-rate capability, represents a promising new approach for in-beam monitoring. However, its application in CIRT remains largely unexplored, and experimental validation under realistic irradiation conditions is still lacking. In this study, we report the preliminary experimental validation of an in-beam All-Digital PET system for range verification in CIRT. A discrete 7x7 raster-scanned carbon-ion field (261.03 MeV/u) with a spot spacing of 5 mm and a delivery of 2000 MU per spot was delivered to a polymethyl methacrylate (PMMA) phantom (6x6x15 cm³). To simulate localized anatomical heterogeneity, a cylindrical air cavity with a diameter of 1 cm was embedded at a depth of 5 cm from the beam entrance, introducing an approximate local range shift of 10 mm. Two-dimensional distal range maps successfully revealed the irradiation field and clearly identified the cavity boundary. Within the region of interest, the measured localized range shift was 9.14 mm, with a residual error of only 0.86 mm compared with the ground truth. These results demonstrate the feasibility of All-Digital PET for accurate in-beam range verification under clinically relevant raster-scanned carbon-ion delivery conditions and support its potential for future CIRT applications.

Abstract

Targeted radionuclide therapy employs therapeutic radiation-producing radionuclides bound to targeting agents to selectively damage cancer cells. Recent interest has expanded beyond conventional beta emitters to include alpha emitters and Auger electron emitters. We investigated the therapeutic potential of the ¹³⁴Ce/¹³⁴La pair by comparing its DNA damage induction with commonly used radionuclides. A simplified human fibroblast cell model was implemented using Monte Carlo track-structure simulations to evaluate single-strand and double-strand DNA break yields. Particular attention was paid to the various spatial distributions of the radiation source. In this simulation, radionuclides were distributed at the cell surface, cytoplasm, and nucleus surface; both physical and chemical stages of radiation interaction were modeled, including indirect damage mediated by hydroxyl radicals. The results show that double-strand break yields increase as radionuclide localization approaches the nucleus, while single-strand break yields remain relatively stable. Indirect effects dominate DNA damage formation across all radionuclides and spatial distributions. The ¹³⁴Ce/¹³⁴La pair demonstrates DNA damage yields comparable to established beta and Auger electron emitters under all tested conditions. Although these findings do not directly translate to therapeutic efficacy, they indicate that this radionuclide pair has competitive radiobiological effectiveness and warrants further investigation for potential application in targeted radionuclide therapy.

Abstract

Particle therapy with protons and carbon ions offers conformal dose delivery, but its clinical benefit strongly depends on accurate verification of the beam range inside the patient. The TIARA (Time-of-Flight Imaging ARray) detector has been developed to address this challenge through the exclusive measurement of Prompt-Gamma (PG) Time-of-Flight (TOF). The system combines a fast plastic beam monitor with arrays of Cherenkov gamma detectors based on PbF₂ crystals coupled to SiPMs, enabling sub-nanosecond timing performance and real-time treatment verification.

An 8-channel prototype has been tested with proton and carbon-ion beams delivered by cyclotrons, synchro-cyclotrons and synchrotrons. Measurements with 63 MeV protons at the MEDICYC cyclotron demonstrated a range accuracy of 3.3 ± 0.1 mm (2σ) for irradiation spots of about 2×10⁷ protons, corresponding to roughly 3200 detected PG events. Experiments with carbon ions at the CNAO synchrotron achieved a range accuracy of 4.74 ± 0.36 mm (2σ) at clinical intensity by grouping four spots of 2.4×10⁶ ions each, with approximately 5600 detected PGs. These results confirm the high sensitivity of the TOF approach for proton and carbon-ion therapy.

Beyond range monitoring, TIARA implements the Prompt Gamma Time Imaging (PGTI) method to identify anatomical changes occurring between treatment planning and dose delivery. An iterative reconstruction algorithm was developed to merge the TOF information from multiple detector modules and reconstruct both the PG emission profile and the proton beam average velocity evolution inside the patient. The method was validated using a RANDO head phantom irradiated at the Proteus One accelerator under two anatomical configurations simulating an empty and filled sinus cavity. The reconstructed PG distributions enabled localisation of the induced anatomical change, demonstrating the potential of TIARA for adaptive particle therapy and in vivo tissue identification.

Abstract

Accurate in-vivo range verification is essential for realizing the full dosimetric advantage of carbon-ion therapy [1]. Secondary electron bremsstrahlung (SEB), a low-energy photon emission produced along the beam trajectory inside the patient, has attracted attention as a prompt signal for non-invasive range monitoring [2]. While our prior work demonstrated the principle using a pinhole collimator, the limited collection efficiency of a pinhole aperture imposes a fundamental constraint on range estimation precision [3]. A knife-edge slit geometry offers a one-dimensional focusing advantage that substantially raises the collected photon flux compared to the pinhole-based approach. In this study, a knife-edge slit camera equipped with a Ce:GAGG scintillator array was constructed and evaluated through irradiation experiments with a 290 MeV/u carbon-ion beam at the Heavy Ion Medical Accelerator in Chiba (HIMAC). The camera employed a tungsten knife-edge slit collimator with a slit width of 1 mm and a thickness of 48 mm, positioned 490 mm from the beam axis with a collimator-to-detector distance of 49 mm. The Ce:GAGG scintillator array consisted of 1.5 × 1.5 × 1 mm³ crystals arranged in a 16 × 16 matrix with a pitch of 1.6 mm. One-dimensional SEB profiles were measured over a range of beam range conditions by varying the range-shifter thickness, and representative results are presented. The acquired distributions exhibited a systematic dependence on beam range, with a 5 mm increment reliably distinguished by visual inspection. Quantitative range estimation yielded an uncertainty of 2.89 mm for 10⁸ ions; further details will be presented at the conference.

Abstract

Proton therapy requires accurate in vivo dose verification because proton range uncertainties may compromise treatment precision. This study aimed to develop and evaluate a deep learning-based method for three-dimensional (3D) dose reconstruction using positron emission tomography (PET) in proton therapy.

A dual-panel PET system with time-of-flight and depth-of-interaction (TOF-DOI) capability was modeled using GATE Monte Carlo simulations. PET activity images and corresponding ground-truth dose maps were generated for single pencil beams with energies across 60–120 MeV. A U-Net-based neural network was developed to reconstruct the 3D dose distributions from reconstructed PET activity images. Quantitative evaluation was performed using 3D gamma passing rates (GPRs) and distal range metrics (R₅₀ and R₉₀).

The proposed method accurately reconstructed 3D dose distributions while reduced the impact of limited-angle artifacts inherent to the dual-panel PET geometry. GPRs exceeded 99% under the 3%/3 mm criterion and 98% under the 2%/2 mm criterion for all validation datasets. Submillimeter accuracy was achieved for distal range verification, with maximum deviations of 0.15 mm for R₅₀ and 0.11 mm for R₉₀.

The results demonstrate the feasibility of using a dual-panel PET system with a deep learning method for accurate 3D dose reconstruction and proton range verification, supporting its potential in proton therapy quality assurance.

Abstract

In-beam PET is a promising technique for range verification in particle therapy because positron emitters are produced along the beam path through nuclear fragmentation reactions. Over the past years, QST has developed a human-scale OpenPET system that enables three-dimensional PET imaging during and/or immediately after carbon-ion therapy. In parallel, LMU has developed an analytical approach based on a pencil beam algorithm for irradiation-plan-based PET image prediction. In the present study, we combine these two developments toward the clinical implementation of OpenPET for range verification in carbon ion therapy. The approach involves comparing measured OpenPET images with predicted PET images.
Phantom experiments were performed at HIMAC. OpenPET data were acquired during and after irradiation. To simplify PET image prediction, the present analysis was limited to the post-irradiation period, and list-mode data acquired from 0 to 600 s after the end of irradiation were extracted and reconstructed as the measured OpenPET image. A predicted PET image corresponding to the same 0-600 s post-irradiation period was generated from the irradiation plan and PET measurement conditions by modeling positron-emitter production and physical decay. 
The predicted PET images showed an activity distribution corresponding to the planned irradiation fields, and the comparison indicated reasonable agreement between the predicted and measured images. These results demonstrate the feasibility of combining OpenPET with the irradiation-plan-based PET prediction, and they support the potential of the clinical use of OpenPET for range verification in carbon-ion therapy.

Abstract

Accurate quality assurance (QA) is essential in proton therapy, where the steep dose gradient of the Bragg peak requires precise knowledge of beam position to spare healthy tissue while ensuring full tumor coverage. This work presents the experimental evaluation of easyPET.3D, a compact and cost-effective preclinical PET scanner, as a potential tool for pretreatment QA in proton therapy using phantoms.Homogeneous PMMA phantoms were irradiated at the Paul Scherrer Institute (PSI) Gantry 1 facility with proton energies ranging from 138 to 195 MeV, delivering up to 6.4 × 10¹¹ protons per irradiation. The resulting 3D β+ activity distributions were reconstructed and analyzed to characterize both longitudinal range and transverse beam profiles. A temporal convergence study determined the minimum acquisition time required for stable range estimation. In addition, a controlled displacement experiment, with longitudinal offsets of 1, 2, 3, and 5 mm, was performed to assess the system’s ability to detect range variations.Transverse analysis revealed energy-dependent beam broadening, with mean FWHM values decreasing from 12.3 mm at 138 MeV to 9.6 mm at 195 MeV, and a characteristic widening near the distal end consistent with proton stopping at the Bragg peak. Range estimation converged within 10–12 cumulative frames (~30–35 min). Sensitivity tests showed that the system could detect longitudinal phantom displacements as small as 1 mm, with a linear response and a systematic overestimation of 13%, attributed to manual positioning uncertainties.These results demonstrate that easyPET.3D is a promising and accessible solution for pretreatment QA in proton therapy workflows.

Abstract

Multi-time-point SPECT/CT is used for patient-specific dosimetry after Lu-177-PSMA therapy. However, inter-scan motion, anatomical deformation, and time-dependent uptake variations can compromise volume-of-interest (VOI) activity extraction and time-integrated activity (TIA) estimation, thereby requiring accurate inter-time-point SPECT registration. This study proposes a structure-weighted CT-guided registration method for longitudinal SPECT/CT. Organ-specific dosimetric priorities and an uptake-derived activity-concentration term were integrated into a spatial weight map, which was incorporated into the CT registration similarity metric to emphasize structures with greater dosimetric relevance. The resulting CT-derived transformation was applied to SPECT images using physical-coordinate inverse mapping. Compared with Elastix and an unweighted baseline, the proposed method improved SPECT VOI Dice similarity coefficient (DSC) from 0.82 with Elastix and 0.87 with the unweighted baseline to 0.90. It also reduced fitting-residual-based TIA normalized root-mean-square error (nRMSE) across all evaluated structures, with additional reductions relative to the unweighted baseline of 15.8% in kidneys, 23.2% in prostate, and 49.2% in tumors. These results indicate that incorporating dosimetric structure weighting into CT-guided registration can improve dosimetry-relevant spatial consistency and temporal fitting stability in multi-time-point SPECT/CT workflows for radionuclide therapy.

Abstract

Boron Neutron Capture Therapy (BNCT) is a binary radiation treatment relying on the selective uptake of B-10 in tumors, requiring a high thermal neutron flux with minimal fast neutron and gamma ray contamination. The reaction between the thermal neutron and the B-10 releases an alpha and lithium nucleus, two high-LET particles that deposit their energy inside the cell causing lethal damage. This work characterizes a BNCT-optimized irradiation field at a TRIGA Mark II reactor and compares the effects on various cell samples to standard radiation fields. The irradiation setup was localized within the graphite thermal column of the TRIGA reactor (Mainz). Gamma contamination was mitigated using lead shielding, while additional graphite bars further thermalized the neutrons. Dosimetric profiling was conducted using radiochromic films paired with specific radiation converters. Additionally, PADC plates were employed to quantify the alpha particle tracks emitted by the BNCT reaction, as well as proton tracks generated by the elastic scattering of fast neutrons. For biological validation, FaDu tumor cells and isolated lymphocytes were exposed to nominal doses (1–6 Gy) to evaluate clonogenic survival and DNA double-strand breaks (DSB) via the γH2AX foci assay. Experimental data were cross-referenced with Monte Carlo (GATE/Geant4) cellular microdosimetric models. Dose contributions from primary beam components were assessed across different shielding configurations to optimize the thermal neutron field, and microdosimetric models were applied to estimate biological sample doses. Results demonstrated linear dose-response relationships for lymphocyte DNA damage, while tumor cell survival was reduced compared to photon reference data, due to relative biological effectiveness of thermal neutrons.
Ultimately, this integrated experimental and computational methodology validates the reactor facility for advanced BNCT research.

Abstract

Compton cameras are a potential tool for assessing translational targeted alpha therapy (TAT) treatments using ²²⁵Ac in murine models. For this application, gamma cameras perform sub-optimally due to the extremely low activities used. The IRIS group at IFIC-Valencia is developing Compton cameras based on lanthanum bromide (LaBr₃) crystals, capable of visualizing ²²⁵Ac distributions through the detection of the 440 keV gammas emitted during the decay of its short-lived daughter nuclide ²¹³Bi. Following successful imaging tests with ¹³¹I in live mice at CIC biomaGUNE in San Sebastián, Spain, this work presents
measurements performed using the current Compton camera prototype, MACACO III+, and a mouse phantom filled with ²²⁵Ac. The system was shown to be capable of visualizing ²²⁵Ac distributions in the phantom with two combinations of organ compartments filled with activity. Images obtained using the Compton camera are of comparable quality to those obtained using commercial pre-clinical SPECT systems at 440 keV. Monte Carlo simulations of a mouse phantom filled with ²²⁵Ac have also demonstrated the improved image quality that can be achieved with an augmented Compton camera, MACACO IV, featuring a larger first plane detector and currently under development. Measurements using MACACO III+ for imaging live mice grafted with AR42J tumor cells and injected with [²²⁵Ac]Ac-DOTATATE are planned for the near future.

Abstract

Boron Neutron Capture Therapy is a binary cancer treatment modality in which a tumour-selective boron-carrying drug is administered to the patient prior to irradiation with a neutron beam. The highly localised nuclear reaction between thermal neutrons and ¹⁰B deposits a lethal radiation dose selectively within tumor cells. After recent advances driven by accelerator-based neutron sources and improved boron compounds, a critical gap remains: there is currently no method to image boron distribution in vivo during treatment, preventing real-time dose verification.

We present the development of a Compton camera system designed to detect the 478 keV prompt gamma rays emitted following ¹⁰B neutron capture, thereby enabling direct spatial mapping of boron dose deposition. The detector is based on LaCl₃ and CeBr₃ scintillators arranged in a pyramidal geometry, inherited from nuclear physics instrumentation developed at CERN-n_TOF, which provides intrinsically low neutron sensitivity and high gamma-ray detection efficiency.

Experimental validation was carried out at Institut Laue-Langevin (ILL, Grenoble). A boron sample in water at 65 ppm, within the clinically relevant concentration range, was successfully imaged, yielding a signal-to-background ratio of 12. This constitutes, to our knowledge, the first experimental demonstration of Compton imaging of boron at clinical concentration levels.

At treatment-level neutron fluxes, however, the original detector suffered from count rate saturation. To address this, we redesigned the detector incorporating pixelated crystal arrays. The pixelated prototype was subsequently validated at the LENA reactor (U. Pavia) achieving a reduction in per-crystal count rate by a factor of approximately 250.

These results represent a substantive step toward a clinically deployable real-time boron imaging system, with implications both for treatment monitoring of patients and preclinical pharmacokinetic studies with animal models.

Abstract

Proton therapy enables conformal tumor irradiation while reducing dose to surrounding healthy tissues. But anatomy, motion, and delivery uncertainty can cause differences between planned and delivered dose. PET can detect proton-induced positron-emitting radionuclides such as ¹⁵O, ¹³N, & ¹¹C, providing a method for post-irradiation in-vivo dosimetry verification. We evaluated a mobile PoC-PET system for imaging validation & MLEM-based radionuclide decomposition.

PoC-PET used 8 TOF detector modules. Imaging performance was evaluated with a Derenzo-style tumor phantom 3.3-11.4mm lesions at a 20:1 tumor:background ratio and with a FDG-injected mouse. To evaluate radionuclide decomposition, three vials with known ¹⁵O, ¹³N, & ¹¹C activities were imaged for 60 minutes. Counts were binned in 30s windows and reconstructed into PET image series. ROIs generated decay curves modeled as mixtures of exponential decay components using an MLEM-based fitting approach.

PoC-PET resolved all tumor clusters by the VPR < 0.735 criterion. Mouse imaging clearly showed regional FDG uptake with spatial agreement to MicroPET. In the multi-isotope experiment, the MLEM fit matched the large-ROI decay curve with a ~1% nRMSE. The fitted isotope allocation captured expected isotope behavior: ¹⁵O dominating early, ¹³N showing intermediate decay, and ¹¹C persisting throughout the scan. The injected mixture was 29% ¹¹C, 16% ¹³N, and 55% ¹⁵O, the MLEM estimated 27.7%, 15.7% and 56.6% respectively.

These results demonstrate the feasibility of using a mobile PoC-PET system and radionuclide decomposition for post-irradiation radionuclide characterization.

Abstract

Multi-time-point SPECT/CT imaging enables improved characterization of radiotracer kinetics for quantitative dosimetry, particularly in emerging radiopharmaceutical therapies. However, activity quantification is affected by partial volume effects (PVE), which introduce time-dependent bias in both lesion and organ dose estimation. This issue is especially relevant for novel radionuclides such as ¹⁶¹Tb, where complex emission characteristics may lead to varying sensitivity to image correction across time. In this study, we analyze the temporal behavior of lesion and organ dosimetry for ¹⁶¹Tb-FAPI-2286 and evaluate the influence of partial volume correction at multiple imaging time points (2 h, 24 h, and 96 h post-injection). Patient-specific SPECT/CT data were used, with lesions and organs segmented using semi-automatic and deep learning-based approaches. Two iterative correction methods, Richardson–Lucy and reblurred Van Cittert (RVC), were applied prior to dosimetric analysis. Absorbed dose values were estimated using Monte Carlo simulations, and relative differences were computed between corrected and non-corrected images. Results indicate minimal impact of correction at early imaging, with relative differences generally below ±3%. In contrast, substantial and heterogeneous variations were observed at later time points, with lesion-level dose changes ranging from approximately −30% to +38%. Bone lesions exhibited the largest positive deviations, whereas lung lesions showed consistent reductions at delayed imaging. Both correction methods demonstrated similar temporal patterns, with slightly lower variability for the RVC approach. These findings highlight the time-dependent nature of PVE in ¹⁶¹Tb-based dosimetry and emphasize the importance of incorporating correction methods in delayed imaging. Temporal-aware quantitative strategies are therefore essential for improving the reliability of lesion-level dose estimation in emerging theranostic applications.

Abstract

The development of integrated strategies for simultaneous cancer detection and treatment is driving advanced theranostic platforms. In this context, the COXIRIS system (Confocal Orthovoltage X-Ray Induced Radiation Integral System) introduces a large-scale X-ray fluorescence (XRF) imaging architecture designed to combine real-time functional imaging with precision-guided therapy.This work evaluates the feasibility of COXIRIS for imaging and guiding treatment of tumors labeled with high-Z biomarkers or nanoparticles (e.g., Au, Gd). The system uses XRF detection with a collimated X-ray source directed toward a focal inspection region, while a synchronized high-sensitivity detection system performs three-dimensional scanning and tumor localization. The partially confocal detection geometry improves spatial selectivity at macroscopic scale and enables guidance of therapeutic modalities including radiotherapy, ablation, drug delivery, and surgical assistance.Experimental results demonstrated tumor localization at depths up to 15 cm in tissue-equivalent media using low radiation dose levels. Gadolinium-based targets were detected in less than 2 seconds at this depth. Although acquisition times are currently comparable to some nuclear medicine procedures, scalability through larger detector areas and optimized geometry indicates potential for sub-second detection.COXIRIS represents a novel large-scale XRF imaging and theranostic guidance platform that addresses limitations of existing micro-scale XRF techniques. Its compatibility with conventional radiotherapy systems, including linear accelerators and orthovoltage units, supports its translational potential for real-time and personalized cancer treatment.

Abstract

A novel flat-panel total-body walk-through PET system under development at Ghent University aims to improve patient throughput and cost-effectiveness compared to conventional total-body scanners. Its high sensitivity enables low-dose imaging protocols, where CT may be omitted to minimize radiation exposure. In this setting, an alternative CT-less attenuation correction method is required. The high sensitivity and large number of LYSO detector elements in the walk-through geometry also result in a substantial and potentially exploitable intrinsic background signal from ¹⁷⁶Lu.

This GATE simulation study investigates the feasibility of leveraging this intrinsic ¹⁷⁶Lu background activity in the LYSO detectors for transmission-based attenuation estimation. The limited angular coverage inherent to the flat-panel design introduces additional challenges for implementation of this method. To address this, a maximum likelihood transmission reconstruction (MLTR) algorithm was implemented. MLTR is particularly interesting for low-dose applications because it is independent of emission data and computationally efficient, making it suitable for clinical integration.

The effect of limited-angle artifacts on attenuation map convergence was evaluated, along with their impact on attenuation correction in PET image reconstruction, using a full angular coverage flat-panel geometry with side panels as reference. Simulations with a NEMA image quality phantom and the XCAT phantom demonstrated that limited angular coverage has minimal impact on attenuation correction in emission image reconstruction. These transmission-based attenuation maps may further serve as initial estimates for more advanced approaches, such as maximum likelihood reconstruction of attenuation and activity (MLAA) or physics-informed deep learning methods.

Abstract

A DOI-enabled total-body TOF PET system with a 200-cm axial field of view was developed using 2-mm-pitch LYSO crystal arrays and light-sharing-window-based DOI encoding. The system consists of seven detector rings with a transaxial diameter of 70 cm. Phantom experiments, including SZBL logo, Derenzo, and dynamic line-source phantoms, were performed to evaluate imaging performance. The reconstructed logo phantom demonstrated clear visualization of 2-mm structures, while rods down to 2.0 mm could be distinguished in the Derenzo phantom, indicating high spatial resolution consistent with the detector crystal pitch. Dynamic imaging results demonstrated continuous tracer propagation across consecutive 5-s frames, indicating high sensitivity and short-frame dynamic imaging capability. In addition, TOF resolution measurements based on the NEMA NU 2-2018 method achieved approximately 330 ps for both single-ring and dual-ring configurations. These results demonstrate the feasibility of integrating DOI capability, high spatial resolution, dynamic imaging capability, and stable TOF performance within a long axial field-of-view total-body PET system.

Abstract

Total-Body Positron Emission Tomography systems offer substantially increased sensitivity due to their extended axial FOV, enabling reduced dose, shorter acquisitions, or improved image quality. Incorporating TOF information further enhances image reconstruction by improving Signal-to-Noise Ratio. We have recently developed and installed a novel TB-PET scanner (IMAS) at Hospital La Fe in Valencia (Spain). The system is based on semi-monolithic LYSO crystals and provides TOF and Depth of Interaction (DOI) capabilities. In this work, we present the implementation and evaluation of TOF and scatter correction for the IMAS system, and we also present ongoing developments related to random, dead time, and quantification corrections are presented.

For the TOF information, an iterative skew correction was used. Scatter correction was applied using a TOF-based scaling method. For the random, dead-time and quantification correction study, the acquisition mode was modified allowing the acquisition of coincidences and singles at the same time.

A system CTR of 690 ps was measured at the NECR peak. The combined use of TOF and DOI information and the inclusion of scatter correction demonstrated a reduction in positional errors for off-center lesions and an improvement of the SNR, contributing to improving the image quality of the IMAS system.

Abstract

The NEMA NU 2 scatter phantom (70 cm, ~20 kg) was designed for conventional PET scanners with axial fields-of-view (AFOV) under 30 cm. For long axial field-of-view (LAFOV) systems exceeding 70 cm, this phantom provides only a localized assessment that fails to capture out-of-FOV scatter and randoms from the full body. Scaling the phantom to 1.2–1.75 m would increase its weight to 34–50 kg, rendering routine use impractical. We propose a single-scan multi-position protocol that exploits the idle intervals between static acquisitions in a standard decay experiment to reposition the phantom axially within each time slot. A center-plus-off-center demonstration was performed on a uMI Panorama GS (1.5 m AFOV). The 70-cm phantom was measured at both the AFOV center and 10 cm beyond the AFOV edge, with the off-center position sampled twice per slot to enable interpolation to a common activity level. The two off-center acquisitions showed close agreement across all count rate metrics and scatter fraction (SF), confirming reproducibility. The SF at the off-center position was elevated due to out-of-FOV scatter contributions, consistent with total-body imaging conditions. This approach enables evaluation of LAFOV systems at multiple axial positions without additional phantom preparation, offering a practical basis for NEMA NU 2 evalution metrics related to the scatter phantom experiment.

Abstract

Lung cancer (LC) is increasingly recognized as a systemic disease, with metabolic alterations that may vary across histological subtypes. Smoking is a major LC risk factor, but its effects on systemic metabolism and tumor phenotypes across LC subtypes remain insufficiently characterized. We integrated systemic metabolic networks and tumor habitat imaging from static and dynamic total-body ¹⁸F-FDG PET/CT to characterize smoking effects across LC subtypes. This multicenter study included 125 LC patients and 200 healthy controls from three centers. Thirty-seven regions spanning the central nervous system and peripheral organs were analyzed. Regional SUL_mean values and time–activity curves were harmonized using ComBat and adjusted for demographic covariates. Population-level metabolic networks were constructed from bootstrapped Pearson correlations of regional uptake, while individual-level networks were derived from residual time-activity curve correlations after third-order polynomial fitting to calculate organ-specific abnormality strength. Tumor habitats were defined using PET/CT intensity and entropy features, and their associations with smoking and organ abnormalities were assessed by multivariable regression and mediation analyses. Both network analyses showed that smoking disrupted systemic metabolic coordination in a subtype-dependent manner, with the liver emerging as a central hub. Tumor habitat imaging revealed a smoking-associated shift toward hypometabolic tumor microenvironments. Mediation analysis suggested that habitat composition modulated the association between smoking and hepatic metabolism, linking local tumor heterogeneity to systemic metabolic responses. These findings demonstrate that integrated metabolic network and tumor habitat imaging from total-body ¹⁸F-FDG PET/CT provides quantitative insight into smoking-related systemic metabolic reprogramming in LC and may support imaging-based characterization across histological subtypes.

Abstract

Deep learning models for PET image enhancement often struggle to generalize across unseen scanner configurations due to limited training data diversity. We introduce a Domain Randomization (DR) framework that utilizes Monte Carlo simulations via the SimPET platform to randomize scanner geometries and reconstruction protocols within clinically operational ranges. To mimic biological variability, this approach is supplemented by a tailored activity augmentation pipeline, resulting in a training dataset of 232 activity maps. To validate this methodology, we trained a 2.5D Attention U-Net to map simulations to high-quality activity maps. Performance was assessed across incremental data regimes ranging from 20 to 236 training instances. On unseen scanner domains, the DR strategy consistently outperformed the no-DR baseline in SSIM and relative contrast. Conversely, on the in-distribution source domain, the DR model achieved perceptual equivalence to the baseline, despite a minor trade-off in absolute pixel precision. Real-world validation with a Hoffman Phantom acquisition on a Philips Vereos scanner suggested a remaining sim-to-real gap on real data distributions. By demonstrating improved generalization without compromising structural fidelity on synthetic domains, this work presents a promising pathway for future research and democratizes access to high-quality, annotated training data through scalable and cost-effective deep learning tools.

Abstract

Spatiotemporal Emission Reconstruction from Prompt Gamma Timing data (SER-PGT) has the potential to verify the stopping power through passive measurements during proton therapy. The underlying inverse problem is ill-posed and offers some unusual challenges compared to conventional emission or transmission reconstruction scenarios.In this work, we introduce an unsupervised deep learning framework for reconstructing prompt gamma emission distributions from simulated data, without requiring ground-truth images during training. Our method combines a physics-informed Poisson likelihood loss in the data domain with a UNet++ architecture with ResNeXt101 encoder. It operates on pseudo-inverse reconstructions from a system matrix modeling 110 detectors, 160 depth bins, and 100 time bins. Evaluation on 200 test samples yields comparable performance to MLEM in terms of Mean Absolute Error and Peak-Signal-to-Noise-Ratio in both image and data domains, demonstrating that accurate reconstruction is achievable through self-supervised learning. The approach showcases fast execution speed and a flexible framework. This makes it a promising candidate to move SER-PGT closer towards stopping power estimation in clinical scenarios.

Abstract

This study investigates the cross-site generalization of deep learning for Tc-99m ethyl cysteinate dimer (ECD) brain perfusion single-photon emission computed tomography (SPECT), where site-specific imaging variability often limits cross-site generalization of deep learning models. To address this, a multi-site supervised contrastive learning (SupCon) pretraining framework was developed using a 3D ResNet-18 encoder, which was trained on 2,242 images collected from five sites across eight diagnostic categories. The pretrained representation achieved a latent-space silhouette score of 0.815, while t-SNE visualization showed that samples from different sites were well mixed within diagnostic clusters, indicating that the learned features were more disease-discriminative than site-specific. To simulate deployment at a new hospital, a sixth site that had been entirely excluded from pretraining was reserved for downstream evaluation. In this hospital-held-out setting, where only 46 balanced samples were available for adaptation, few-shot fine-tuning achieved 80.4% accuracy for Alzheimer’s disease-versus-Normal classification, with a macro F1-score of 0.801 and an area under the receiver operating characteristic curve (AUC) of 0.819. This performance substantially exceeded that of a randomly initialized baseline, which achieved 60.9% accuracy, a macro F1-score of 0.568, and an AUC of 0.639. These results suggest that multi-site SupCon pretraining may provide a practical route toward scalable cross-site deployment of brain SPECT classifiers with minimal site-specific annotation, potentially enabling rapid clinical adoption at new hospitals.

Abstract

This paper presents an approach for using large language models to automatically generate descriptions for 3d medical images. It describes a novel method for encoding volumetric data into a discrete set of tokens, which are then used to condition the text generated by a pre-trained foundational language model. Design considerations and model architecture are outlined, and the feasibility of the technique is demonstrated for generating descriptive captions of image volumes with information related to imaging modality, administered radiopharmaceutical and anatomical scan range.

Abstract

Synthetic computed tomography (sCT) generation from magnetic resonance imaging (MRI) is an important component of MRI-only radiotherapy planning because MRI provides excellent soft-tissue contrast but lacks quantitative Hounsfield unit (HU) information for electron-density assignment. However, MRI-to-sCT synthesis remains challenging due to the indirect MRI representation of bone, the ambiguity between intra-body air and external background, and the heterogeneous HU distributions across anatomical regions. This study proposes a region-aware 3D supervised MRI-to-sCT synthesis framework using paired CT-referenced MRI and CT volumes. The network combines 3D convolutional feature extraction with a Project-and-Excite (PE) attention module to enhance volumetric anatomical representation. To improve HU fidelity in clinically relevant regions, MRI masks were used for input background suppression, CT body masks were used to define valid supervision and evaluation regions, and HU-derived masks were used to impose region-weighted constraints on body, background, intra-body air, and bone regions. The method was evaluated on 24 paired MRI/CT cases, including 16 training, 5 validation, and 3 independent testing cases. In ablation analysis, the full body–background–air–bone loss achieved the highest PSNR of 19.65 dB, while SSIM decreased to 0.53, suggesting improved control of large HU deviations rather than simple maximization of structural similarity. These preliminary results support the feasibility of region-aware HU optimization for MRI-to-sCT synthesis, while larger-scale validation, region-wise HU analysis, and dosimetric assessment remain necessary before clinical application.

Abstract

Tau positron emission tomography (PET) is strongly linked to cognitive impairment and disease progression in Alzheimer’s disease, but direct tau imaging is not always available in routine clinical or research workflows. This study evaluated a multimodal deep learning framework that uses amyloid PET and fluorodeoxyglucose (FDG) PET to estimate binary ¹⁸F-flortaucipir (FTP) positive/negative status. A total of 340 participants from Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Geneva University Hospital cohort with available amyloid, FDG, and FTP PET were included. Images were registered to Montreal Neurological Institute space and used as two-channel volumetric input to a SwinUnetR-based classifier. The multimodal model setting was evaluated using stratified five-fold cross-validation on the combined cohort. The model achieved an accuracy of 0.853, sensitivity of 0.823, specificity of 0.871, and area under the receiver operating characteristic curve of 0.904. These findings suggest that combined amyloid and FDG PET contains complementary information relevant to tau status and may provide practical decision-support evidence when direct tau PET is unavailable or when the indication for tau imaging remains uncertain.

Abstract

In this work, a physics-aware neural network architecture is proposed for Compton camera image reconstruction. The model incorporates physics-based features from coincidence events and uses backprojection as a prior.  It is applied both to mono energetic sources as well as to targeted alpha therapy using Actinium-225 where the presence of multiple gamma-ray lines makes image reconstruction more challenging for existing medical imaging devices, such as single-photon emission computed tomography. The input data are preprocessed and categorized into scattering-only, absorption-only, and coincidence events. These data are used to generate feature maps through neural network layers and a transformer-based model, followed by an image reconstruction decoder. Training and testing data were generated using the Geant4 simulation toolkit. The model was first evaluated using controlled single-source simulations with different source sizes, energies, positions, and event statistics. It was then tested under more realistic conditions using a mouse phantom with Actinium-225 activity distributed in tumor and background regions. CPPF-Net achieved a mean peak-localization error of 1.0 ± 0.7 mm and a structural similarity index of 0.99 ± 0.01. These results indicate that CPPF-Net can provide accurate Compton camera image reconstruction for Actinium-225 targeted alpha therapy. Experimental validation with measured data is foreseen for the near future.

Abstract

Prostate cancer is one of the most commonly diagnosed cancers in men worldwide. Accurate detection using ¹⁸F-prostate-specific membrane antigen (PSMA) positron emission tomography/computed tomography (PET/CT) enables reliable identification of the disease and supports clinical staging. In this study, we propose a deep learning framework for the slice-level classification of ¹⁸F-PSMA PET/CT images using convolutional neural networks trained on axial, coronal, and sagittal views. The models achieved high discrimination performance, with an area under the receiver operating characteristic curve of 0.98 and a specificity of 0.98, while showing orientation-dependent differences in sensitivity and classification balance. However, lower performance in the hold-out dataset indicates limited generalization capability, highlighting the importance of imaging plane selection in prostate tumor classification. The classification model serves as a preliminary step for downstream segmentation by enabling slice-level localization of the region of interest.

Abstract

Low-count Positron Emission Tomography (PET) suffers from severe Poisson photon-counting noise, yet existing deep learning denoisers either rely on supervised regression networks that tend to oversmooth anatomical detail or on diffusion-based generative models designed for Gaussian corruption that require iterative, noise-level-dependent sampling at inference time. Equilibrium Matching (EqM) was recently proposed as a generative modeling framework that learns a time-invariant equilibrium gradient field toward the data manifold, but has not yet been explored for medical image restoration. We present a novel adaptation of EqM to PET denoising, reformulating its interpolation scheme for Poisson-corrupted acquisitions and introducing noise-level conditioning via Adaptive Group Normalization, and conduct a controlled comparison of its two parameterizations under identical training conditions: the signal form, which predicts the clean image directly with hyperparameter-free inference, and the velocity form, which predicts the noise residual via iterative Euler updates. Signal EqM achieves up to +5.26 dB PSNR gain at severe noise (N_peak = 50) with no step-size tuning. Velocity EqM consistently outperforms signal EqM on both PSNR and FID at all seven noise levels, with the PSNR gap growing from 0.13 dB at N_peak = 50 to 1.23 dB at N_peak = 5000, mirroring the x- vs. velocity-prediction trade-off in flow-based generative models.

Abstract

Accurate detection and localization of traumatic injuries in abdominal CT remain challenging because voxel-level annotations are limited and expensive to obtain. We present a label-efficient framework for 3D abdominal trauma detection that combines self-supervised pretraining with semi-supervised transformer-based detection. First, we use Masked Image Modeling (MIM) on 1098 CT volumes to pretrain a 3D U-Net encoder for anatomical representation learning. Next, we adapt V-DETR to dense volumetric CT through a feature adapter that converts the encoder feature grid into a compact token sequence for transformer decoding. The pretrained encoder is then integrated with V-DETR and 3D Vertex Relative Position Encoding (3D V-RPE) to improve the localization of irregularly shaped injuries. Finally, semi-supervised teacher–student consistency regularization leverages 2,000 additional unlabeled volumes during detector training. To the best of our knowledge, this is the first application of a 3D DETR-style detector to the RSNA abdominal trauma detection task. On this benchmark, the proposed method achieves 31.33% test mAP@0.50 using only 78 labeled training volumes, corresponding to a 1.53× improvement over supervised-only training. These results show that combining medical-domain pretraining with semi-supervised learning is an effective strategy for label-scarce 3D medical detection.

Abstract

High breast density is a known risk factor for breast cancer. Current gold standard methods of density measurement are highly variable, and as such, reporting of density is not standard practice worldwide. The use of artificial intelligence has shown promise, but a lack of a definitive ground truth leads to less reliable predictions. This work combines artificial intelligence with spectroscopic X-ray imaging, which preserves compositional tissue information, as a novel density estimation method. A proof-of-concept study was performed by simulating the breast imaging process for breast tissue-equivalent phantoms to build a training set for a custom machine learning tool. This gave good results, having a mean squared error in the model’s output of 0.013 +/- 0.002 %. Based on these results, further work focused on increasing the realism of the simulation and training data by modelling the charge transport in the detector as well as including anthropomorphic breast phantoms. This work will provide the foundations towards building a reliable and accurate way to measure breast density, allowing early cancer detection by tailoring patient pathways, and improving health outcomes.

Abstract

The positron range (PR) is a fundamental resolution-limiting factor in positron emission tomography (PET), particularly for high-energy radionuclides such as ⁸²Rb. In cardiac PET, the long PR of ⁸²Rb causes spatial blurring, activity spill-out, and reduced myocardial contrast, especially across heterogeneous tissue interfaces. Deep learning-based PR correction (PRC) provides a practical post-reconstruction strategy, but supervised regression networks often produce over-smoothed outputs by attenuating high-frequency stochastic fluctuations. We propose a model-consistent framework that decouples structural PR correction from texture preservation by combining PRC, matched smoothing estimation, and deterministic texture reinjection. In simulated ⁸²Rb PET studies, the proposed framework improved the recovery coefficient by 13.8% relative to the uncorrected ⁸²Rb input, while restoring noise levels closer to the reference (9.74 ± 1.85% vs. 10.62 ± 0.82%), supporting structure–texture decoupling as a promising strategy for PET imaging.

Abstract

Nasopharyngeal carcinoma (NPC) is common in southern China, Southeast Asia, and North Africa. Radiotherapy treatment planning (RTP) relies on magnetic resonance (MR) imaging and computed tomography (CT). MR provides superior soft-tissue contrast for target delineation, whereas CT provides electron density information for dose calculation. However, separate acquisition and subsequent registration of MR and CT not only increase time, cost, and radiation exposure, but also introduce registration and dosimetric errors. Diffusion model-based MR-to-CT synthesis offers an efficient, cost-effective, and radiation-free alternative for NPC radiotherapy. However, the nasopharyngeal region has highly complex anatomical structures. As a result, existing methods often fail to preserve anatomical details adequately and may suffer from mode averaging between adjacent slices, reducing slice-specificity and distributional separability. Accordingly, a wavelet-enhanced trajectory-refined flow matching model (WETR-FM) was proposed. A wavelet fusion convolution (WFConv) is introduced to strengthen the extraction of multi-frequency information in the feature space, particularly high-frequency structural details. A wavelet enhanced attention (WE-Attn) module is designed to enhance the network’s ability to model critical multi-scale structural features through attention interactions among different frequency components. A trajectory refinement loss (TR loss) is incorporated to enlarge the differences among the denoising trajectories and terminal distributions of different slices. This helps alleviate distributional confusion between adjacent slices, while improving slice-specificity. Experimental results demonstrate that WETR-FM outperformed existing approaches (cycleGAN, sc-cycleGAN, pix2pix, DDPM, FM, SM) in terms of MSE and MAE by more than 35%, and improved by 4%-10% in PSNR. These findings demonstrated promising potential for application in NPC radiotherapy.

Abstract

This paper presents a genetic-algorithm-inspired methodology for denoising positron emission tomography (PET) sinograms from a single noisy acquisition. The method assumes that a noise-free sinogram (NFS) is embedded within the noisy sinogram (NS) and formulates denoising as the search for chromosome-encoded parameters capable of recovering it. Each chromosome generates a candidate NFS by fitting continuous exponential curves to the measured sinogram data, with the exponential order and other generation parameters encoded in a reduced-size chromosome. For each candidate, the residual noise is computed as the difference between the NS and the generated NFS. The selection criterion is based exclusively on the statistic al properties of this residual noise, without requiring a ground-truth image or supervised training data. Since PET measurements are governed by photon-counting statistics, the residual noise is expected to exhibit Poisson-like behavior. Accordingly, the fitness function evaluates the mean-to-variance ratio over homogeneous count regions obtained by percentile-based stratification of the residual noise map. A ratio close to 1.0 indicates consistency with Poisson statistics. Experiments on simulated brain phantoms show a strong correlation between this noise-statistical criterion and the inverse of PSNR and SSIM after filtered back-projection reconstruction. These results indicate that chromosomes whose residual noise better satisfies Poisson statistics tend to produce higher-quality reconstructed images. The proposed method therefore provides a reference-free and interpretable framework for PET sinogram denoising, reducing dependence on annotated datasets and offering a mathematically evaluable alternative to supervised deep-learning approaches.

Abstract

The choice of cross-validation (CV) strategy can significantly affect reported performance in deep learning studies, but systematic comparisons of CV methods for medical image classification remain limited. This study systematically benchmarks seven common cross-validation strategies across nine mainstream deep learning models and a logistic regression baseline to classify major adverse cardiac events (MACE) utilising stress-phase PET MPI polar map data. For each model and CV combination, evaluated on identical data splits, we report average precision (AP) and area under the receiver operating characteristic curve (AUC) as primary metrics, with confidence intervals estimated by bootstrapping. Quantitative analysis shows that Stratified Shuffle consistently yields the highest and most stable AP and AUC scores, exemplified by SqueezeNet achieving an AP of 0.52, AUC of 0.85, and delivering reduced variability across models compared to the bootstrap approach. These results highlight the critical impact of CV strategy on reliable and reproducible model evaluation for imbalanced and limited-size biomedical datasets.

Abstract

Optimizing detection efficiency and angular resolution in Compton cameras (CC) remains a challenge for its use in nuclear medical imaging of targeted alpha therapy (TAT) and prompt-gamma (γ) Positron Emission Tomography (PET) isotopes. A trade-off between detection efficiency and angular resolution must be considered when designing scattering layer material, thickness, and distance to the absorbing layer. We propose a novel approach by repurposing a cross-strip edgeon Cadmium Zinc Telluride (CZT) PET detector system as a high-sensitivity, single-layer CC. By leveraging 40 mm of CZT thickness and binning coincidences based on scattering distance, the system functions as a virtual multi-stage CC within a single volume. GATE simulations of γ’s from 137Cs (662 keV) and 133Ba (356 keV) yield event sensitivities up to 4,508±7 and 1,239±4 cps/MBq and angular resolutions of 10.8° and 16.3° respectively. This work demonstrates the advantages of high energy resolution edge-on CZT, operating as a dual-mode PET/CC, providing a high-sensitivity where angular resolution can be further enhanced without sacrificing sensitivity. Future work aims to improve the existing design with advanced reconstruction algorithms, depth-of-interaction (DOI) correction, and detector optimization.

Abstract

A new microscopic nuclear imaging instrument using radioluminescence microscopy (RLM) has been developed to image living cells at high resolution with radiotracers tagged with positron emitters. These are used in biological samples to mimic clinical applications in order to evaluate their effectiveness to target receptors, antigens, and proteins that are overexpressed during growth of cancerous cells. The methodology has demonstrated impressive results for planar applications using optical reconstruction of β-ionization tracks (ORBIT); however, 3D imaging presents a challenging problem of reconstructing the 3D distribution of the source requiring backtracking of the ionization track of the positron in the scintillator in order to determine its position and 3D orientation at the detector surface. Four microlens arrays are used to record projections, limiting the angle of projection measurements to 360° sampling in the azimuthal angle φ and 40° sampling in the polar angle θ. The angle information acts as a collimator but the limited angle acquisition presents a challenging reconstruction problem which increases in difficulty for sources further from the scintillator. We consider reconstruction of the image x as a hierarchical maximum a posteriori (MAP) estimation problem under a Poisson data model with a total variation (TV) prior. In our algorithm we treat the prior weight β as a variable and jointly estimate x and β by minimizing the sum of the negative Poisson log-likelihood and a TV penalty. An alternating optimization scheme is used in which x is updated by MAP-TV reconstruction and β is updated from the current image total variation using a deviance-based adaptive update method. Good results are demonstrated using Geant4 simulations of point sources as far as 98 μs from the face of the detector. The novelty is the microscopy application to image individual cells three-dimensionally with radiopharmaceuticals.

Abstract

X-ray fluorescence computed tomography provides element-specific imaging, but weak fluorescence peaks are often obscured by a broad Compton-scatter continuum in photon-counting spectra. The objective of this study was to improve fluorescence projection extraction before reconstruction by replacing local empirical background fitting with a physics-informed hierarchical full-spectrum unmixing method. Each measured spectrum was modeled under Poisson counting statistics as the sum of a Klein--Nishina-constrained scatter template and a calibrated fluorescence-line template. The estimator was built hierarchically, progressing from pixel-wise spectral fitting to detector-plane joint unmixing and cross-angle projection-stack refinement. The method was evaluated on a simulated polymethyl methacrylate phantom with known fluorescence projections and on an experimental spectral data set. Under the default simulated Poisson-noise setting, the proposed method reduced projection root-mean-square error from 7.98 to 3.54 and background standard deviation from 3.92 to 1.39 relative to polynomial background extraction. On experimental data, it achieved the lowest contrast-to-noise-ratio-based detection limit, 0.305 mg/ml, compared with 0.370 mg/ml for polynomial extraction. These results indicate that structured full-spectrum unmixing can improve weak fluorescence extraction and downstream detectability while remaining compatible with existing reconstruction pipelines.

Abstract

[¹⁸F]FDG and [¹¹C]choline ([¹¹C]CH) are used in hepatocellular carcinoma (HCC) lesions for more accurate characterization across various differentiation stages. In this study, we investigate a simultaneous dual tracer protocol, in which [¹¹C]CH is administered immediately following [¹⁸F]FDG scanning, and assess whether dual tracer images can provide diagnostic information comparable with single tracer [¹⁸F]FDG and [¹¹C]CH. We simulated single and dual tracer datasets to evaluate the impact of residual [¹⁸F]FDG activity on [¹¹C]CH quantification. Realistic activity distributions were used, and lesions of varying sizes (1–4 cm) were inserted in the liver and spleen with different tumor-to-background ratio (TBR) uptake patterns. Analytic simulations were performed using STIR library for a long axial field-of-view PET scanner. Results indicated that all lesions remained detectable in dual tracer images, and TBR preserved differentiation dependent patterns. Averaged relative differences and standard deviation for SUV_mean, SUV_max and SUV_peak between the dual tracer and [¹¹C]CH images were -21.1±10.5%, -22.5±10.6% and -18.0±8.2%, respectively. To improve quantification, voxel-wise [¹⁸F]FDG subtraction was applied as a separation step, reducing the mean absolute relative difference to 1.2±1.1% for SUV_mean, 4.4±4.1% for SUV_max and 1.2±1.08% for SUV_peak. These results demonstrate the clinical feasibility of a dual tracer [¹⁸F]FDG:[¹¹C]CH imaging protocol and motivate further optimization of separation methods for simultaneous dual tracre clinical implementation.

Abstract

The Nuclear Plasma Science Society (NPSS) Educational Committee (EduCom) oversees educational initiatives and manages the NPSS Schools as its main vehicle for sustainable education. One such initiative is the EduPET project, which aims to develop an educational system to be used in EduCom's instrumentation schools. The EduPET system was designed to demonstrate the fundamental principles of Time-of-Flight Positron Emission Tomography (ToF-PET) using modern detector instrumentation. The current prototype consists of two opposing detector modules, each comprising a 2×2×30 mm³ LYSO:Ce crystal coupled to a Hamamatsu S13360-2050 silicon photomultiplier (SiPM), with signal digitization performed by a DRS4-based WaveDREAM board. Initial measurements included energy spectroscopy, coincidence timing, and ToF-based source localization using a ²²Na point source positioned at multiple locations along the detector axis. Coincidence time resolution (CTR) values ranging from 463 to 591 ps FWHM were obtained across the different positions. The timing resolution achieved was sufficient to resolve source displacements and demonstrate the probabilistic nature of ToF-based position estimation, confirming the suitability of EduPET system as an effective educational platform for illustrating key ToF-PET concepts and advanced instrumentation techniques. Future developments will focus on expanding the detector geometry toward a complete tomographic system.

Abstract

Intraoperative radiotracer imaging requires gamma cameras to maintain adequate imaging performance at relatively large source-to-camera distances, because surgical targets are often separated from the detector by tissue, instruments, surgical access constraints, or sterile operating space.

In this work, we propose a Telephoto-Lens Magic-Cube gamma camera based on a hybrid encoding mechanism. The proposed system integrates front-end long-hole geometric encoding with detector-side self-collimation encoding, allowing distant radioactive sources to generate distinguishable detector responses.

The proposed system was evaluated using numerical simulations and radioactive-source prototype experiments. In simulations, the Telephoto-Lens Magic-Cube maintained better spatial resolution at large source-to-camera distances than the Magic-Cube system without the Telephoto-Lens and two conventional parallel-hole collimator systems. It also successfully detected low-activity hot spots under background activity at large distances, whereas the reference systems failed to provide reliable detection. Prototype experiments demonstrated a mean sensitivity of 118 cps/MBq over 50–200 mm, dual-source resolution of 3 mm at 100 mm and 5 mm at 150 mm, and accurate lateral localization of 10 μCi point sources. In an anthropomorphic prostate lymph-node phantom, the system visualized five low-activity deep lymph-node-like sources under background activity and organ-uptake interference. These results suggest that the hybrid-encoded Telephoto-Lens Magic-Cube gamma camera is promising for intraoperative localization of deep low-activity lesions or lymph nodes.

Abstract

Whole gamma imaging (WGI) is a novel imaging concept combining both PET and Compton camera using a dual-ring detector arrangement. We previously developed a series of WGI prototypes for proof-of-concept. In WGI Prototype IV (WGI-IV), reported in 2025, we targeted imaging of 909 keV γ-rays from ⁸⁹Zr, achieving a 3.0-mm rod resolution by Compton imaging. In this study, we aimed at further improving Compton imaging performance by using thin scatterer detectors, considering a potential extension to typical SPECT tracer imaging, and we developed WGI Prototype V (WGI-V). Compton imaging performance is determined mainly by the spatial resolution and energy resolution of scatterer detectors. Therefore, we used fine and thin 0.9×0.9×2.0 mm³ HRGAGG scintillators to obtain high spatial resolution and high energy resolution. The outer absorber ring (304 mm diameter), which was composed of 216 fast-LGSO crystal blocks, was the same as that in WGI-IV. The inner scatterer ring (109.4 mm diameter) was remodeled with 10 HRGAGG crystal blocks (22×22 array). We evaluated energy resolution for scatterer detectors using ¹³³Ba, ²⁴¹Am, ^99mTc and ¹⁷⁶Lu (background). Then, we evaluated small rod resolution using a ⁸⁹Zr-filled rod phantom (rod diameters: 1.0-4.8 mm). Energy resolutions for scatterer detectors were 37.6% at 31 keV, 27.3% at 59.5 keV, 21.5% at 81 keV, 15.0% at 141 keV, and 7.5% at 307 keV. These were superior to those for the previous scatterer detector (HRGAGG, 1.45×1.45×15 mm³, 14×14 array). In the ⁸⁹Zr-filled rod phantom measurement, the 2.2-mm rods were resolved by WGI-V, whereas they could not be resolved by WGI-IV. High-resolution Compton imaging (2.2-mm rod resolution) was achieved by WGI-V with 0.9×0.9×2.0 mm³ HRGAGG scatterer detectors. An application of the WGI-V to SPECT tracers will be presented at the conference.

Abstract

The measurement of the three-photon to two-photon annihilation rate ratio (R_3γ/2γ) in positronium decays emerges as a promising diagnostic indicator for accessing tissue hypoxia [1, 2, 3]. The triplet state of positronium (ortho-positronium) is highly sensitive to local molecular oxygen concentrations [4, 5]; thus, the R_3γ/2γ ratio could reflect on the oxygenated level of tissues [6, 7]. This study presents methodology for determining the absolute values of this annihilation ratio using the Jagiellonian Positron Emission Tomograph (J-PET), the first PET scanner based on plastic scintillators [8]. We performed measurements with sodium-22 source enclosed in metal to characterize the detector response and geometrical acceptance for the R_3γ/2γ ratio. Since the positronium formation is suppressed in metals the resulting annihilation ratio serves as a detector-dependent reference value.  After calibrating with sodium-22 in metal, we have performed the studies with porous polymer material where the positronium formation is enhanced. The two-photon and three-photon annihilation events selection are described, and their respective lifetime distributions are presented. The results presented validate the event selection methodology and establish the detector calibration baseline for future quantitative R_3γ/2γ studies in biological sample with J-PET.

Abstract

Compton cameras (CCs) can be designed as compact and portable systems that can be employed in combination with other imaging modalities. These imaging systems provide 3D images with just one device. However, the resolution in the axis perpendicular to the first detector is degraded with respect to the parallel plane, especially at positions far from the detectors. Placing two CCs orthogonally can improve the uniformity of the 3D spatial resolution and also increase the usable size of the field of view (FOV). Following an initial study based on simulations carried out by the IRIS group, the current work aims at an experimental verification of operating two CCs positioned orthogonal to each other in the dual-CC MACACOX2 system. In this work, simulations are performed using a geometry that replicates the laboratory setup. Coincidence data are simulated with sources located at the center of the geometry, such that the scatterers of both CCs are positioned 30 mm from the center of the FOV. The image reconstruction algorithm has been developed and optimized by the group specifically for this imaging modality and is being applied in this work. Images of a ²²Na point source array obtained using the simulated two head CC system and reconstructed at 511 keV are obtained. Point sources separated by 8 mm can be clearly resolved at two, three, and four cm from the detector surface. Such successful findings provide foundation for the ongoing laboratory experiments with the MACACOX2 system.

Abstract

Medical thermal imaging has long been explored as a non-invasive diagnostic modality for detecting localized sub-surface abnormalities associated with pathological processes. However, despite decades of investigation, conventional thermography has not achieved widespread clinical acceptance due to inconsistent and unreliable results. In this work, we experimentally validated a framework for medical thermal tomography combining high-precision thermal measurements in a controlled and monitored environment with physics-informed tomographic reconstruction. Following software-phantom studies, the methodology was validated using a dedicated automated experimental setup together with a suite of calibrations and temporal and spatial corrections, reducing the uncertainty in panoramic thermal projections to approximately 25 ^◦C. Reconstruction was performed using physics-informed forward thermal diffusion models within the AMIAS/RISE framework. Using this methodology, localized hot-spots with temperature differences as low as ∆T = 1.5 ^◦C were successfully detected and reconstructed at depths up to 3.3 cm in hardware phantoms. The reconstructed positions and temperatures were in good agreement with independent thermocouple and μCT measurements. These results provide strong evidence that thermal tomography is feasible when combined with precision thermal measurements and physics-informed reconstruction methods.

Abstract

Quantum-inspired formulations have emerged as a promising approach for medical image denoising and segmentation by recasting image processing tasks as eigenvalue problems. While computationally demanding on classical hardware, these methods align naturally with quantum computing paradigms, particularly the Variational Quantum Eigensolver (VQE). In this work, we assess whether quantum simulation on
classical hardware is justified in the absence of real quantum devices by comparing classical computing (CPU-based diagonalization), quantum simulation on CPU, and GPU-accelerated quantum simulation using Qiskit. Experiments on synthetic data and a clinical PET/CT case show consistent reconstruction and segmentation quality across implementations. Quantitatively, classical methods are faster for small problems
(e.g., 0.0063 s vs 0.0229 s), while GPU-based quantum simulation reduces this gap (~0.003 s) and becomes slightly faster at larger scales (660.78 s vs 669.1 s). These results indicate that quantum simulation is not advantageous for small cases, but can become competitive or beneficial for larger problems within a quantum-inspired framework.