Online program Overview Session: MIC-10

Abstract

Lead-loaded plastic scintillators are promising candidates for next-generation positron emission tomography (PET) detectors, combining the fast timing response and low cost of organic scintillators with enhanced gamma-ray stopping power due to heavy-element doping. This study evaluates the 511-keV photon detection performance of EJ-256, a commercially available plastic scintillator containing 5 wt% lead, relative to a conventional polyvinyltoluene (PVT) scintillator using Monte Carlo simulations. All simulations were performed in Geant4 (v11.1.1). Monoenergetic 511-keV gamma rays were incident on scintillator blocks with a fixed cross-section of 4 × 4 mm² and thicknesses ranging from 20 to 200 mm in 5-mm increments. For each configuration, 100,000 primary events were simulated, and the resulting energy-deposition spectra were analyzed in ROOT. Detection efficiency was defined as the fraction of incident photons depositing 505–515 keV within the scintillator volume. EJ-256 exhibited a distinct fullenergy peak at 511 keV, whereas the reference PVT scintillator showed only a broad Compton continuum. Detection efficiency for EJ-256 increased from 0.9% at 20-mm thickness to 2.9% at 200 mm, consistent with enhanced photon absorption at greater thicknesses. However, efficiency gains became marginal beyond approximately 135 mm, indicating practical saturation. The energy resolution remained approximately constant at ~10% full width at half maximum across the investigated thickness range. These results indicate that moderate lead loading enhances 511-keV photon detection in plastic scintillators without compromising spectral stability, supporting the potential of EJ-256 as a cost-effective candidate for PET detector modules.

Abstract

Time of flight positron emission tomography (TOF-PET) plays an important role in the diagnosis of Alzheimer’s disease and cancer because it enables localized estimation of lesion positions. Improving coincidence time resolution (CTR) is essential for enhancing the precision of lesion position estimation. Currently, commercially available PET systems achieve a CTR of approximately 200 ps full width at half maximum(FWHM), while the next goal is to reach 100 ps FWHM at the system level. However, commercial PET systems employ thick scintillators for high detection efficiency, resulting in a trade-off between CTR and detection efficiency due to photon transport time spread. To overcome this trade-off, we proposed xDetector, a readout scheme in which detector groups are orthogonally stacked. xDetector employs shorter scintillators than conventional detectors to improve CTR, while maintaining detection efficiency through multiple layers stacked. In addition, by orthogonally stacking detector groups via air coupling, it is possible to estimate the interaction position within the scintillator, which contributes to further CTR improvement. In previous study scintillator crystals and multi pixel photon counters (MPPC)s with 4.2 mm pitch were optimized for CTR measurements together with a high-frequency readout circuit. However, it was demonstrated that the CTR of xDetector with 3.2 mm pitch was 10% faster than that with 4.2 mm pitch, which can be attributed to faster time response of 3 mm MPPC than 4 mm. Thus, in this study, we aimed to further improve CTR by leveraging geometric flexibility and adopting small crystals with the same cross-sectional pitch of MPPCs compared to previous configurations. Through this investigation, we explore the potential of xDetector towards realizing 100 ps CTR.

Abstract

Depth-of-interaction (DOI) estimation can reduce parallax error in PET imaging, but conventional pixelated detectors usually require additional DOI-encoding hardware or modified detector geometries. This study investigates a cost-effective neural network approach to recover DOI information from a state-of-the-art air-coupled pixelated PET detector using only compact engineered features, which could be extracted directly by front-end ASICs without digitizing all silicon photomultiplier (SiPM) channels. A Geant4 optical simulation was performed, and neural networks were trained for both 3D position regression and DOI-bin classification. The feature-only model achieved DOI performance comparable to models using the full raw SiPM counts, with an MAE_z of 1.107±0.001 mm, and an F1 score of 0.790 for DOI-bin classification, with most classification errors occurring between neighboring bins. These results demonstrate a practical and cost-effective pathway to enable DOI capability without changing the detector geometry, while reducing readout complexity, system cost and power consumption.

Abstract

Photon-counting computed tomography (PCCT) has attracted attention because of its potential for high spatial resolution, low-dose imaging. Although semiconductor detectors are widely studied for PCCT, indirect-conversion detectors with fast and bright scintillators combined with silicon-photomultipliers arrays (SiPMs) have recently emerged as promising alternatives owing to their potential for the scalability and cost effectiveness. Though Gd₃(Al,Ga)₅O₁₂ or (Lu,Y)₂SiO₅ shows an excellent scintillation properties, Kα escape peak of Gd (43 keV), and Lu (55 keV) degrade the accurate photon counting. Instead, Mg co-doped Y₃(Al,Ga)₅O₁₂:Ce (YAGG) garnet scintillators with lower Kα escape peak (<15 keV) is developed for X-ray imaging applications. The effect of Mg co-doping on scintillation properties results in representing 5-7 folds increase in the light yield of 35,000 photons/MeV@662 keV and the decay time of <50 ns. In addition, synchrotron X-ray imaging performance demonstrates that the spatial resolution of Mg co-doped Y₃Al₃Ga₂O₁₂:Ce is comparable to that of LuAG:Ce with the 1 mm sample thickness. Although the density of YAGG:Ce,Mg (5.08–5.33 g/cm) is lower than that of LuAG:Ce (6.67 g/cm), there is a concern regarding X-ray stopping power and deterioration of spatial resolution, the linear attenuation coefficients at 12.4 keV are sufficiently high (273–350 cm⁻¹ for YAGG:Ce,Mg and 943 cm⁻¹ for LuAG:Ce), resulting in X-ray absorption efficiencies exceeding 99.99% for 1 mm thickness samples, indicating negligible influence on imaging performance in the lower energy range. Moreover, we also demonstrate the feasibility of YAGG:Ce,Mg to medical field from the success in clear imaging of small, dried fish. These results show that YAGG:Ce,Mg scintillators are promising candidates for next-generation X-ray imaging detectors and can be applied for indirect-conversion PCCT applications.

Abstract

In positron emission tomography, time-of-flight information can localize source positions along lines of response. Cherenkov-radiator-integrated microchannel-plate photomultiplier tubes (CRI-MCP-PMTs) with a 3.2-mm-thick lead glass window (n~1.5) have achieved 30-ps coincidence time resolution (CTR). However, improving coincidence detection efficiency remains a challenge, as increasing the radiator thickness degrades the CTR. In this study, we focused on the radiator refractive index. Based on the Frank-Tamm formula, we hypothesized that ultraviolet-transparent materials with a high refractive index would increase Cherenkov photon production, reducing event loss. Furthermore, increased Cherenkov photon detection would improve effective single-photon time resolution, thus enhancing CTR. We tested this hypothesis using Geant4 simulations of point-source measurements with paired CRI-MCP-PMTs incorporating optical and detector characteristics. Assuming wavelength-independent refractive indices from 1.7 to 3.0, we found coincidence detection efficiency increased from 1.1% to 3.0% whereas CTR deteriorated from 41.5 to 49.6 ps. Contrary to our expectations, the reduced speed of light in the medium increased timing jitter, overriding the CTR benefits from detecting more Cherenkov photons. A similar simulation for lead fluoride radiators (n~1.7–2.1)—expected to improve efficiency while maintaining the CTR—showed a worse CTR (46.4 ps) was obtained than for the wavelength-independent cases (~42.0 ps). Wavelength-resolved analysis showed a CTR of 53.3 ps for 240–340 nm photons, which was unexpectedly worse than that of 39.9 ps for 540–640 nm photons, despite the higher theoretical photon production in the shorter band. One possible cause is the steep refractive index dispersion in the shorter band. In summary, while higher refractive indices improve coincidence detection efficiency, they can degrade CTR. Minimizing refractive index dispersion is critical for fast timing.

Abstract

Diamond-based dosimeters are popular for medical dosimetry in view of their fast response, radiation hardness and tissue equivalence. However, some design improvements to achieve better accuracy in signal readout and approach tissue equivalence are possible by the removal of the metallic contacts, sometimes covering the active area of said devices.
In this contribution, we explore the use of a simple processing and scalable solution to obtain non-metal, low Z electrical contact. Thus, the fabrication of diamond detectors with graphitic contacts at the Microelectronics Institute of Barcelona are shown. Several approaches are explored, some of which never utilised before for radiation detector fabrication, namely pyrolytic photoresist film, nickel catalysis and ion surface graphitisation.
A prototype of such device using the pyrolytic photoresist film process for the first time has been produced and was tested under a 60 kV X-ray tube. We report the fabrication and preliminary performance results of such device. Although the substrate material quality is not ideal for radiation detection, it was possible to observe radiation response using the PPF as an interface material. The prototype features a linearity better than 3% as tested in a laboratory X-ray chamber, proving it a promising technology for detector operation. Finally, the technology challenges are finally discussed to assess the feasibility of PPF-diamond technology for dosimeter batch production.

Abstract

Accurate localization of annihilation γ-ray interaction positions within detector block is essential for achieving high spatial resolution in positron emission tomography (PET). Conventional high-resolution PET detectors often use finely pixelated scintillator arrays resulting in increased system cost due to the larger number of crystal elements and number of readout-channels. Semi-monolithic scintillator detectors offer an alternative solution by enabling signal-based position estimation with reduced readout. In this work, we experimentally evaluated a classification-based multilayer perceptron (class-MLP) for estimating the γ-photons interaction position direction within each semi-monolithic LYSO detector slab across a cassette of 40 mini-modules (mM), each comprised of 16 LYSO slabs coupled to 8x8 SiPMs and read by 16 readout channels, developed for the Ultra-High-Performance Brain PET (UHB-PET) system. Calibration data were acquired from each of the 40 mM, and operated on a benchtop setup using a 0.5mm slit-collimated source translated in 0.5mm steps. The class-MLP used 24 multiplexed readout and classified events into 49 position bins spanning 0-24mm along the monolithic. Independent MLPs were trained for the individual mM. After correction for the finite collimator resolution, the mean positioning-error FWHM was 1.29±0.13mm (range: 0.85 to 1.41mm). The corresponding FWTM was 3.91±0.29mm (range: 3.29 to 4.73mm). Our results demonstrate that class-MLP can achieve near-millimeter detector mean spatial resolution along the monolithic dimension in 20mm-thick semi-monolithic LYSO detectors with minimal variability across all mM of a UHB-PET cassette. Future work will focus on acquiring calibration data from all cassettes in the fully assembled UHB-PET scanner to characterize its system-wide detector spatial resolution using the current calibration strategy per mM.

Abstract

Laser engraving of monolithic scintillators to create pixel boundaries has been around already for some time, with the promise of good pixel identification as in conventional pixelated detectors, but at a much lower cost. However, their use is hindered by the fact that the pixel boundaries are semi-transparent, making it difficult to identify the pixel of interaction with the algorithms applied to fully pixelated detectors.   The RETIMAGER project proposes laser-engraved (LE) pixelated LYSO blocks, with optimized transparency of pixel boundaries that might be exploited to obtain depth-of-interaction (DOI) based on the light spread in the SiPM array, while keeping crystal pixel identification. Hexagonal SiPM arrays, scintillators and pixel cells are introduced, improving packing fraction and reducing gaps in between detectors, but adding difficulty when applying ordinary vertex identification strategies. We present the experimental DOI performance  for several RETIMAGER detectors, engraved with different optical densities (and thus trasparencies) por hexagons of 1.25 mm side. Data is acquired in a testbench with a collimated source, and employed to train an NN algorithm, to predict DOI. For a 13 mm thick crystal, DOI assignation uncertainty ranges from 2.4 to 3.2 mm FWHM, depending on the trasparency of the boundaries, with the highest trasparency detectors exhibiting the best results. Experimental results follow the trend of the simulations, which can be used to further optimize the DOI performance and to reduce bias.

Abstract

Bismuth germanate (BGO) is a high-density and cost-effective scintillator for PET detectors, but depth-of-interaction (DOI) encoding with single-ended readout remains challenging due to its relatively low light yield. This study investigated a light-sharing based BGO detector using paired crystals for DOI encoding. The effects of reflector geometry and lateral surface treatment were evaluated to optimize detector performance. Among the unpolished configurations, the trapezoidal reflector provided the best DOI encoding capability, achieving an average DOI resolution of 5.29 mm FWHM and an energy resolution of 20.4%. Partial surface polishing improved energy and timing performance but degraded DOI encoding capability. After applying timing correction based on DOI and rise-time correction, the timing resolution of the unpolished trapezoidal configuration improved from 427 ps to 353 ps FWHM. These results suggest the feasibility of a single-ended DOI approach for BGO PET detectors with potential time-of-flight (TOF) applications.

Abstract

Time-of-flight (TOF) positron emission tomography (PET) performance is governed by the coincidence time resolution (CTR), which is fundamentally constrained by the number and arrival-time distribution of scintillation photons reaching the photosensor. To enhance both quantities, we fabricated nano-scale line- and hole-type photonic crystal (PhC) patterns with a 1.2 μm period on the polished 3×3×20 mm³ surfaces of lutetium-yttrium oxyorthosilicate (LYSO) scintillators via flex-mold soft lithography using a UV-curable optical resin (n = 1.5). Six SiPM coupling conditions were evaluated: bare LYSO with and without grease, and line- and hole-type PhC patterns each with and without grease.

The intrinsic scintillation photon arrival-time distribution was reconstructed by deconvolving each measured waveform with the single-photoelectron (SPE) response of a photomultiplier tube (PMT), and the absolute light gain of each PhC geometry was independently assessed in an airgap PMT configuration by comparing the photopeak before and after patterning. PhC-patterned crystals achieved a 42.8% reduction in deconvolved rise time and light gains of 1.20 (line) and 1.18 (hole). To quantify the redistribution of photon arrivals, fractional yields were computed within three time windows — early (0–1 ns), mid (1–10 ns), and late (10–50 ns) — normalized to the total yield over 0–50 ns. The early-window yield increased from 0.7% to 0.9% (a 33.5% relative gain), the mid-window yield rose from 19.8% to 21.2%, and the late-window yield decreased from 79.5% to 77.9%, confirming a redistribution of photon arrivals toward the earlier windows.
In SiPM CTR measurements, the line-type PhC combined with grease achieved the best CTR of 180.6 ps, a 29.5% improvement over the bare crystal. These findings demonstrate that nano-scale PhC patterning improves CTR in LYSO-based TOF-PET by enhancing both the absolute light yield and the prompt photon density.

Abstract

X-ray fluorescence imaging (XFI) is an emerging molecular imaging technique that combines high sensitivity with high spatial resolution, offering significant penetration depth in tissues while minimizing ionizing radiation exposure compared to other existing methods. This makes XFI a promising tool for tumor characterization and personalized treatment. The IMMPRINT project aims to develop a proof-of-principle demonstrator that leverages X-ray fluorescence computed tomography as a novel hybrid modality for tumor profiling, with a focus on breast cancer. The pnCCD silicon sensors produced by MPG HLL, Munich, have been chosen for the proof-of-principle studies. Fluorescence photons are generated when gold nanoparticles are excited by incident X-rays and are detected by the pnCCD. A specific system has been designed to operate up to four pnCCDs, including a cooling system to stabilize sensors’ temperature between 0ºC and 20ºC, and it is under construction. Simulations are also performed which support the system design.

Abstract

In the context of the “Small Animal Proton Irradiator for Research in Molecular Image-guided Radiation Oncology” (SIRMIO) project, a high-resolution depth-of-interaction (DOI) LYSO-based detector has been developed for a small-animal in-beam PET scanner. In addition to PET applications, the DOI detector may also be suitable for the detection of high-energy prompt gamma rays originating from excited nuclei such as 12C and 16O, enabling potential applications in Compton camera imaging for ion-beam therapy monitoring. However, the interaction mechanisms of high-energy gamma rays differ from those of 511 keV annihilation gamma rays,  particularly in terms of the relative contributions of photoabsorption and Compton scattering processes. In this study, we experimentally investigated the performance of the DOI detector for high-energy gamma-ray detection to gain a better understanding of its applicability to prompt gamma imaging. The detector consists of a three-layer pixelated scintillation block coupled to a 64-channel silicon photomultiplier array. The use of the PETsys TOFPET2 ASIC, employed in the Compton camera MACACOp developed by the IRIS group, is a potential alternative to the current electronics employed in the SIRMIO project, which enables individual channel readout. The study demonstrates the feasibility of operating the detector with PETsys electronics, and characterizes its response with 60Co and 22Na sources.

Abstract

Positron emission tomography systems are becoming increasingly complex due to the growing number of detector channels and the large amount of data generated, especially in total-body imaging configurations. In this work, we present PETRARCA, a positron emission tomography detector system integrating an analog neural network ASIC (ANNA) for real-time edge computing of gamma-ray interaction coordinates within a monolithic scintillation crystal. The system consists of a 32 mm × 32 mm × 10 mm monolithic LYSO crystal coupled to an 8 × 8 SiPM matrix read out by dedicated front-end electronics. The detector architecture includes both a conventional digital acquisition path and a parallel analog inference path based on the ANNA ASIC, enabling simultaneous acquisition of sensor data and hardware neural-network predictions. A first evaluation of the system has been carried out using simulated data and the real ASIC response to emulate a reduced-scale configuration using a 4 × 4 SiPM array coupled to a 26 mm × 26 mm × 10 mm monolithic crystal. The performances of the ASIC in predicting the interaction coordinates were tested, and proved to have a mean absolute error (MAE) of around 1 mm on both the x and y axes. The preliminary results demonstrate the feasibility of analog neural-network edge computing for spatial localization in PET detectors and support further developments toward fully integrated real-time detector systems.

Abstract

Lutetium oxide–based scintillators are of growing interest for positron emission tomography (PET) because of their high density and high effective atomic number, which provide efficient stopping of 511 keV annihilation photons. Lanthanum-doped (Lu,Y)₂O₃ (LYO:La) is a transparent ceramic variant that retains the favorable stopping power of Lu₂O₃ while being easier to fabricate, making it an attractive candidate for next-generation PET detector arrays. In this work we report a scintillation characterization of LYO:La pixels in two geometries relevant to PET, 1×1×10 mm³ for high-resolution preclinical and dedicated-organ imaging, and 3×3×10 mm³ for whole-body clinical systems. Pixels were cut from LYO:La ceramic disks, polished on all faces, and evaluated with a range of side reflectors (Teflon, Tyvek, ESR film, and Spectralon) in both flat and upright coupling orientations. We measured X-ray-excited radioluminescence emission spectra, ²²Na (511 keV) pulse height spectra, relative light yield, energy resolution, scintillation decay time profiles, and timing resolution. The results illustrate the influence of pixel geometry, coupling face, and reflector choice on light collection and energy resolution, and show that LYO:La pixels offer a promising combination of stopping power, light yield, and timing performance for PET detector modules.

Abstract

Dynamic medical X-ray imaging requires detectors that combine high quantum efficiency, low residual lag, and high spatial resolution without the optical spread and afterglow of scintillator-based flat panels. We report a fully inorganic perovskite-based direct-conversion perovskite X-ray imager fabricated from phase-pure orthorhombic Cesium Lead Bromide (CPB) sensor material and integrated with active-pixel readout arrays. Phase-pure CPB was obtained by controlled crystal growth and converted into dense blade-coated films; non-phase-pure starting materials containing secondary phases produced nonuniform films and degraded images, confirming the importance of phase control for reproducible detector performance. The optimized CPB detector demonstrated high linear X-ray response, sensitivity of 55,400 μC Gyair⁻¹ cm⁻² at 30 keV and 0.5 V/μm, a limit of detection of 417 pGyair s⁻¹, a mobility-lifetime product of 8.0 × 10⁻⁴ cm² V⁻¹, and a calculated detective quantum efficiency near 85% at practical dose levels. In flat-panel imaging, CPB imagers reached 6.9 lp/mm at 20% modulation transfer function on a 100 μm-pitch amorphous-silicon array and resolved 20 lp/mm after integration with a CMOS readout. Most importantly for low-lag medical applications, improvements in phase purity and charge-transport stability reduced residual detector lag from 8.9% to 1.1%. These results show that all-inorganic perovskite direct conversion can provide a path to low-dose, low-lag, high-resolution imaging for fluoroscopy, tomosynthesis, dental cone-beam computed tomography, and other image-guided procedures.

Abstract

Large-area microchannel plate photomultipliers (MCP-PMT) such as the Large-Area Picosecond Photon Detector (LAPPD) and High-Rate Picosecond Photon Detector (HRPPD), are attractive for time-of-flight positron emission tomography (TOF-PET) due to their superior single-photon time resolution (SPTR) and low-cost potential. However, when coupling them with popular TOF-PET scintillators, it was previously reported worse-than 250 ps coincidence time resolutions (CTR) for annihilation photons. In this work, by applying a custom high-frequency (HF) amplifier to the outputs of an HRPPD (10×10cm² window) coupled to a 3×3×3 mm³ LGSO crystal and an SiPM-based reference detector, we measured a 144 ps CTR. In addition, we showed that individual crystals of a 4×4 array of 3×3×10 mm³ LYSOs can be resolved using this HRPPD whose output was 4×4 pads having a 25 mm pitch. We also constructed a detector model for investigating key factors affecting CTR. These factors include scintillation yield, light transport in the crystal and HRPPD window, photon detection efficiency and charge collection of the HRPPD, and electronics contributions, with parameters derived from analytical calculations, Geant4 simulations, or experimental measurements. This model predicted a CTR that was consistent with the measured value. Based on this model, a development roadmap with quantitative performance targets and near-term milestones is presented.

Abstract

Compton scattering in Positron Emission Tomography (PET) is commonly seen as an important degrading factor for the image quality. In particular, object scattering is known to reduce  image contrast and distort quantification. Conventional scatter correction techniques do not model scatter directly in the system matrix, despite its potential to better handle low-count data scenarios. In this work, we propose an analytical model for single Compton scattering events and subsequently derive expressions of the system matrix and sensitivity. We validate the obtained expressions with Monte Carlo simulations. Ongoing work includes studying the impact of such approach by using low-count simulated data, and a comparison of the proposed method with conventional scatter correction techniques.

Abstract

Dual energy computed tomography (DECT) utilizes transmission measurements through a sample with two distinct X-ray spectra to produce a two material decomposition. By introducing a path-length constraint to DECT measurements, it is possible to obtain a three material decomposition. In both cases, the stability of the resulting decompositions and noise properties of the virtual monochromatic images (VMIs) are sensitive to acquisition choices like beam energy. We aim to extend a theoretical noise-propagation framework for choosing optimal beam energies and VMI energy for two material decompositions to the three material case. We then demonstrate the utility of this framework for a proposed use-case of contrast enhanced breast tomography. In particular, by assuming linear propagation of noise through material sinogram estimation and filtered backprojection, a compound Poisson noise process in X-ray measurements, and Gaussian-distributed distance constraints, we obtain analytic expressions for the Jacobian of the material sinogram-to-data transformation, the variance of VMI pixels, and the VMI contrast to noise ratio (CNR). We then use the VICTRE software to generate a realistic computational breast phantom, modeled as consisting of glandular tissue, adipose, and iodine contrast agent. For this phantom, we find that tube potential pairings of 30-40 kVp and 55-80 kVp give a high value of the material sinogram-to-data transformation Jacobian determinant, indicating good stability and minimal noise propagation. For this phantom and spectra, we compare analytically predicted and empirically realized VMI CNR versus monoenergy, and find good agreement between them with a maximum CNR of 9.6 for a VMI monoenergy of 65.5 keV in both cases.

Abstract

Multi-pinhole SPECT enables static data acquisition without rotating the detectors, reducing patient burden. In brain SPECT, particularly for diagnosing neurodegenerative diseases like Alzheimer's, high spatial resolution is critical for visualizing minute regional functions. However, increasing the number of pinholes to secure sufficient sensitivity causes projection overlaps on the detector. Furthermore, physical factors such as gamma-ray attenuation and scattering within the object degrade the quantitative accuracy of reconstructed images. In addition to these issues, a significant trade-off exists regarding the pinhole aperture size. While a larger diameter (e.g., 5 mm) increases photon counts and reduces statistical noise, it exacerbates projection blur and degrades spatial resolution. Conversely, a smaller diameter (e.g., 2 mm) preserves resolution but suffers from increased noise. Previously, we proposed a convolutional neural network (CNN) approach using U-Net++ for a three-detector multi-pinhole SPECT that simultaneously performs overlap separation and various physical corrections. In this study, we applied this framework to brain SPECT to determine the optimal aperture size by comparing 2-mm and 5-mm pinhole diameters. Our results demonstrated that the 2-mm aperture, which maintains spatial resolution, outperformed the 5-mm aperture across quantitative image quality metrics, including NMSE, PSNR, and SSIM, confirming its superiority for detailed brain imaging.

Abstract

We have developed a 3-layer staggered LYSO de tector with DOI capability within the project SIRMIO (Small animal proton irradiator for research in molecular image-guided radiation-oncology) for proton range verification via PET imag ing. To evaluate the feasibility of detecting prompt gamma rays for quasi-instantaneous monitoring of proton beams, we propose a workflow that includes the validation of optical simulation settings and the implementation of the saturation effect of the SiPM photonsensor in the Geant4 simulation. Measurement data with ¹⁵²Eu, ²²Na and ¹³⁷Cs radioactive sources are utilized for modeling the saturation effect of SiPM, and also for validating optical boundary settings in Geant4. The measured spectra of ²²Na and ¹³⁷Cs point sources are well reproduced with the proposed simulation workflow. The responses to prompt gamma rays are then simu lated. According to the floodmaps and profiles, LYSO pixel can be identified at 2.2, 4.4 and 6.2 MeV, but the contrast decreases strongly when energy reaches 6.2 MeV. The peak-to-valley ratios drop by up to 53% at 4.4 MeV and 62% at 6.2 MeV (relative to 2.2 MeV), primarily due to the energy deposition occurring more frequently in multiple LYSO pixels. These predictions suggest the detector’s potential for detecting prompt gamma rays, while accurately recovering or rejecting inter-crystal scatter events is essential to maintain detector performance at higher gamma energy.

Abstract

Mesh-type Reference Computational (MRCP) Phantoms provide anatomically realistic tetrahedral organ structures, but their practical use in OpenGATE-based imaging and dosimetry simulations remains limited. In our previous work, a simplified tetrahedral phantom was imported using a tessellated-solid-based workflow and evaluated with heart and body geometries. In this study, we extended the approach to a whole-body tetrahedral phantom framework using parameterised tetrahedral geometry construction and region-aware dose scoring. The updated workflow reads mesh, material, organ-region, and visualization information, constructs the complete phantom in OpenGATE, and enables selected-organ three-dimensional dose scoring while preserving full-phantom particle transport. A copy-number-based filtering method was introduced to associate tetrahedral elements with anatomical regions, avoiding material-name-based selection. The complete phantom, including all anatomical organs, was successfully loaded with a memory usage of 23.84 GB. Three-dimensional deposited energy and dose distributions were generated for selected organs. Multi-threaded performance was evaluated using 8 million x-ray beam particles, reducing computation time from 2453.13 s with 1 CPU thread to 429.43 s with 8 CPU threads, corresponding to a 5.7-fold speed improvement. In the 8-thread simulation, dose scoring for four selected organs, including blood, heart wall, and both lungs, was completed in approximately 5 minutes. These results demonstrate the feasibility of multi-threaded, organ-specific MRCP dosimetry and provide a basis for extending mesh-type phantom simulation to CT, PET, SPECT, and linac workflows in OpenGATE.

Abstract

This study integrates GEANT4/GATE Monte Carlo simulations with hardware experiments to investigate and correct inter-crystal scattering (ICS) in Lutetium Yttrium Oxyorthosilicate (LYSO) Positron Emission Tomography (PET) detectors. In the simulation phase, 10⁷ valid events from a 511 keV gamma-ray point source were tracked across a 16 x 16 LYSO array (2.1 mm pitch) to analyze 3D energy spread and depth of interaction (DOI) characteristics. To correct ICS-induced parallax errors, we designed a physics-guided Convolutional Neural Network (CNN) with an energy masking constraint to recover the initial interaction coordinates. Furthermore, we established a physical experimental platform using arrays of various pitches (2.1 mm, 3.2 mm, and 4.1 mm). Deploying the pretrained CNN on physical hardware and conducting preliminary iterative training with empirical data demonstrated the predictive potential of machine learning for measured signals. During system image reconstruction, perfect CNN prediction of ICS events could theoretically improve spatial resolution from 1.45 mm to 1.21 mm. By synergistically applying algorithmic ICS correction and DOI information, the full width at half maximum (FWHM) can be further compressed to 0.99 mm. This study provides a conceptual solution and initial attempt for mitigating inter-crystal scattering in future high-resolution PET detectors.

Abstract

SPECT is useful for investigating a range of neurological and psychiatric conditions, such as Parkinson’s disease, epilepsy and schizophrenia. Most clinical brain SPECT studies are performed with rotating gamma-cameras with parallel-hole collimators, which results in sub-optimal image quality. Multi-pinhole (MPH) collimators can provide an improved sensitivity vs. resolution trade-off, while Cadmium-Zinc-Telluride (CZT) detectors offer better intrinsic spatial resolution as well as superior energy resolution compared to scintillation detectors. In collaboration with Kromek Ltd, we have developed a molecular breast imaging system based on CZT detectors and MPH collimators. Here we report initial results on designing a dedicated brain SPECT scanner using similar technology. The system will consist of a polygon of planar detectors and will be completely stationary, allowing for dynamic imaging. We performed computer simulations for various design configurations, based on a 20-cm diameter spherical field-of-view (FOV), assuming an intrinsic detector resolution of 0.4x0.4 mm and a depth-of-interaction (DOI) resolution of 1 mm. The simulations were based on an analytical system model, and images were reconstructed using an ML-EM algorithm. Our initial evaluations involved two different digital phantoms: A grid of 7x7 point-sources with a non-zero background; and the Hoffman brain phantom. With the target resolution set to 9 mm at the center and a degree of multiplexing of ~30%, the sensitivity was ~6×10^-4. An average reconstructed resolution of ~5 mm FWHM was obtained, going from 9 mm at the center down to 2-3 mm towards the edge of the FOV. In conclusion, our initial results show that high image quality can be obtained with this type of SPECT system.

Abstract

Silicon photomultipliers (SiPMs) coupled to scintillators are widely used in medical imaging, particularly in fast-timing applications such as time-of-flight positron emission tomography (TOF-PET) . While significant progress has been made in modeling detector response and correlated noise, recent advances in SiPM technology and fast readout electronics require a reassessment of the full analog signal formation process. In particular, the transition from idealized digital timestamps to realistic analog signals introduces information loss due to bandwidth limitations, electronic noise, and correlated effects such as optical crosstalk.
We present a comprehensive simulation framework that bridges this gap by modeling both analog waveform formation and generation-dependent optical crosstalk with explicit temporal structure. The framework enables systematic separation of different degradation mechanisms affecting timing performance. We investigate the impact of bandwidth limitation, electronic noise, and correlated noise on coincidence time resolution, as well as the information loss transitioning form digital to analog signals.
Our results show a clear degradation of timing resolution when moving from ideal timestamps to realistic analog readout conditions, in agreement with experimental measurements. Furthermore, we observe that optical crosstalk can, in certain scintillator configurations, improve timing performance by effectively enhancing early signal amplitude, while simultaneously degrading higher-level temporal information relevant for event characterization. The presented framework provides a flexible tool for disentangling these effects and guiding SiPM and readout optimization for future fast-timing detector systems.

Abstract

MERMAID is the first PET scanner prototype dedicated to study small aquatic animals. Particularly, it aims to support the use of adult zebrafish as a model of human diseases. The current prototype, which consists of four detection modules composed of LYSO:Ce crystals, has demonstrated the capability of providing images with submillimeter resolution. An extension of MERMAID by another four modules is planned, aimed  to increase the system sensitivity and allow for shorter scan times, making it more suitable for in-vivo imaging. In the presented work, various configurations for the extended MERMAID system are characterized by means of simulations and tailored reconstruction framework; modifications of imaging protocol for the system are  investigated.

Abstract

Single cell radiation measurements are an integral part of interrogating the biological heterogeneity of cancerous tumors. Identifying cell types present in bulk tumors via FDG uptake can elucidate cell types resistant to treatments, that would potentially be obscured via macroscopic FDG PET imaging. The BetaBox 1.0 was previously constructed to perform these single cell radiation measurements, but suffers from a lack of scale and inability to isolate cells after radiation measurement for further study. To address these challenges, we are constructing the BetaBox 2.0, which will be a high-sensitivity, low-noise device capable of performing many simultaneous single cell radiation measurements via the detection of beta particles in a pixelated array of organic scintillators coupled to SiPMs. The design of BetaBox 2.0 is considered and modeled in GATE simulations, which are used to determine an optimal organic scintillator thickness for the device. These simulations are then expanded on via experimental measurements taken with a SiPM pixel read out by PETsys readout electronics. These datasets jointly suggest the BetaBox 2.0 will be capable of isolating the signal from single cells, and quantitatively assessing their radiation uptake.

Abstract

Optical photon transport in scintillator-based imaging detectors dominates TOPAS Monte Carlo wall time because each primary spawns tens of thousands of secondary photons, and TOPAS itself has no released GPU backend. We present MetalOpticalPhoton, a TOPAS extension that ports the optical physics chain to Apple Metal hardware ray tracing while reusing existing TOPAS material/surface/geometry tables. On an Apple M3 Ultra (80 GPU cores), execution wall time on a 100K-event scintillator-plus-lens benchmark is 25–28× shorter than a 30-thread M3 CPU baseline and about 10× shorter than a 90-thread AMD 7995WX baseline, sustaining roughly 33×10⁶ photons/s GPU throughput. Two corrections to vanilla OpenTOPAS—a parallel-world envelope coincident-boundary loop suppressing 75–80% of the sensor-integrated CPU optical fluence, and a primary polarization initialization typo—were required for a self-consistent CPU reference and are relevant to TOPAS optical Monte Carlo independently of GPU acceleration. After patching, CPU–GPU agreement reaches the sub-1% level reported for production GPU optical Monte Carlo (sensor-integrated fluence sum ratio 0.993 ± 0.002 over five seeds).

Abstract

We have developed a 3-layer staggered LYSO DOI detector for the SIRMIO (Small animal proton irradiator for research in molecular image-guided radiation-oncology) project, with a crystal pitch of 1.0 mm. Due to the small crystal size, inter-crystal scatter (ICS) events degrade spatial resolution, while discarding them causes significant sensitivity loss. In our detector, ICS events account for approximately 50% of all interactions at 511 keV and result in a median lateral positioning error of approximately 2 mm. To solve this problem, we propose a dual-head convolutional neural network (CNN) for separating the 8×8 MPPC total response of two-pixel ICS events into two individual response matrices, from which the pixel positions and energy fractions can be obtained—a precondition for recovering the ICS events. The CNN correctly identifies both scattering pixels in 69.3% of test events, and in 26.4% of cases one pixel is correctly identified. Among the events with one pixel mismatched, the lateral error of the misidentified pixel falls within 1.0 mm in 91.0% of cases. Meanwhile, the median energy fractions prediction error is 0.016 per pixel. The results demonstrate that CNN-based MPPC response decomposition can serve as the enabling foundation for complete ICS recovery in 3-layer staggered DOI detectors, addressing a limitation explicitly acknowledged in recent physics-based ICS studies.

Abstract

PET systems require (among others) normalization corrections for the geometric sensitivity of the scanner, the result of the spatial arrangement of the detector crystals and energy readout area. It is estimated for a given extent of the photon energy acceptance window, and must be re-estimated if the lower energy threshold (LET) is changed, e.g. when optimizing the number of noise-equivalent counts for a given object, or to improve joint estimation of activity and attenuation. To investigate the effect of changing the LET on the geometric sensitivity, we performed a simulation study in SimSET and compared sensitivity sinograms, reconstructions and properties of accepted photons. We found that the effects are dependent on detector location and photon incidence angle, and likely due to the finite extent of the energy readout area.

Abstract

Monte Carlo simulations play a vital role in many aspects of proton therapy treatment, including proton depth of penetration monitoring. A key limiting factor of full simulation engines is computation time: simulations can take several days, far longer than the time between treatment fractions. This project aims to develop a Monte Carlo simulation software, MIMOSA, for prompt gamma-ray based proton range verification, that is fast enough to be deployed before the next treatment fraction. Clinically acceptable gamma pass rates above 99% for 2%/2 mm and 97% for 1%/1 mm with respect to TOPAS were achieved on both homogenous and heterogenous phantoms. To benchmark computational speed, state-of-the-art GPU codes MOQUI and FRED were tested on the same computer. Comparatively, MIMOSA achieved a notable speed increase, able to simulate 1e9 primary protons with a 200 MeV energy in approximately 3 minutes. Along with maintaining accuracy and increasing computational speed, the project uses the hardware agnostic SYCL standard, and will be available with an open-source license and extensive documentation. Ongoing work is devoted to simulating realistic radiotherapy treatment plans, and including prompt gamma-ray transport and detector response.

Abstract

Computed tomography image quality is degraded by scattered photons, which reduce image contrast and quantitative accuracy. Time-of-flight-based scatter rejection has been proposed as a method to distinguish primary and scattered photons using their arrival-time differences, but the influence of beam filtration under polychromatic x-ray spectrum has not been fully evaluated. This study investigated the effects of tube voltage and beam filtration on time-of-flight-based scatter rejection using Monte Carlo simulation. X-ray spectrum were generated at 40 and 120 kVp, and four filtration conditions were compared: no filter, 1.06mm aluminum, 0.15mm copper, and 0.6mm tin. Polymethyl methacrylate phantoms with thicknesses of 50 and 200 mm were simulated. Scatter rejection performance was evaluated using the scatter rejection rate as a function of timing jitter from 0 to 1000 ps, and the area under the curve was used as a summary metric. At 120 kVp, tin filtration increased the area under the curve by 11.1% in the 200mm phantom and by 30.6% in the 50mm phantom compared with the no-filter condition. In contrast, at 40 kVp, the filtration effect was small, with a maximum area under the curve difference of 0.010 in the 50mm phantom. These results indicate that tin filtration improves timeof-flight scatter rejection at 120 kVp, whereas filter optimization has a limited effect at 40 kVp.

Abstract

Hadron therapy is a highly conformal cancer treatment technique that enables superior dose delivery to the tumor while sparing surrounding healthy tissues. However, its full clinical benefit is still limited by uncertainties in proton range within tissue, mainly due to the indirect estimation of ion stopping power from X-ray CT images. These uncertainties lead to safety margins that reduce treatment precision. Real-time, high-accuracy ion-range verification is essential to improve dose delivery and further enhance the therapeutic potential of proton therapy. Among the available approaches, PGI and PET are regarded as the most promising, as they provide complementary information. PGI offers direct correlation with the Bragg peak during irradiation, whereas PET yields high-resolution activity distributions from beta+ emitters produced along the proton path.

In this work, we report recent developments and experimental results obtained with a hybrid PGI-PET system for in situ proton range verification. The system is based on four high-efficiency two-plane Compton imagers arranged in a cross configuration, specifically optimized to enhance PGI sensitivity at the expense of a reduced PET field of view. Experimental campaigns were carried out at commercial IBA Proteus Plus and Proteus One proton therapy facilities using 100 and 125 MeV pencil beams delivered at minimum clinical currents. Polyethylene phantoms were irradiated at controlled positions and data were acquired during both in-spill and off-spill periods.

The first reconstructed 2D images demonstrate that PGI can directly identify the Bragg peak position through the maximum of the PGI distribution over a broad field of view, while PET provides a narrower and higher-resolution distribution associated with the induced beta+ activity. These results confirm the feasibility and complementarity of the hybrid PGI-PET approach and support its potential for accurate, clinically relevant proton range verification.

Abstract

Cell tracking in vivo is a key challenge in translational medicine. Positron emission tomography (PET) and magnetic resonance imaging (MRI) offer complementary strengths for this task, yet the validation of tracking algorithms against a reliable ground truth remains unresolved. Fluorescence microscopy can fulfill this role, but its integration into the highly constrained PET/MRI environment poses significant technical challenges. Here, we present a fluorescence microscope (FM) designed for simultaneous operation inside a small-animal PET/MRI scanner, with all custom structural and housing components 3D printed in polylactic acid (PLA). The assembled instrument has a total length of 95 cm and fits within the 35 mm PET insert bore, accommodating sample diameters of up to 21 mm. The imaging probe is rotated via a non-ferromagnetic, friction-driven mechanism, enabling multi-angle acquisition without repositioning the sample, and reliable rotation over the full 360° range was demonstrated under manual control. The system follows an epi-fluorescence configuration with 10× magnification, an excitation wavelength of 488 nm, and a 500 nm longpass filter in the detection path. Characterization with a calibration slide and a Convallaria majalis test specimen yielded a pixel pitch of 0.144 μm/pixel (6.96 pixels/μm), a spatial resolution of 2.32 μm (FWHM), a signal-to-noise ratio (SNR) of 7.21, and a field of view (FoV) of 0.55 × 0.31 mm2 (3816 × 2160 pixels), sufficient to resolve individual cells. These results establish the feasibility of integrating fluorescence microscopy into a PET/MRI scanner and provide the basis for future multimodal cell tracking studies with a reliable optical ground truth.

Abstract

We report a design study for a combined preclinical scanner that integrates PET, SPECT, and MRI for simultaneous acquisition of PET, SPECT, and MR images for broad-band multi-isotope imaging capability critical for theranostics applications. The PET subsystem provides a very high-sensitivity and benefits from a wide variety of existing tracers. The SPECT subsystem allows for broad-band multi-isotope imaging capability critical for imaging therapeutic radioisotopes and their daughters. MRI offers unmatched soft-tissue contrast and high-resolution anatomical reference. The PET/SPECT insert consists of 12 detector rings. The central four PET rings are integrated with four rings of micro camera elements (MCEs) that are “looking at” the object through the small gaps between the PET detector blocks. This compact design provides approximately 16% PET sensitivity at the center and 1-mm PET imaging resolution across the 8-cm transverse by 28-cm axial FOV. The 64 CZT SPECT cameras provide approximately 1-mm spatial resolution over a 6-cm transverse and 9-cm axial FOV. To support simultaneous PET/SPECT operation, the design incorporates our previously developed anti-coincidence rejection strategy to suppress PET-induced contamination in SPECT data. Prior F-18/Tc-99m PET/SPECT simulations demonstrated up to 83.7% rejection of scatter contamination in projection-domain analysis, supporting the feasibility of acquisition-level cross-modality rejection. These design elements resulted in a compact and high performance tri-modality imaging platform for simultaneous PET/SPECT/MR imaging for preclinical theragnostic applications.

Abstract

We previously proposed a Broadband Ultrasensitive Gamma Imager (BUGI) that uses a sensing collimator imager to encode and detect gamma rays without a metal collimator. When combined with fast scintillators, this approach enables simultaneous high-sensitivity SPECT and time-of-flight-capable PET imaging, with potential applications in monitoring alpha-particle radiopharmaceutical therapy (RPT). Unlike conventional simultaneous PET/SPECT systems, which use high-energy collimators that reduce PET sensitivity to SPECT levels, BUGI raises SPECT sensitivity toward that of PET.In this study, we evaluated two preclinical BUGI detector designs using GATE Monte Carlo simulations: a pinhole-inspired geometry and a parallel-hole-inspired geometry. The scanner consists of an octagonal arrangement of 8 detector blocks, each containing 1 × 4 arrays of self-collimating mini-modules built from 3×3×3 mm³ GAGG crystals. A dual-isotope cylindrical phantom was simulated with a ⁹⁹ᵐTc background, a ⁹⁹ᵐTc spherical tumor, and an ¹⁸F spherical tumor, with a tumor-to-background ratio of 4:1. SPECT events and PET coincidence events were reconstructed independently using a GPU-based MLEM framework.Both detector geometries produced usable simultaneous SPECT and PET images with comparable tumor contrast . The parallel-hole design showed better background uniformity, improved preservation of the cylindrical phantom shape, and sharper tumor edges.The pinhole design showed an octagonal artifact pattern that may be reduced with increased angular sampling. PET images were sharp and visually similar between the two designs. Future work will focus on developing and validating multi-energy-window SPECT and PET reconstruction methods for the BUGI system.

Abstract

Building upon previous works that explored the feasibility of super-resolution (SR) positron emission tomography (PET) by applying the principle of structured illumination microscopy (SIM), in this work we propose the concept of a dual-ring SR PET system design. Such a PET system contains a rotatable inner detector ring that serves the dual purpose of creating the needed measurement modulation for performing SIM-based SR as well as increasing event detection to eliminate a major drawback of the previous approach in which the modulator could considerably diminish the imaging sensitivity. We simulated a 2d PET system comprising 576 4.2 mm (transaxial) x 4.2 mm (axial) x 20 mm (radial) LYSO crystals on a 770 mm diameter ring. On the other hand, the inner ring consisted of repeating 2.8 mm-width attenuating material or detector and 5.6 mm-width air. The attenuating material was either 5 mm thick tungsten (W) or 10 mm thick LYSO. The outer detector was 20 mm thick LYSO. Axially, the inner ring was uniform and had the same 4.2 mm length as the outer detector ring. An SR dataset was made of three measurements acquired at three rotational positions of the inner ring. We also considered point-spread function (PSF) reconstruction when not using the inner ring, and also the situation when detector wobble motion was introduced to increase sampling. Our simulation results show that the dual-ring image can resolve 1.2 mm sources in the case of noise-free data and the image quality is visually superior to the PSF and Wobble+PSF images. For a noisy data set, it yields an overall better image quality than the other two SIM-based SR images using W or LYSO modulators for it has higher coincidence detection sensitivity. Quantitative performance evaluation based on the contrast-recovery-coefficient versus the background noise also shows superior results with the dual-ring system.

Abstract

Dedicated brain positron emission tomography (PET) systems can achieve higher sensitivity and spatial resolution than whole-body PET scanners due to their compact detector geometry. However, the reduced detector aperture introduces parallax error, making depth-of-interaction (DOI) capability important for maintaining spatial resolution uniformity across the field of view. In this work, we present a comprehensive performance evaluation of a DOI-enabled dedicated brain PET system based on light-sharing window (LSW) detector technology. The system consists of ten detector panels with an axial field of view of 280.6 mm and employs 2.0 × 2.0 × 20.0 mm³ LYSO crystal arrays with DOI encoding capability.

Abstract

General-purpose positron emission tomography scanners often face challenges with the high-frequency motion and small target volumes inherent in cardiac imaging. This study introduces a heart-dedicated positron emission tomography prototype featuring twenty detector panels in a cylindrical geometry with a 493.28 millimeter aperture and a 236.4 millimeter axial field of view. The detectors utilize lutetium-yttrium oxyorthosilicate crystals coupled to multi-pixel photon counters with depth-of-interaction capability. Performance was validated through Geant4 Application for Tomographic Emission simulations, detector bench tests, and in vivo experiments. Results demonstrate a coincidence resolving time of 234 picoseconds full width at half maximum and a spatial resolution of 1.0 millimeter under idealized conditions. Furthermore, electrocardiogram-gated imaging in rat and rabbit models successfully captured eight cardiac phases from end-systole to end-diastole, providing distinct myocardial structures without significant motion blur. These results validate the system’s high sensitivity and timing precision, establishing a foundation for extracting functional parameters such as ejection fraction. Future work will implement depth-of-interaction correction across the full field of view to maintain high peripheral resolution for precise cardiovascular research and clinical diagnostics.

Abstract

Positron emission tomography (PET) is an indispensable tool in neuroscience, yet its efficacy in early lesion detection and advanced brain research is often constrained by the fundamental trade-offs between spatial resolution and sensitivity. To surpass these limitations, this study introduces the NeuroEXPLORER with Insert (NXI), a high-performance dedicated brain PET system that integrates Virtual Pinhole (VP-PET) technology into the NeuroEXPLORER (NX) platform. We developed a comprehensive quantitative reconstruction framework incorporating a full suite of physical correction techniques tailored for the NXI’s asymmetric geometry. The system's performance was rigorously validated through Monte Carlo simulations using GATE. Point-source evaluations demonstrated a enhancement in overall sensitivity, driven by an increased proportion of high-resolution insert-scanner (IS) coincidence events in the vicinity of the insert. An ablation study performed on uniform cylinder data, demonstrated the effectiveness of the integrated correction suite, including normalization, attenuation, random, and scatter corrections. In Derenzo phantom experiments, the NXI system successfully resolved 0.6 mm hot rods, a benchmark that remained unresolvable by the standard NX configuration. Furthermore, high-resolution digital brain phantom studies confirmed the system's clinical potential, showing a 5.74% increase in data acquisition volume and a 19.31% improvement in the contrast-to-noise ratio (CNR) of gray and white matter compared to the NX system. By synergizing optimized hardware design with specialized reconstruction algorithms, this work verifies that the NXI architecture substantially elevates both sensitivity and spatial resolution, providing a robust technical pathway for the development of next-generation dedicated brain PET instrumentation.

Abstract

To improve diagnostic accuracy in hybrid breast imaging, we develop a dedicated PET/MR breast insert for integration into a clinical whole-body PET/MR system. The final insert consists of two cylindrical detector rings (⌀ = 178 mm, height = 95 mm), each comprising five rings with 32 detector blocks per ring. A sparsely populated (18/32 blocks) single-ring prototype was realized, using dual-layer, depth-of-interaction capable 10×10/9×9 LSO scintillator arrays (1.6×1.6×9.8 / 5.2 mm³) and SiPM readout connected via 30 cm flexible cables to MRI-compatible front-end electronics. Detector calibration and bias voltage optimization demonstrated changes in CRT of up to 100 ps between 42-47V. A bias voltage of 45 V was selected as a stable operating point, maintaining consistent energy resolution and coincidence count rates. Reliable crystal identification was achieved for both layers, with a peak-to-valley ratio of 3.7. At this operation point, the prototype achieved energy resolutions of 13.1 ± 1.8% FWHM (top layer) and 11.8 ± 2.4% (bottom layer). CRT depended on the applied energy window, yielding 383 ± 34 ps FWHM for 430–610 keV and improving to 349 ± 32 ps for a 511±50 keV window, while degrading to 483 ± 175 ps for 250–760 keV. Compared to previous system configurations (>500 ps), the updated readout enabled a substantial improvement in timing performance. These results demonstrate robust energy and timing performance of a sparsely populated, DOI-capable PET/MR insert under realistic integration constraints, including extended signal routing and MRI-compatible system design. The findings support the scalability towards a fully populated double-ring system currently under construction, enabling systematic evaluation of PET/MR performance under stress-test sequences and realistic breast imaging conditions within a clinical PET/MR scanner.

Abstract

This work presents a low‑cost small‑animal PET system based on BGO crystals with depth‑of‑interaction (DOI) capability. A 10×8 BGO array (1.5×1.5×20 mm³) is coupled to an MPPC array in a 2×2 configuration, and a 6‑mm light‑sharing window (LSW) encodes DOI‑dependent light sharing. Using a rapid calibration method (vertical line source + histogram matching), the detector achieves an average DOI resolution of 4.9 mm FWHM and an energy resolution of 17.4% at 511 keV. System‑level evaluation shows a spatial resolution of 1.43 mm FWHM at the center and resolves 1.5‑mm hot rods. The design eliminates labor‑intensive mechanical collimation, offering an affordable, maintainable solution for high‑sensitivity preclinical imaging.

Abstract

The J-PET (Jagiellonian Positron Emission Tomography) collaboration is developing a large Field-Of-View (FOV) Total-Body J-PET (TB-J-PET) scanner [1,2,3,4] enabling, among others, low dose imaging, short scan times, and dynamic imaging. The key distinctive feature, i.e., the utilization of cost-effective plastic scintillators, could facilitate a wider adoption of large FOV scanners. In this work, we study the benefit of combining the TB-J-PET with a Brain PET Insert (BI) via Monte Carlo simulations. We show that the proposed detector configuration has unique advantages enhancing the spatial resolution. For this purpose, we first developed a rigorous and accurate scheme to estimate spatial resolution from Maximum Likelihood Expectation Maximization (MLEM) reconstructions. We thereby address known challenges of MLEM-based spatial resolution assessment [5], which is discouraged by NEMA standards that continue to rely on the suboptimal Filtered Back-Projection (FBP). We then investigate the impact of the detector geometry on the resolution. The BI uses narrower scintillators than the TB-J-PET (4 vs. 6 mm). We show that this is a resource‑efficient approach to achieve high resolution inside the BI, since the coincidences between TB-J-PET and BI (~50% of sensitivity) have particularly low uncertainty. We also show that this detector configuration is geometrically optimal and suppresses the non-collinearity effect. We back up our findings with theoretical analyses. We also investigate other factors influencing spatial resolution, including event selection, detector scattering, phantom scattering, and positron range. We provide intuitive insights, offering compelling arguments for the experimental realization of the detector geometry under investigation.

Abstract

PET remains a cornerstone of molecular imaging,yet the trade-off between spatial resolution,sensitivity,and system versatility persists.While integrating high-resolution inserts into existing systems can enhance performance within a targeted field-of-view,conventional designs lack geometric symmetry in insert-scanner(IS)and insert-insert(II)LORs.This asymmetry necessitates exhaustive,configuration-specific normalization data,severely limiting the mobility and clinical adaptability of the insert.To overcome these limitations,we propose the Adapt-PET(Adaptive-geometry PET)design strategy,which utilizes a "Hardware-Algorithm Co-evolution" approach featuring a movable,DOI-capable insert.The insert geometry is derived by proportionally downscaling the base scanner’s transaxial dimensions while maintaining axial consistency,followed by systematic crystal subdivision.This configuration enables a unified geometric mapping strategy that projects IS and II LORs onto a standard scanner-scanner(SS)space.By exploiting intrinsic symmetries,our approach achieves geometric compression ratios comparable to baseline systems,effectively decoupling system normalization from hardware geometry and eliminating the need for re-normalization upon insert relocation.The proposed strategy was validated through GATE simulations of two prototypes integrated with a Siemens Biograph Vision scanner.Positioned at an inner radius of 200-mm,each insert comprised 24 modules of 30×40 subdivided crystal arrays(1.047×0.749×15 mm³)with a 6-bin,2.5-mm DOI resolution.Evaluations using a uniform cylinder show that the system achieves near-quantitative reconstruction accuracy.Derenzo phantom imaging results reveal that the peak-to-valley ratio can be improved by up to 20%,while torso phantom results demonstrate a substantial enhancement in the detectability of 4-mm lesions.This study establishes a practical framework for flexible,organ-targeted PET imaging through hardware-algorithm synergy.

Abstract

Poor spatial resolution and partial volume effect (PVE) limit PET’s ability to accurately quantify radiotracer uptake in small brainstem nuclei such as the locus coeruleus, one of the earliest sites to undergo neuropathological changes in the early stages of tau accumulation associated with Alzheimer’s disease. Prism-PET is designed to improve spatial resolution and mitigate the PVE by combining ultrahigh-resolution detector modules with high scanner sensitivity in a practical and cost-effective system.

Latest generation of Prism-PET detector module was evaluated using both qualitative and quantitative performance assessments. The module’s flood map was examined to assess crystal separation, pattern stability, and the reliability of crystal identification across the detector module. The tested module produced a clear and stable flood map with well-separated crystal patterns, indicating robust event localization. Quantitatively, we measured the detector’s energy resolution, depth-of-interaction (DOI) resolution, and coincidence timing resolution. The tested module achieved 10.34% energy resolution, 3.23 mm FWHM DOI resolution, and 337 ps coincidence timing resolution. In the future work, we will report on the performance of the third generation Prism-PET scanner.

Abstract

Patient motion during PET acquisition is a well-known source of image degradation. In dedicated brain PET systems such as CareMiBrain (Oncovision), patients occasionally slide downward during acquisition despite head immobilization, compromising diagnostic quality in a clinically relevant subset of studies. We present MoDeCo (Motion Detection and Correction), a modified iterative reconstruction algorithm that detects and compensates for intra-acquisition rigid head motion without requiring external tracking hardware. The listmode data are split into temporal fragments; fast preliminary reconstructions are co-registered to the first fragment to estimate rigid transformations, and a modified MLEM/OSEM algorithm applies them within each subiteration to combine the partial back-projections. Head motion was quantified on 256 [18F]FDG patient studies acquired at Hospital Clínico San Carlos (Madrid), and a qualitative clinical comparison was performed on 15 selected studies. Maximum head displacement during acquisition exceeded 5 mm in 14% of patients. In the high-motion subset (n = 10), MoDeCo was rated significantly superior to standard OSEM in 60% of cases and moderately superior in 20%; mild-motion studies (n = 5) were rated as equivalent. No case showed standard reconstruction outperforming MoDeCo. MoDeCo improves image quality in patients with significant motion while preserving equivalence in those without appreciable motion.

Abstract

Photon-counting detectors (PCDs) can perform quantitative imaging by utilizing the differences in the individual energies of X-ray photons, but this analysis requires various corrections, such as charge sharing, to determine the correct attenuation amount. The purpose of this study is to quantitatively clarify the importance of beam hardening and detector response corrections, which can be applied by software processing, when performing imaging with polychromatic X-rays using a line-sensor-type PCD having an anti-coincidence mode. Our PCD was a CdTe detector manufactured by Varex Imaging (pixel size 0.1 mm, thickness 0.75 mm). The response characteristics were calculated using the Monte Carlo simulation code EGS5, and the charge summing effect related to the "anti-coincidence mode" in the four clustered pixels was reproduced in a simple way. The evaluation method involved examining the reproducibility of quantitative values in the effective atomic number (Z_eff) images. Simulations using various virtual phantoms and experiments using a non-destructive testing system with the PCD-type line sensor were conducted. As a result, we demonstrated that by including our proposed correction procedure with the anti-coincidence mode, it is possible to generate highly quantitative Z_eff images. In conclusion, this study verified that when using PCDs, it is important to implement software-based beam hardening and detector response function corrections in addition to using the anti-coincidence modes.

Abstract

This study evaluated a lead-free BaSO₄–silicone composite for diagnostic X-ray shielding (60–100 kVp) using LD-V1 film, combining MCNP6 simulation with experimental validation. Evaluation of a 2 mm thick composite demonstrated shielding efficiencies ranging from 68.79% to 83.50%. Simulated results showed good agreement with experimental measurements, with a maximum deviation of 10.28%. Furthermore, the two-dimensional dose distribution map revealed an average radiation reduction of approximately 60.34% across the material. This result confirms effective attenuation of X-ray radiation across the irradiated field. The dose reduction was also spatially uniform over the entire protected area. In addition, the consistent dose reduction pattern indicates a homogeneous dispersion of BaSO₄ particles within the silicone matrix. These findings establish the BaSO₄–Si composite as an effective lead-free alternative and demonstrate the utility of LD-V1 film for both quantitative dosimetry and spatial validation in clinical X-ray systems.

Abstract

DXA has evolved beyond standard BMD measurement to include sarcopenia diagnosis and bone quality assessment via the TBS. As TBS is now a recognized modifier in the WHO FRAX Tool , image resolution has become the paramount technical factor determining the performance of DXA systems. This study evaluates the image resolution of iNSiGHT C510 (Osteosys, Korea), the novel DXA model designed for high-resolution whole-body imaging within 3 minutes.

The image resolution of iNSiGHT C510 was quantitatively assessed using  a  line phantom. The system achieved a vertical resolution of 0.63 lp/mm and a horizontal resolution of 0.9 lp/mm, demonstrating superior resolution capabilities compared to the industry-leading GE Prodigy. Anatomical discriminability was further verified using the spine phantom  with increasing BMD levels. Analysis of peak-to-valley ratios, which represent the contrast clarity between bone and intervertebral discs, confirmed that iNSiGHT C510 maintains excellent discriminability despite its longitudinal scan architecture. These quantitative results suggest that the system minimizes ROI setting errors, ensuring high clinical reproducibility for TBS calculations that are highly sensitive to subtle signal variations.

Abstract

Beam hardening in cone-beam computed tomography (CBCT) of high-density rock causes cupping artifacts and nonlinear CT value deviations, hindering quantitative mineral analysis. Conventional deep learning methods often over-smooth microstructures due to a lack of physical constraints. We propose a physics-informed Retinexformer framework that reformulates the polychromatic X-ray degradation as a multiplicative Retinex model. Under this paradigm, the non-uniform hardening field is treated as a degraded illumination map, while the pristine rock structure represents the ideal reflectance. A pre-trained Denoising Diffusion Implicit Model (DDIM), combined with mask-guided moment matching, is utilized to explicitly generate a statistically aligned physical hardening field prior. Furthermore, a multi-scale feature encoder extracts high-frequency structural details from the residuals between the input and the prior to guide the denoising process. Experiments on a 10-cm diameter underground rock dataset show that our method significantly corrects CT value deviations while preserving the fine structure. Quantitatively, it achieves a PSNR of 37.13 dB and an SSIM of 0.940, reducing RMSE to 0.0149—significantly outperforming SOTA baselines (NAFNet, Retinexformer, MambaIRv2). Notably, our model achieves this with only 2.00M parameters and 19.42G FLOPs, offering a highly efficient and interpretable solution for CT artifacts correction in dense media.

Abstract

Dual-energy X-ray absorptiometry (DXA) is the clinical standard for bone mineral density (BMD) assessment. Modern wide fan-beam DXA architectures (fan angle ≥ 20°) shorten whole-body acquisition to under three minutes but enlarge the irradiated volume, raising the scatter-to-primary ratio with field size. Conventional DXA omits anti-scatter grids, relying on the long object-to-image-detector distance (OID > 580 mm) for scatter rejection via the air-gap effect. Whether this assumption holds for wide fan-beam DXA has not been directly tested. We investigated residual scatter in an experimental wide fan-beam DXA system using three complementary experiments on step-wedge and anthropomorphic phantoms. First, we implemented an outer-channel scatter probe and detected measurable scatter at outer channels even at OID > 580 mm. Second, with an 8:1 anti-scatter grid we quantified the dual-energy mass attenuation difference Δ(μ/ρ); at the maximum phantom configuration (25 cm polycarbonate + 4.9 cm aluminum), scatter caused an inversion of Δ(μ/ρ) in the no-grid configuration (−0.018 cm²/g); the 8:1 grid restored Δ(μ/ρ) to 0.076 cm²/g — a partial recovery only. Third, a 14:1 grid successfully suppressed the spurious BMD elevation. Compared to the no-grid baseline which showed a severe thickness-dependent BMD bias (coefficient of variation of approximately 19%), the 14:1 grid reduced the variability to 3% across 0–140 mm of added polycarbonate layers. These findings establish that explicit scatter compensation — whether through anti-scatter grids or software-based correction — is essential for thick-patient measurements at wide fan-beam DXA.

Abstract

Contrast agents play an important role in spectral CT because they enhance the visibility of vessels, organs, tumors, and lesions, thereby improving image interpretation and lesion identification. However, accurately quantitative contrast-agent decomposition remains challenging. Conventional material decomposition algorithms still suffer from limited quantitative accuracy even in contrast-agent imaging. This problem becomes more severe when multiple contrast agents coexist. To address this limitation, we propose an empirical multi-agent decomposition method. The proposed method first applies conventional empirical decomposition to obtain initial decomposition results and estimate the corresponding decomposition errors. Then, bivariate error surfaces are constructed to model the relationship between iodine and gadolinium concentrations and their decomposition errors. These fitted error surfaces are subsequently used to compensate for the errors produced by conventional empirical decomposition. Simulated experiments demonstrate that the proposed method can simultaneously decompose iodine and gadolinium with improved quantitative accuracy, and the global MAPE and NRMSE are effectively reduced compared with the image-domain decomposition method for the test phantoms.

Abstract

X-ray computed tomography (CT) relies fundamentally on accurate transmission measurements to reconstruct meaningful attenuation maps. However, detector response non-linearity systematically biases transmission measurements away from the values predicted by the polychromatic forward model, propagating these errors into reconstructed data. To address this, a combined spectral modelling and empirical calibration framework was developed for a high-resolution flat-panel detector, enabling prediction of measured transmission values across a broad range of sample thicknesses and compositions. A polychromatic spectral model was constructed by combining tube-spectrum estimates with Beer–Lambert attenuation using NIST/XCOM coefficients, convolved with an effective CsI scintillator response function. Predicted transmission values were compared against experimental measurements obtained using high purity elemental foils spanning a wide range of thicknesses. Consistent and reproducible discrepancies between modelled and measured values revealed a structured detector response non-linearity not captured by the forward model. A first-order empirical correction was derived from the model–experiment residuals, yielding a fitted detector-response exponent of α = 1.47 applied to model predictions to reproduce experimental detector behaviour. The calibrated model can therefore generate realistic transmission values consistent with measured detector output across a range of sample configurations. These findings highlight the broader potential of this framework, particularly for dual-energy CT, where material decomposition relies critically on the relative contrast between acquisitions at different tube voltages. A calibrated forward model of this kind offers a pathway toward improved relative CT threshold accuracy, supporting more reliable separation of materials with overlapping attenuation properties and more robust quantitative interpretation of reconstructed data.

Abstract

This paper develops a pipeline for generating synthetic cone-beam computed tomography (CBCT) projections from real dental CBCT acquisitions, providing a reference for supervised learning. To this end, we segment and synthesize fine structures from real CBCT acquisitions to obtain a material map of the head volume. Using the characteristics of the CBCT scanner, notably the effective spectrum, we simulate the acquisitions by modeling a polychromatic X-ray source and adding compound Poisson noise. The results demonstrate the realism of the generated synthetic CBCT acquisitions, as well as the plausibility of the simulated artifacts.

Abstract

The O-arm fluoroscopy machine, an application of cone beam computed tomography (CBCT) for interventional procedures, lacks a dedicated image quality (IQ) phantom for system image performance evaluation. This study aimed to investigate the viability and limitations of using the SedentexCT IQ phantom, a phantom originally designed for dental CBCT, to assess the image quality of resultant CBCT images of an interventional O-arm unit. Six CBCT image quality tests were performed (i.e. Uniformity, Image Noise, Geometric Accuracy, Voxel Intensity Values, Contrast Resolution, and Spatial Resolution), and corresponding parameters were assessed as baseline data in the low dose, standard dose, high dose, and enhanced cranial modes of the O-arm machine. The procedures were structured to the EFOMP-ESTRO-IAEA CBCT quality control protocol. An observed limitation of the phantom was the evaluation of the spatial resolution (lp/mm), since the lowest resolution of the SedentexCT IQ phantom is 1.0 lp/mm. Overall, since the phantom was applicable for the required CBCT image quality tests, the SedentexCT IQ phantom can be considered as a viable tool for image quality testing of the O-arm machine.

Abstract

The purpose of this work is to validate basis material decomposition in breast cancer screening mammography using a conventional single-kV scan. We employ mammography system physics modeling and a pathlength constraint to enable estimation of the basis material images. Using 3D breast phantoms and low noise Monte Carlo projection data generated with the FDA VICTRE software, we decompose transmission data into pathlength basis images of adipose and glandular tissue. The decomposition recovered adipose and glandular maps with mean absolute errors of 3.47 mm and 1.83 mm respectively. Accuracy along the skinline and nipple is degraded due to material mismatch. The results indicate material decomposition in mammography using a single energy acquisition is feasible to enable the quantitative recovery of adipose and glandular material maps.

Abstract

Time-of-flight dual-energy X-ray absorptiometry (TOF-DXA) is proposed as a method for enhancing material decomposition by incorporating timing information into conventional dual-energy attenuation measurements. In this study, a 64-channel energy and timing acquisition system was developed for TOF-DXA.

The detector consisted of a 64-channel LYSO array with a pixel size of 3 × 3 × 3 mm³, coupled to four 16-channel Broadcom NUV-MT SiPM arrays. Detector signals were split into independent energy and timing paths. For the energy path, signals were processed through differential amplifiers and digitized using 10-bit multichannel analog-to-digital converters. Timing signals were processed using front-end comparators and a Kintex-7 FPGA-based tapped delay line (TDL) time-to-digital converter. The TDL was implemented using a single-chain architecture based on CARRY4 primitives.

This dual-path configuration enabled synchronized acquisition of energy and timing information on an event-by-event basis. The FPGA logic measured energy values through peak detection and synchronized them with timing data stored in a timing FIFO to ensure event-by-event alignment. The energy resolution was evaluated using a Co-57 source at 122 keV, while the timing resolution was evaluated in coincidence with a reference detector using a Na-22 source. X-ray spectra were acquired using an X-ray generator operated at 83 kVp and 0.2 mA with a Ce filter (K-edge: 40.4 keV) to evaluate system performance in the DEXA energy range.

The energy resolution was 16.3% at 122 keV, and the coincidence timing resolution was 104 ps at 511 keV. The low- and high-energy peaks were clearly resolved in the acquired X-ray spectrum.

These results demonstrated that the developed 64-channel acquisition system could simultaneously measure energy and timing information for TOF-DXA. Optimization of the acquisition board is currently in progress, and comprehensive system performance results will be presented at the conference.

Abstract

Static computed tomography (SCT) enables fast, task-specific imaging through ultrafast and flexible firing among distributed focal spots. However, it introduces irregular and incomplete data acquisition, which poses great challenges to reconstruction. In this work, we propose Geometry-Aware Reconstruction Method (GARM) for SCT across diverse scanning strategies. The Geometry-Guided Dimension Reduction (G2DR) Network first transforms irregular 3D projections into slice-wise 2D fan-beam projections using geometric information derived from ray positions. A 2D dual-domain network then suppresses sparse-view artifacts guided by geometric and Radon-space information. Simulated experiments demonstrate that GARM achieves high-quality reconstruction from irregular and incomplete projections with strong generalization capability across diverse scan strategies.

Abstract

Compton cameras are imaging systems that detect Compton-scattered gamma rays using two position-sensitive detectors, typically called the scatterer and the absorber. 
Each interaction pair defines a Compton cone enabling three-dimensional image reconstruction.
In this work, we investigate how the selection of low scattering angle events influences image quality. 
A GATE-based simulation of a Si/CdTe Compton camera system was developed to generate Compton cone data. 
A 511 keV rod phantom with source cylinders of radii from 1.5 mm to 0.75 mm was implemented, and 100 M primary photons were simulated. 
Images were reconstructed from subsets obtained by applying a maximum threshold on Compton electron energy in the scatterer. 
Energy selections of 5, 10, and 15 keV (corresponding to scattering angles of 8°, 11.4°, and 14°) were used, along with a reference dataset without any energy cut for standard CC reconstruction comparison.
We show that for two rods of 1.5 mm and 1.25 mm radius, the peak-to-valley ratio (PVR) without any energy selection is approximately 3.5. By selecting energy deposits below 15 keV (14°) the PVR nearly triples, with further improvements observed for lower energy thresholds, reaching a value of 13.8 for a 5 keV energy cut
For smaller rod radii, the PVR remains below 3, although low-energy event selection still yields enhancements of approximately 60%, 30%, and 17% for radii of 1 mm, 0.75 mm, and 0.51 mm, respectively.
These findings support event-selection strategies for CC systems capable of triggering at low energies in the scatterer.
Restricting to the most informative events can reduce computational load and improve reconstruction performance, with the strongest benefits observed for rods with radii above 1 mm.

Abstract

Dual-view SPECT planar imaging is a widely used method in clinical bone scan screening due to its efficiency and low radiation dose compared to whole-body SPECT/CT imaging. However, its diagnostic efficacy is often limited by anatomical overlap and the absence of depth information. Achieving high-quality 3D image from only two projections remains a challenge for traditional methods. Furthermore, existing deep learning-based approaches for ultra-sparse-view SPECT reconstruction are mainly designed for cardiac imaging and cannot be directly adapted to bone scintigraphy. To address these limitations, we propose a deep learning framework for 3D SPECT reconstruction from orthogonal biplanar projections and CT Scout views. The framework comprises a 3D condition map generation module and a structure-guided 2D conditional diffusion module, facilitating a two-stage reconstruction process including an initial coarse reconstruction followed by iterative refinement via a diffusion model. Quantitative results demonstrate that the proposed method achieves better recovery coefficients for bone tissues and metastatic lesions, as well as improved LPIPS scores for perceptual quality. Qualitative assessments further reveal that the reconstructed images exhibit highly realistic textures and shows strong agreement with standard full-view scans. This method provides a fast, low-radiation 3D reconstruction strategy with considerable potential to optimize clinical bone scan workflows, enhance diagnostic accuracy and improve patient comfort.

Abstract

Reconstructing the emission distribution of prompt gamma radiation in time and space offers a unique advantage: it reveals the primary particle motion inside the irradiation target during particle therapy. Similar to typical tomographic emission reconstruction scenarios, the measurement process based on Prompt Gamma Timing can be described by a linear system model and Poisson-distributed measurement data. However, the structure of the emission domain (emission time and position of the prompt gammas) and the detection domain (detector index and event time difference relative to a primary particle signal) are unique. In this work, we inspect some features of this inverse problem and analyze related fundamental limitations based on reconstructions with ideal data and the singular value decomposition of the system matrix. The results show a need for regularization and help identify possible improvements to the approach. Additionally, understanding this inverse problem in-depth also supports the development of other Prompt-Gamma-Timing-based techniques.

Abstract

The Software for Tomographic Image Reconstruction (STIR) is a long-standing open-source platform for PET and SPECT reconstruction that has supported tomographic imaging research for more than two decades. Public STIR documentation dates to the early 2000s, and the project remains actively developed, with a current public release in November 2025 and a GitHub repository reporting 9,409 commits, 146 stars, and 108 forks.   Recently there have been major contributions, that have expanded substantially the framework in scanner support, motion correction, and raw-data interoperability. Advances include support for the Siemens Biograph Vision Quadra long axial field-of-view PET/CT system, a data-driven rigid head motion-correction workflow for brain PET, and PETSIRD as a vendor-independent raw-data interface. Here, we demonstrate the impact of motion on clinically relevant quantitative measures. Together, these developments reinforce STIR’s position as a mature, community-driven platform for modern PET reconstruction research.

Abstract

The technology of amorphous silicon-based micro-channel plates (AMCPs) uses a versatile approach to stack hydrogenated amorphous silicon in thicknesses up to 100 μm, using plasma-enhanced chemical vapor deposition  (PECVD) and etch microchannels of diameter below 2 μm every 5 μm, on a hexagonal pattern, by deep reactive ion etching. With gains up to 1000 for reverse bias less than 300 V, the design is possible to build on arrays of sensitive transimpedance amplifiers, integrated in read-out circuits with high channel density, such as for instance 15 by 15 μm. This ensures single photon detection with quantum efficiency up to 30% and noise rates below 50 Hz/cm², if combined with state of the art photocathodes such as the hi-QE from Photonis. Current fabrication runs focus on direct integration of the stack on ASICs, to be integrated with the photocathode over the next months. Measurements show timing response with jitter below 10 ps, making this high-precision photodetector ideal for radiation detectors combining ultra-fast photon emission and low light yield, such as Cherenkov.

Abstract

This paper presents a universal calibration method based on spatial vector decomposition to address the challenge of dynamically changing and difficult-to-calibrate geometric parameters in cone-beam computed tomography (CT) systems with complex scanning trajectories. The proposed method enables high-precision and robust calibration for a broad range of CT imaging systems. First, the projection matrix at each scanning angle is computed using a calibration phantom. This matrix is then decomposed into spatial vectors representing the system's geometric state, including the position of the X-ray source, the detector reference point, and orientation vectors. These vectors can be directly used for image reconstruction or converted into device-specific parameters. A vector-based detector mapping algorithm is also introduced to facilitate efficient projection data correction. Experiments on Micro-CT, CL and robotic-arm CT systems confirm that the method significantly improves image clarity, suppresses artifacts and distortions, and accurately reconstructs the actual motion trajectories of the X-ray source and detector. The method demonstrates high accuracy and versatility, providing a reliable calibration solution for high-degree-of-freedom CT systems.

Abstract

CT‑less attenuation correction in PET has gained increasing interest as a means to reduce patient radiation exposure, with recent AI‑based methods providing anatomically plausible attenuation estimates from emission data and complementary inputs. However, residual inaccuracies in such attenuation maps can still impact PET quantification.

Joint estimation of activity and attenuation provides a physics‑based framework to mitigate these effects, but remains less well constrained in the absence of time‑of‑flight (TOF) information.

This work evaluates the use of AI‑generated CT‑less attenuation maps as spatial priors for joint estimation of activity and attenuation in TOF and non‑TOF PET. Using a maximum‑a‑posteriori formulation, the CT‑less attenuation map is employed both as an initial estimate and as a regularising prior.

Experiments on paediatric PET datasets show that this approach stabilises joint reconstruction and recovers anatomically plausible attenuation maps, including in the non‑TOF case. These results support the use of CT‑less attenuation maps as effective priors for robust PET reconstruction.

Abstract

Dual-panel positron emission tomography (PET) scanners offer compact, cost-effective, and organ-dedicated imaging geometries, but their limited angular coverage introduces anisotropic resolution loss and elongation artifacts. This work presents a convergence-informed artificial intelligence (AI)-guided iterative reconstruction framework for the ProVision dual-panel PET system. The proposed approach exploits voxel-wise convergence information (VCI) extracted from the first 15 iterations of list-mode multi-ray maximum likelihood expectation maximization (MLEM) reconstruction. VCI was incorporated into five reconstruction strategies: analytic VCI-guided MLEM, multilayer perceptron-based VCI-guided MLEM, standalone 3D Swin-UNet reconstruction, AI-guided maximum a posteriori reconstruction (AI-MAP), and AI-initialized MLEM (AI-MLEM). Training and evaluation were performed using Monte Carlo simulations, a scaled NEMA image quality phantom, an external simulated patient test set, and an experimental anthropomorphic pelvis phantom. Compared with conventional MLEM, the proposed hybrid methods substantially improved contrast recovery, background uniformity, and geometric fidelity. AI-MLEM increased the contrast recovery coefficient from 53.16% to 82.35%, reduced background variability from 12.38% to 8.25%, and reduced the elongation ratio from 1.48 to 1.15. On the external simulated patient test set, AI-MLEM improved structural similarity from 0.446 to 0.816 and reduced normalized mean absolute error from 0.405 to 0.215. Experimental phantom results further confirmed improved lesion conspicuity and reduced vertical spreading. These results demonstrate that early-iteration convergence behavior provides useful voxel-wise guidance for mitigating limited-angle artifacts while preserving data consistency in dual-panel PET reconstruction.

Abstract

We propose a continuous Gaussian representation for PET reconstruction, modeling the activity distribution with normalized Gaussian basis functions. This reformulates PET reconstruction as a compact differentiable optimization problem by coupling an analytical Gaussian forward model with Poisson likelihood. The key component is an anatomically guided adaptive density control strategy, which dynamically prunes, splits, and clones Gaussian primitives according to contribution, gradient, and MRI-derived mask constraints, thereby allocating representation capacity to informative regions while suppressing redundant components. Image-domain Huber-TV regularization is incorporated to stabilize reconstruction. Experiments on 2D simulations demonstrate competitive performance across count levels and clear advantages under low-count conditions, suggesting the potential of adaptive continuous representations for PET reconstruction.

Abstract

The latest generation of simultaneous positron emission tomography and magnetic resonance imaging (PET/MRI) scanners is now at the forefront of hybrid imaging. As of now, the reconstruction of both modalities is independent, but research is ongoing in MRI-guided (MRIg) PET reconstruction, where information from the MRI is used as prior information in the iterative PET reconstruction.
We present three different clinical cohorts: meningioma patients (n=8) scanned with [⁶⁸Ga]DOTATOC, a single pituitary adenoma scanned dynamically with [¹⁸F]FET and a neurodegenerative cohort (n=10) scanned with [¹⁸F]FDG. The cohorts have been analyzed with standard OSEM reconstruction (b=0) versus MRIg reconstruction. The meningioma cohort showed an increased volume with standardized uptake values (SUV) above 2.5 when using MRIg reconstruction, and an increased SUV_max in the tumor. The pituitary adenoma showed increase time-activity curve both for β=1 and β=2. The neurodegenerative cohort showed stable SUV ratios of the cerebral-to-cerebellar cortex down to 3 minutes.

This preliminary work shows the clinical value in utilizing the information from the MRI in PET reconstruction with applications both to increase image quality and resolution, but also to facilitate faster imaging.

Abstract

This study presents the development of a hybrid Neural Network designed to simultaneously reconstruct medical images from their sinograms and to predict diagnostic parameters for neurodegenerative disorders, specifically Parkinson’s disease. The training pipeline utilizes simulated Dopamine Transporter Scan (DaTscan) brain images and their corresponding sinograms, generated via a custom Monte Carlo based framework that employs analytically driven functions to model realistic hotspot and background noise characteristics. The architecture integrates image reconstruction and disease-specific parameter estimation into a unified framework. Key diagnostic features include the bilateral asymmetry of radiotracer uptake within the caudate nuclei; deviations in spatial distribution or intensity within predefined Regions of Interest (RoIs) serve as primary indicators of neuropathology. This dual-function approach advances automated, AI-assisted functional neuroimaging, offering a robust pathway for early, non-invasive diagnosis of neurodegenerative conditions.

Abstract

Limited-angle computed tomography (LACT) reconstruction is highly ill-posed because missing projection views lead to severe artifacts and incomplete boundaries. This work proposes a global-local latent diffusion framework for limited-angle CT reconstruction. The proposed method performs diffusion reconstruction in the latent space of a variational autoencoder and jointly incorporates U-Net-based local modeling and transformer-based global modeling into the denoising process. The U-Net component captures multiscale spatial details and boundary continuity, whereas the transformer component models long-range dependencies and global structural relationships associated with missing-angle information. These complementary priors are adaptively integrated to guide latent-space denoising. During sampling, a physics-guided data-consistency constraint is further introduced. Specifically, the intermediate latent prediction produced by the model is decoded back to the image domain, processed with a projection-consistency operation, and then re-encoded into the latent diffusion trajectory. Preliminary experiments on a simulated data set show favorable reconstruction quality, indicating that the proposed global-local diffusion framework can improve structural fidelity and projection consistency for limited-angle CT reconstruction.

Abstract

Proton CT is a promising alternative to X-ray CT for proton therapy treatment planning, allowing direct estimation of the relative stopping power map within the patient body without relying on conversion from Hounsfield units. Conventional list-mode proton CT scanners measure the energy loss of each individual proton to estimate the integral of the relative stopping power along its path, the water-equivalent path length, but do not meet the requirements for clinical use due to their low acquisition rates. An alternative proton CT scanner design was recently proposed where the time-of-flight of each proton is measured between two detectors located before and after the patient along the proton beam. The main advantages of this sandwich time-of-flight design are its compactness and high acquisition rates. However, conversion from sandwich time-of-flight data to water-equivalent path length is not possible. In this work, we propose an iterative reconstruction of the sought relative stopping power map based on the non-linear forward problem. The method is assessed using Monte Carlo simulations of the acquisition of four simple phantoms. Although computationally expensive, this method accurately recovers the relative stopping power map.

Abstract

A high-sensitivity, high-resolution brain PET system with a multi-resolution detector design has been developed. However, image reconstruction without normalization correction produces pronounced image artifacts. In this study, we implemented list-mode reconstruction with direct normalization correction using normalization factors derived from simulation data for this multi-resolution PET system. The effect of normalization correction on reconstructed image quality was evaluated using Monte Carlo simulations and real scan data from a modified image-quality phantom. The results showed that normalization-corrected reconstruction substantially reduced image artifacts and improved image uniformity in simulated data, while residual nonuniformity remained in real scan data due to the lack of experimentally measured crystal efficiencies. Overall, this work demonstrates the feasibility of normalization correction for the dual-layer brain PET system and provides a basis for further experimental normalization using measured crystal efficiencies

Abstract

In low-dose (LD) myocardial perfusion imaging (MPI) with single-photon emission computed tomography (SPECT), radiation exposure is reduced at the expense of image quality, potentially impairing perfusion defect detection. Deep learning methods have shown promise in reducing noise in LD MPI images. However, clinical data exhibit varying cardiac orientations, increasing the complexity of the learning task for  denoising models. We hypothesize that standardizing cardiac orientation with respect to a common reference prior to training can reduce this complexity and improve denoising performance. We evaluated this approach with a previously proposed edge-aware convolutional autoencoder (EdgeCAE) across four levels of trade-off between edge preservation and mean-squared error (MSE) fidelity. The model was trained on 738 clinical subjects and tested on 300 noise realizations derived from two independent patient anatomies, with and without pre-reorientation. Our results demonstrated that cardiac reorientation consistently reduced the MSE compared to the full-dose reference images. This was also supported by visual assessment on an independent patient cohort.

Abstract

Multiplexing in SPECT renders the measurement operator underdetermined. In symmetric multi-source configurations, iterative ambiguity reinforcement sustains persistent ghost artifacts. Established mitigation strategies require supplemental single-beam nonmultiplexed projections, constraining collimator design and sacrificing sensitivity. Sparsity in gradients for common tracer distributions motivate total-variation (TV) regularization as a natural demultiplexing strategy: ghost artifacts produce extra gradient support that TV prior penalizes, breaking the ambiguity cycle. We demonstrate that MAP-TV (maximum a posteriori total variation) acts as an active reconstruction-stage demultiplexer, a supplemental approach to modifying the projection set in the data acquisition stage.

Abstract

In this paper, we propose a new projection method for PET, called Monte Carlo Point Projectors (MCPP), which back-projects measured gamma interactions in the detector by sampling potential positions of the event source within the field of view (FoV). We apply this method to scattered events using a projector for scattered lines of response (ScLoR), as well as a combination of scattered lines of response and Compton cones of response (CoR). The simulated system is a cylinder with a radius of 15~cm and a length of 16~cm, composed of 12 Cadmium Zinc Telluride panels, with a 7~cm radius water sphere at the center to simulate in-patient scatter. We show that for this system, the point spread function (PSF) resolutions were 8.3~cm for CoR and 5.3--4.7~cm for the proposed ScLoR projector and the combination with the CoR, demonstrating the advantage of using scattered events with the MCPP method. We also studied how the missing energy of scattered events influences the PSFs obtained with MCPP, noting that lower missing energy yields better resolution. Its combination with a scatter classification method on experimental data is being evaluated. In conclusion, the MCPP approach facilitates the integration of different projectors to process a maximum of detected events.

Abstract

Dynamic PET imaging is becoming increasingly valuable in neuroimaging and cardiac imaging, thanks to its ability to track functional changes over time.
Conventional approaches to dynamic PET imaging need to find a balance between frame duration and noise: shorter temporal frames can provide important information on rapid physiological processes, but at the expense of increased noise.  Traditional filtering strategies aim at reducing noise, but inherently cause loss of spatial and/or temporal resolution.
In this work, we represent dynamic PET images using a subspace model during image reconstruction to substantially reduce noise while preserving spatial and temporal information, thereby exploiting the spatiotemporal correlations in dynamic images.  We demonstrate the power of the proposed approach on an in-vivo scan of a nonhuman primate performed on the high-sensitivity, high-resolution NeuroEXPLORER scanner.  Dynamic images reconstructed at 1-second frames exhibit high resolution and dramatically reduced noise compared to conventional reconstructions, enabling the visualization of early tracer perfusion, respiratory motion and contrast changes in the whole body, without any gating.  The proposed method shows promise for short-frame imaging of the brain and body.

Abstract

Dynamic ^99mTc-diethylenetriaminepentaacetic acid (DTPA) planar renography is widely used for quantitative renal function assessment, but Gates-based glomerular filtration rate (GFR) estimation depends on manually defined kidney and background regions of interest (ROIs), which is time-consuming and operator-dependent. We developed RFVGG, an ROI-free hierarchical deep learning framework that learns directly from dynamic renography image sequences using routinely available Gates-derived labels. This retrospective study included 728 consecutive patients who underwent dynamic ^99mTc-DTPA planar renography. RFVGG was built on a shared 3D VGG-style backbone with hierarchical task-specific heads for three-, five-, and six-category GFR stratification. Model development followed five-fold cross-validation, and RFVGG was compared with CNN, ResNet-18, and DenseNet under identical evaluation settings. Classification performance was evaluated using accuracy, precision, recall, and specificity. RFVGG achieved accuracies of 0.842±0.020, 0.705±0.043, and 0.606±0.042 for the three-, five-, and six-category tasks, respectively. The corresponding precisions were 0.846±0.017, 0.668±0.037, and 0.614±0.034; recalls were 0.843±0.024, 0.689±0.045, and 0.642±0.039; and specificities were 0.915±0.012, 0.920±0.012, and 0.920±0.008, respectively. These results indicate that RFVGG can perform automated ROI-free GFR category stratification from dynamic ^99mTc-DTPA planar renography without contour-based kidney or background ROI tracing

Abstract

Quantitative brain PET studies often require arterial input functions (AIF), which can be derived from radiotracer concentrations through the internal carotid arteries (ICAs) in dynamic PET images. Image-derived input function (IDIF) methods avoid invasiveness; however, challenges in precise quantification of dynamic flow conditions and correction of partial volume effects (PVE) persist. Recently developed dedicated PET systems for brain imaging, such as the RETIMAGER System, offer excellent spatial and temporal resolution, which is of high relevance for IDIF determination. In addition, hybrid systems offer complementary information for improved IDIF validation. This study shows the design and development of digital and physical phantoms for future validation of IDIF methods in Brain PET systems. A two-phase approach employs the design and construction of a cylindrical phantom with an ICA insert matching real structural geometries. Phase 1 uses a digitalized phantom with Monte Carlo (MC) simulations to validate and optimize acquisition and reconstruction parameters. Phase 2 is in progress for experimental versus simulation validations using a physical phantom with an ICA insert to simulate flow dynamics using an injector. Additionally, a 3D head phantom incorporating gray matter (GM), white matter (WM), scalp, and ICA structures has been designed and is in fabrication for future comprehensive validation.

Abstract

Spatial resolution assessment in positron emission tomography (PET) typically relies on the full-width at half-maximum (FWHM) of point-like line sources reconstructed with filtered back-projection (FBP), yet modern clinical imaging predominantly uses iterative reconstruction methods, motivating the need for phantoms that evaluate resolution under clinically relevant protocols. To address this, we present TRACE (Tapered Rod Array for Continuous-resolution Evaluation), a 3D-printed phantom enabling simultaneous spatial resolution assessment via four complementary methods – FBP FWHM, modulation transfer function (MTF), Rayleigh criterion, and visual assessment – in a single acquisition. TRACE consists of a tapered cone with continuously decreasing fillable conical rod diameters (4.8–1.2 mm over 10 cm) and three integrated capillary slots for NEMA-compatible line source measurements. The phantom was scanned on the uEXPLORER total-body PET scanner in transverse (XY) and axial (XZ) orientations. FBP FWHM yielded mean transverse resolutions of 3.0 ± 0.2 mm (radial) and 3.4 ± 0.2 mm (tangential), and 2.9 ± 0.5 mm axially. Under OSEM reconstruction, the Rayleigh criterion and visual assessment resolved rods down to 1.8 mm (XY) and 2.0 mm (XZ), while the 50% and 10% MTF were achieved at 2.8 mm and 1.6 mm (XY), and 2.6 mm and 1.7 mm (XZ), respectively. TRACE bridges standardized and clinically relevant resolution assessment, delivering all four metrics in a single acquisition and offering a comprehensive framework for PET spatial resolution characterization

Abstract

Conventional methods for assessing PET spatial resolution, such as the NEMA NU-2 protocol and ACR guidelines, provide limited insight into the effective resolution achieved with modern clinical reconstruction techniques. In this work, we propose a simple and low-cost fan-pattern paper phantom enabling modulation transfer function (MTF)-based analysis for continuous spatial resolution characterization. The phantom consists of laser-cut paper in a fan geometry with gap sizes decreasing from 5 mm to 1 mm which is briefly immersed in an F-18 solution.
Reproducibility and repeatability were evaluated on a uEXPLORER PET/CT system. Images were reconstructed using filtered back projection (FBP) and multiple clinical reconstruction algorithms, including OSEM and artificial intelligence (AI)-based methods. Spatial resolution full width at half maximum (FWHM) was derived from the MTF using the 50% frequency criterion. Results demonstrate an average radial resolution of 3.23 ± 0.034 mm for reproducibility and 3.22 ± 0.054 mm for repeatability, consistent with NEMA NU-2 measurements but with lower uncertainties, which can be attributed to the robustness of the MTF evaluation method. The proposed approach enables rapid (<5 min), robust, and continuous assessment of spatial resolution, including under clinical and AI-based reconstruction conditions. The proposed phantom provides a practical tool for PET system characterization and reconstruction optimization and will be made available to facilitate wider adoption.

Abstract

Abstract—Lutetium-177, a combined gamma and beta emitter,
is gaining significant attention as a theranostic isotope with clinical
applications in metastatic prostate cancer and neuroendocrine
tumors. Quantitative imaging of ¹⁷⁷Lu in a preclinical setting
is important for therapy optimization. This study evaluates the
image quality performance of a small-animal SPECT scanner
equipped with an ultra-high-sensitivity collimator for ¹⁷⁷Lu using
a SPECT IQ phantom adapted from the NEMA NU 4 PET
phantom.
Whole-body acquisitions were performed using a multi-bed
protocol, and projection data were reconstructed with an iterative
algorithm under different post-filtering conditions. Image quality
was assessed using recovery coefficients (RCs) for rods of varying
diameters, and image noise was quantified as the percentage
standard deviation (%STDunif ) within a uniform region.
The results show that RC varies with rod size, iteration
number, and post-reconstruction filtering. Effective recovery
was observed for 1.35 mm rods (RC ∼0.77), while RC values
decreased to approximately 0.38 and 0.23 for 1.15 mm and
0.95 mm rods, respectively, indicating that reliable recovery is
achieved for rod sizes of about 1.5 mm. Image noise increases
with iteration number and decreases with stronger smoothing,
reflecting the expected resolution–noise trade-off.
These initialfindings demonstrate the capability of the SPECT
IQ phantom for quantitative performance evaluation and protocol
optimization in high-resolution SPECT imaging.

Abstract

Energy-based scatter estimation in PET estimates scatter from the energy information of detected photon pairs, without relying on an attenuation map. Recent work demonstrated promising results using a model that decomposes the 2D energy histogram into true and scattered coincidences via three basis functions. This model rests on two fundamental assumptions. First, that the shape of the scattered energy spectrum is sufficiently stable across varying activity distributions and time-of-flight (TOF) bins to be represented by fixed weights on two basis functions (c₁ p₁ + c₂ p₂). Second, that the scattered and true energy distributions remain statistically distinguishable given the finite energy resolution of the scanner: when low-angle Compton scatter dominates (as in LORs through high-contrast hot regions at central TOF bins) the scatter spectrum shifts towards 511 keV and the two distributions may become so spectrally similar that reliable decomposition is mathematically infeasible, regardless of counting statistics. In this work, we investigate the validity of both assumptions using long MCGPU-PET Monte Carlo simulations of an elliptical phantom with hot-sphere inserts of radii 1, 3, and 5 cm and contrast ratios of 1:10 and 1:100. Scattered energy spectra are extracted and inspected across TOF bins, hot-region sizes, and contrast ratios; challenging cases are identified for each assumption; and the feasibility of the nine-component 2D energy histogram decomposition from S.A. Zaman Pour et al. is evaluated.

Abstract

The accurate quantification of amyloid-beta (Aβ) plaques with  Positron Emission Tomography (PET) is critical for the early diagnosis and management of Alzheimer's Disease (AD). However, the diagnostic accuracy of PET imaging is heavily dependent on the image reconstruction algorithm employed. This study provides a quantitative comparison of three distinct PET image reconstruction methods to determine their efficacy for Aβ plaque identification. Using a clinical AD dataset, each of the three algorithms was applied to generate enhanced PET images. The resulting Aβ burden was quantified using the standardized Centiloid scale(CL), and a rigorous statistical analysis was performed to evaluate the performance differences between the methods. Identifying the optimal reconstruction technique is paramount, as its implementation can lead to more sensitive early detection of AD, improved monitoring of disease progression, and ultimately, the enhancement of therapeutic outcomes.

Abstract

Generating accurate attenuation maps from CT images is essential for quantitative PET/CT imaging, while minimising radiation exposure is also important. In conventional CT, the attenuation contributions from Compton effect and photoelectric absorption are combined into a single attenuation value and transformed together from the CT energy spectrum to the PET imaging energy, despite their different energy dependencies. In contrast, spectral CT enables separate electron density and effective atomic number images to be reconstructed. This allows to transform the relevant Compton attenuation contribution independently by applying a scaling factor to the electron density image. In this study, we compare attenuation maps obtained using this approach with those obtained via a bilinear transformation. Measurements of phantoms containing various tissue-equivalent inserts are performed using a PET/CT with twin-beam dual-energy capability and a photon-counting CT. In additional measurements, selected inserts are replaced by compartments containing activity in order to evaluate the effect of the different attenuation maps on reconstructed PET images. The measurements are performed across a wide dose range to investigate how far the radiation dose can be reduced while still obtaining accurate attenuation maps. Preliminary results show good agreement between attenuation values obtained at different dose levels, with the highest discrepancies between both transformation methods observed for bone-equivalent inserts. Using electron density and effective atomic number reconstructions from spectral scanners to generate attenuation correction values may improve quantitative PET imaging.

Abstract

Single scatter simulation (SSS) combined with tail-
fitting is the standard analytical method for scatter correction
in PET. Extending SSS to time-of-flight (TOF) data requires
non-trivial adaptations to the scatter kernel. The open-source
Software for Tomographic Image Reconstruction (STIR) library
has a widely used implementation of SSS for non-TOF data.
Here, we present a fully TOF-aware SSS implementation in STIR
with two contributions: (i) the TOF kernel ε_t in the scatter-arm
emission integral is computed via erf (error function) as opposed
to central point approximation in Watson, avoiding the need for
an extra normalisation step. (ii) re-use of cached integrals from
the non-TOF implementation. The implementation is validated
against MCGPU Monte Carlo on a 9-ring compact PET scanner
(27 TOF bins, 169 ps, uniform phantom), confirming strong
agreement with MC single scatter (r=0.985 full 4-D sinogram;
r=0.994 TOF-summed spatial). This contribution opens the door
to fully quantitative PET reconstructions for TOF-PET in STIR.

Abstract

Energy-based scatter (EBS) correction is increasingly utilized in PET image reconstruction. This work introduces a novel EBS framework that assigns scatter probabilities to individual listmode events via analytical Bayesian calculations. The method utilizes pre-defined energy distributions of scatter kernels as basis functions. By analyzing the global energy spectrum, the algorithm estimates the fractional probability for each event type and subsequently calculates the likelihood of Compton scattering for individual coincidences. The approach was validated using two phantom datasets acquired on a Biograph Vision PET/CT scanner (Siemens Healthineers). Reconstructed images were compared against standard vendor-implemented single-scatter simulation (SSS) with tail-fitting. Results demonstrate high consistency, with scatter fraction estimates differing by less than 3% from SSS. Furthermore, uniform and NEMA image quality phantom analyses showed less than 1% divergence in contrast recovery and background variability. These findings suggest that the proposed EBS method provides a robust, computationally efficient alternative for scatter estimation in clinical PET reconstruction.

Abstract

Accurate attenuation correction (AC) using computed tomography (CT) is critical for quantitative PET imaging. However, in respiratory-gated cardiac PET, respiratory phase mismatch between gated PET and CT can compromise the accuracy of CT-based AC, while acquiring additional phase-matched CT scans is impractical due to radiation dose concerns. To address these limitations, we employed a 3D Denoising Diffusion Probabilistic Model (DDPM) to generate synthetic CT images directly from non-attenuation corrected (NAC) PET data for gated PET AC. The model was trained using paired non-gated Rb-82 myocardial perfusion PET scans and CT-derived attenuation maps, and directly applied to respiratory-gated PET without retraining. Visual evaluation showed that 3D DDPM generated synthetic CT images with superior anatomical fidelity and improved respiratory phase alignment with gated PET compared with 3D GAN, 3D UNet, and 2D DDPM. Regional quantitative analysis near the diaphragm further demonstrated that 3D DDPM achieved improved attenuation correction in regions most affected by respiratory mismatch.

Abstract

Conventional synthetic MRI generates contrast images from a single effective set of voxel-wise quantitative parameters: proton density (PD), T1, and T2, which may be insufficient when any two or all three of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) coexist within the same image voxel. This work presents an exploratory framework using the standard spin-echo signal generation model with a purity-weighted Expectation-Maximization algorithm for tissue prototype estimation that systematically quantifies the prevalence of such multi-tissue voxels. From three co-registered spin-echo MR brain images at distinct Time to Echo (TE), Time to Repetition (TR) settings, per-voxel PD, T1, and T2 relaxation maps are recovered via a robust nonlinear inversion pipeline and validated by re-synthesizing the original images. Each voxel is then decomposed into fractional contributions of GM, WM, and CSF using three methodologically independent models: a Signal-Domain Mixture Model, Parameter-Domain Mixture Model and Mean Parameter-Domain Mixture Model. All three consistently show that 52% – 62% of brain image voxels contain contributions from multiple tissue types, demonstrating model-independently that the single-effective-parameter representation is insufficient in most brain voxels and that conventional synMRI parameter maps can be systematically biased. A tissue-aware synthesis framework is proposed that generates physically grounded MRI contrasts across the full spin-echo (TE, TR) measurement space, providing a quantitative basis for more reliable synthetic MRI. Future work will extend this framework to multi-subject validation and evaluate the synthesized contrast library for machine learning-based lesion diagnosis.

Abstract

In prior work, deep-learning (DL) based denoising was demonstrated to be capable of yielding images comparable to standard-dose studies when the imaging dose is reduced several-fold in pediatric Tc-99m DMSA renal SPECT, based on both quantitative image quality metrics and visual assessment by clinical experts. In this study, we further evaluate the diagnostic utility of DL-denoised images by employing a DL-based classifier to perform a task-based assessment of cortical defects, where a 3D convolutional neural network (CNN) was used to differentiate kidneys with cortical defects from those with normal uptake. We demonstrated this approach with a set of clinical data acquired from 956 subjects, with performance quantified using receiver-operating characteristic (ROC) analysis. The results showed that, without DL denoising, the quarter-dose data achieved a lower classification performance than the full-dose in terms of the area under the ROC curve (p-value < 0.001); however, with DL denoising, the quarter-dose data achieved a statistically comparable performance to that of the full-dose data (p-value = 0.452).

Abstract

While deep learning has become the mainstream approach for coronary artery extraction from coronary computed tomography angiography (CCTA) images, its reliance on training data makes it difficult to accurately extract fine distal vessels using existing datasets with insufficient annotations. Therefore, this study proposes a hybrid extraction method utilizing image feature analysis. This method consists of three stages. First, an initial extraction is performed based on anatomical features and CT values. Second, a chained connection algorithm combining tubular structure enhancement via the Hessian matrix and three-dimensional (3D) skeletonization is employed to connect vessel fragments through local search, thereby recovering continuity. Third, volume expansion is performed using an adaptive threshold corresponding to local intensity to restore the accurate vessel diameter. Application of this method to an open dataset demonstrated extreme robustness against CT value degradation caused by insufficient contrast agent perfusion. We successfully extracted and reconstructed fine distal vessels that were missing in the existing labels while maintaining their continuity. Future work involves collaborating with clinical specialists to reconstruct high-quality ground truth labels and conduct rigorous quantitative evaluations.

Abstract

Optical coherence tomography (OCT) provides non-invasive retinal biomarkers that may reflect the distinct neurodegenerative patterns of underlying proteinopathies. This study evaluates a machine learning framework to classify synucleinopathies (Parkinson’s disease [PD] and multiple system atrophy [MSA]) and tauopathies (progressive supranuclear palsy [PSP]) using retinal layer morphometry. Several classification tasks were assessed, ranging from healthy control (HC) discrimination to the differential classification of proteinopathic groups. XGBoost was employed as the primary classifier and evaluated using nested cross-validation. To minimize intra-subject variability, only the most recent visit per patient was included, with features aggregated as the average of both eyes to represent bilateral neurodegenerative involvement. The model demonstrated notable efficacy in distinguishing synucleinopathies (PD+MSA) from tauopathies (PSP), achieving an accuracy of 0.704 ± 0.057. This performance was comparable to the binary classification of HCs vs. all pathological cases (0.724 ± 0.079). In contrast, multi-class settings proved more challenging. These findings suggest that while fine-grained clinical subtype differentiation remains a challenge, OCT-derived biomarkers are highly effective at capturing the broad divergent impacts of alpha-synuclein and tau pathology on the retina.

Abstract

Theranostic radionuclide molecular imaging represents a core technology in precision nuclear medicine. However, its clinical application is hindered by scatter interference and complex background signals in spectral analysis. To address this bottleneck, this paper proposes a multi-step nonlinear iterative method for energy spectrum background subtraction and full-energy peak analysis, suitable for emission computed tomography detectors. The method realizes high-precision background subtraction and peak stripping via a systematic workflow consisting of peak localization, Log-Log-Square root transformation, interval iterative calculation, and Gaussian fitting. Experimental validation is conducted under both single-nuclide and multi-nuclide conditions. The results demonstrate that the proposed method can effectively separate full-energy peaks submerged by scatter in single-nuclide spectra. For the CsI detector, the energy resolution is improved from 12.9% to10.97%. For multi-nuclide and multi-peak spectra, it successfully separates overlapping peaks, suppresses Compton continuum interference, and optimizes energy resolution to around 10.6%. This method improves detector energy resolution, provides solid technical support for the clinical translation of theranostic radionuclide molecular imaging, and exhibits favorable adaptability and correction accuracy. Future work will focus on optimizing iterative efficiency and extending the method’s applicability to more complex clinical scenarios.

Abstract

Residual X-ray scatter degrades image contrast, uniformity, and quantitative accuracy in cone-beam computed tomography (CBCT). A two-dimensional (2D) anti-scatter grid physically suppresses scattered photons, but conventional gain correction using a matched flood projection only compensates the multiplicative grid-wall footprint and does not remove additive residual scatter, which persists as low-frequency nonuniformity in reconstructed images. This work proposes a projection-domain residual scatter correction method that exploits grid-wall shadows as internal sampling positions for scatter estimation. After gain correction, primary signals equalize between wall-shadow and neighboring hole pixels by design; any remaining signal mismatch is therefore attributed to additive residual scatter. Sparse scatter estimates at wall-shadow locations are expanded into a full-field scatter map via Gaussian-weighted interpolation and subtracted from the raw projection before logarithmic conversion and filtered backprojection reconstruction. This approach requires no additional hardware or dedicated scatter measurements, as it utilizes the intrinsic grid-wall structure already present in every grid projection. We evaluate the method using an ACR 464 phantom scanned at 80 kVp and 50 mA with 720 views on a 1516 × 1516 flat-panel detector (0.28 mm pixel pitch). Compared with no-grid and grid-only reconstruction, the proposed correction reduces residual shading artifacts and improves image uniformity. The contrast-to-noise ratio increases from 1.63 (no grid) and 3.27 (grid only) to 3.69 with the proposed method, with improved contrast visibility. These results demonstrate that grid-wall shadows provide practical sampling positions for estimating and correcting additive residual scatter in grid-based CBCT systems.

Abstract

The differential diagnosis of neurological disorders such as Alzheimer’s disease (AD), frontotemporal dementia (FTD), Parkinson’s disease (PD), and multiple sclerosis (MS) remains a clinical challenge due to overlapping neuroanatomical atrophy patterns in structural Magnetic Resonance Imaging (MRI). This work investigates whether low-dimensional embeddings derived from T1-weighted MRI can reveal a consistent neuroanatomical organisation across these conditions. Using a dataset of 528 subjects, we compared two feature extraction strategies: (i) high-dimensional voxel-wise representations and (ii) regional embeddings based on the Automated Anatomical Labeling (AAL) atlas (116 regions). Dimensionality reduction was performed using Principal Component Analysis (PCA) followed by Uniform Manifold Approximation and Projection (UMAP) to project subjects into a two-dimensional manifold embedding. Our results demonstrate that the neuroanatomical manifold is remarkably scale-invariant as both voxel-wise and regional embeddings yielded consistent topological distributions. Centroid distance analysis in the embedding space revealed a progressive deviation from the HC baseline across disorders. Interestingly, while the model was built on the HC-AD axis, MS and FTD cohorts exhibited the largest structural displacements (d_MS=7.24, d_FTD=5.59) compared to AD (dAD=3.57) and PD (d_PD=4.27). These findings suggest that low-dimensional manifold representations derived from structural MRI are robust to spatial aggregation and may provide an interpretable framework for studying shared and disease-specific brain alterations across heterogeneous neurological disorders.

Abstract

Abstract: This paper presents the development of a high-performance Compton camera for targeted radionuclide therapy within the framework of the European project AIDER. The camera is composed of two detector planes, each made of four LaBr₃ or CeBr₃ monolithic crystals of 50 mm x 50 mm size coupled to Silicon Photomultiplier (SiPM) arrays. In order to determine the optimum pixel size, a hybrid analysis approach based on convolutional neural networks (CNN) and dense networks (DNN) has been developed to estimate the interaction position of the photons. Starting from Monte Carlo simulation data (GATE v7.0) and using a reference based on an analytical model, the proposed alternative incorporates the systematic error of this model as an additional input for the NNs. This achieves active correction of the edge effect and improves the overall spatial resolution (FWHM) from a range of 1.80–2.25 mm down to 0.63–0.70 mm. Results show the benefits of the combined NN approach and indicate the possibility of employing SiPM arrays with large pixel sizes, effectively restricting the amount of readout channels. Experimental tests are ongoing.

Abstract

This study evaluates the suitability of spatial registration in deep learning for neuroimaging, focusing on the balance between reducing inter-subject anatomical variability and preserving data fidelity. We hypothesize that processing images in their native anatomical space avoids morphological distortions introduced by registration, enabling neural networks to detect subtle structural biomarkers that may be altered in normalized spaces. To test this, a standard densely connected neural network was used to classify two balanced cohorts: one focused on the diagnosis of Parkinson's disease (PD) and the other on the diagnosis of Alzheimer's disease (AD). The experimental design compared four input modalities: whole-brain volumes and gray matter maps, each in native space and registered to a standard reference space. For PD classification, native and registered spaces achieved similar accuracy, although registration produced a marked imbalance between sensitivity and specificity. For AD, the best performance was obtained using gray matter in native space, confirming that preserving topological integrity is critical in pathologies with pronounced atrophy. These findings suggest that spatial registration should not be considered an indispensable preprocessing step in deep learning-based neuroimaging, but rather a methodological choice that must be evaluated based on the task and dataset. Processing in native space is an effective alternative that simplifies preprocessing while preserving diagnostically relevant information, opening future research directions on model robustness to anatomical variability without prior transformation.

Abstract

Non-rigid motion correction for gated PET is challenging due to low contrast and high noise in homogeneous regions, where deformation estimates may drift, while strong smoothness regularization can oversmooth boundaries and suppress localized motion near lesions. We propose an edge-guided flow fusion method that combines two deformation fields predicted by identical registration architectures trained with different smoothness weights. For each moving--fixed PET pair, we compute an edge probability map from the gradient magnitude of the Gaussian-smoothed absolute intensity difference, followed by percentile normalization and logistic mapping. After thresholding and morphological dilation, the resulting mask defines an edge-dominant region. We fuse the two predicted flows by selecting the low-regularization flow near edges and the high-regularization flow in homogeneous background regions, then warp the moving PET volume using a spatial transformer. This approach preserves deformation flexibility at boundaries while maintaining smooth motion in low-contrast regions, without changing network architecture.

Abstract

PET detectors often use a Center of Gravity (CoG) algorithm applied to the output signal of sensor
modules to perform coordinate reconstruction. CoG assumes that the spread of optical photons emitted
by a pixelated scintillator crystal is symmetric about the point of interaction. A photoelectric
(PE) event, where energy is deposited within a single scintillator pixel, produces such a
distribution. Inter-crystal scatter (ICS) events, however, occur when a gamma photon deposits energy
into multiple scintillator pixels, producing multiple photon distributions that degrade CoG
accuracy and spatial resolution. Convolutional neural networks (CNNs) have proven effective
for identifying and recovering ICS events in simulated data and in real systems when trained on
experimentally acquired datasets, typically obtained using collimated measurements to isolate
scintillator pixels. However, acquiring collimated datasets is time-consuming and requires high
precision, motivating the use of simulated data as a primary training source. In this work, PET
data was generated using GATE and used to train a small CNN to identify and remove ICS events
based on softmax confidence values for a detector consisting of a 24×24 LYSO array coupled to an 8×8
SiPM array. A NU-4 image quality phantom and Na-22 point source acquired on a clinical system
demonstrated improvements in recovery coefficient (RC) and peak-to-valley ratio for all RC five rod sizes, along
with modest improvements in spatial resolution and spill-over ratio. The results demonstrate the feasibility of using simulated
data as a primary training source for ICS filtering and recovery. Thus, a hybrid approach incorporating
limited real data may further improve performance without requiring a large number of collimated acquisitions.

Abstract

Dynamic PET reconstruction is challenging due to strong temporal variations in count statistics, leading to non-uniform noise and spatial resolution across frames. In this work, we simply extend previous works to propose a temporally and spatially-variant regularization strategy for dynamic penalized maximum-likelihood (PML) reconstruction based on a time-and-space-dependent modulation factor derived from the Fisher information matrix. A modified estimation of the modulation factor is used to improve robustness in low-count conditions by removing dependence on noisy measurements. The method was evaluated using dynamic FDG simulations based on the 2D Zubal brain phantom and compared with OSEM and PML reconstruction without modulation.
Results showed improved preservation of anatomical structures and lower RMSE values in the cortex, putamen, and caudate regions while avoiding the excessive smoothing observed with conventional PML reconstruction. These results suggest that automatic temporally and spatially adaptive regularization can improve image quality and quantitative accuracy in dynamic PET imaging.

Abstract

This study applies the Statistical Agnostic Regression (SAR) methodology to validate predictive model for brain aging using gray matter functional connectivity derived from fMRI data from the UK Biobank database. The method evaluates significant linear relationships between age and functional connectivity correlation matrices, by resampling and estimating an upper bound on the expected risk through a Bayesian threshold under the worst-case scenario. The SAR power quantifies the probability of detecting a true linear relationship using the expected risk statistic, via Monte Carlo simulations under the null distribution. Predictive performance is assessed through Principal Component Analysis (PCA) applied to the global correlation matrices, which are constructed from subjects whose ages fall within the extreme percentiles (below the 15th and above the 85th percentile). This framework enables the identification of connectivity signatures associated with aging within a unified representation of functional brain connectivity dynamics. Functional connectivity matrices of gray matter were computed from preprocessed fMRI data by extracting region-wise time series and estimating pairwise statistical dependencies using Pearson correlation coefficients. The results support the development of a statistically significant model for predicting age from functional connectivity correlation matrices and demonstrate that combining rigorously validated predictive models with SAR and PCA methodologies can characterize age-related variability in functional connectivity and improve understanding of brain aging using large-scale population neuroimaging data.

Abstract

Radiomic features derived from [¹⁸F]FDG PET/CT imaging can quantify intratumoural spatial heterogeneity and have demonstrated prognostic utility across multiple oncological indications, including sarcoma. However, the biological interpretation of radiomic risk scores remains poorly understood, partly due to confounding in clinical datasets and the absence of ground-truth tumour parameters. A simulation-based approach offers a controlled framework for mapping image-derived features to biologically and physically meaningful tumour properties.

In this study, a parametric tumour model incorporating eight biologically motivated parameters—including uptake intensity, radial signal gradients, compartmental structure, multifocal uptake, and morphological deformation—was used to generate 1,024 synthetic tumour volumes across physiologically plausible parameter ranges. In parallel, a prognostic radiomic risk score was derived from [¹⁸F]FDG PET/CT images of 115 sarcoma patients enrolled in the Oxford Precision Oncology for Sarcoma (OxPOS) study. This trained model was subsequently applied to each simulated volume to infer a radiomic risk score, and correlation analysis identified simulation parameters most strongly associated with predicted risk.

Three parameters; baseline uptake intensity, radial signal gradient, and fractional high-uptake volume, exhibited moderate but statistically significant correlations with the model-derived risk score (|r|= 0.41, 0.38, and 0.52, respectively). These findings provide mechanistic insight consistent with known tumour biology; for example, elevated radial signal gradients are characteristic of tumours with necrotic cores.

This work demonstrates the utility of simulation for enhancing biological interpretability and communicability of radiomic prognostic models. Ongoing work aims to expand the simulation framework and integrate genomic data to further elucidate the biological underpinnings of radiomic signatures in sarcoma.

Abstract

Dynamic positron emission tomography (PET) requires dense arterial blood sampling to estimate arterial input functions (AIFs), creating substantial clinical burden and limiting practical implementation. To reduce number of samples, we developed a data-driven method for AIF estimation under sparse sampling conditions using nonlinear mixed-effects modeling. The study dataset included 90-min dynamic brain PET AIF data from 26 healthy subjects acquired with the 11C-DPA-713. A three-exponential nonlinear mixed-effects model was used to characterize the AIF, with fixed effects representing population-level kinetics and random effects capturing inter-individual variability. The AIF was modeled using a deterministic five-point subsampling strategy that selected representative early, intermediate, and late decay-phase observations. Model parameters were estimated using the stochastic approximation expectation maximization algorithm, and reduced-sampling predictions were compared with reference ground truth generated from the full dataset. The reduced five-point strategy produced AIF estimates that closely matched the full-data reference, yielding a root mean squared error of 0.37 and a mean relative error of ranging from 19% to 27.8%. Visual comparisons further demonstrated strong agreement between predicted and reference AIFs across subjects, indicating that plasma activity characteristics were preserved despite reduction in sampling density. These findings suggest that sparse sampling combined with nonlinear mixed-effects modeling can provide reliable AIF estimation while reducing the burden of arterial blood collection in dynamic PET studies.

Abstract

Thyroid eye disease (TED), the most common autoimmune orbital disorder and a major cause of orbital morbidity, is still assessed largely through subjective clinical measures. Although orbital MRI radiomics offers potential quantitative biomarkers for disease activity and treatment response, the reproducibility of these features under real-world imaging variability has not been systematically evaluated, limiting clinical translation. We performed a perturbation-based stability analysis using 150 patients from the public TOM500 dataset, extracting 293 radiomic features from seven Image Biomarker Standardization Initiative (IBSI)-compliant families across nine bilateral orbital structures using original and Laplacian-of-Gaussian filtered MRI images. We used ten perturbations per patient with combined rigid transformations, Gaussian noise, and 1-voxel mask modifications, and assessed stability using ICC(A,1). Across 2,637 structure-feature combinations, mean ICC was 0.72, and 53% of features achieved Good or Excellent stability (ICC ≥ 0.75). The medial rectus (ICC 0.84), followed by superior rectus and optic nerve (both 0.81) were the most robust structures, whereas the eyeball (0.55) and lateral rectus (0.61) were least stable. Laplacian-of-Gaussian filtering improved texture feature reproducibility, increasing median texture ICC from 0.73 to 0.80 and enhancing stability in medial rectus, lacrimal gland, and orbital fat features. Stability rankings were nearly identical between active and inactive disease groups (Pearson r = 0.92), indicating that disease severity did not influence feature reliability. These findings provide the first large-scale reproducibility assessment of orbital MRI radiomics in TED and establish a foundation for future biomarker and predictive modeling studies.

Abstract

The aim of the study was to investigate whether proton therapy fields can be verified in vivo during treatment using film-based transit portal dosimetry. Potential clinical treatment plans were checked for downstream dose contributions. The complete proton portal imaging workflow was verified in experiments with an anthropomorphic phantom.   Head-and-neck, mediastinal, thoracic, abdominal and cranio-spinal axis irradiations (CSIs) yielded downstream dose areas, which enable portal imaging. The cranial fields of CSI plans were deemed as most suitable and tested with the phantom in a simplified setup. Radiochromic films, mounted on top of the patient couch, served as portal imaging detectors. Phantom tests indicate that the patient setup can be verified within an accuracy of ≈1 mm per direction. In conclusion, a novel, straight-forward procedure has been developed for in-vivo dosimetry in proton therapy for some patients with specific indications.

Abstract

We developed a patient-specific volumetric dosimetry system using a tumor-shaped plastic scintillator derived from patient MR images for integral dose verification in Gamma Knife radiosurgery. Three tumor model scintillators (TMSs) representing a vestibular schwannoma (0.39 cm³) were irradiated using a nine-shot plan prescribing 8 Gy to the 60% isodose line. The integral dose, defined as the energy absorbed over the entire tumor volume, was measured and compared with the value calculated by the Leksell GammaPlan® treatment planning system. Two independent calibration methods were employed: calibration at a secondary standard dosimetry laboratory and Monte Carlo simulation-based calibration using Geant4. The integral dose measured by the simulation-based calibration was (2.00 0.76)% higher than the planned value of 5.01 mJ, falling within the acceptable 3% tolerance for Gamma Knife absolute dosimetry. The laboratory-based calibration yielded an integral dose (5.84 3.56)% lower than the reference value of 4.30 mJ; the larger deviation is attributed to probe reassembly uncertainty and the angular dependence of the calibration coefficient. These results demonstrate that the TMS system can serve as a practical tool for verifying the three-dimensional integral dose calculated by a treatment planning system.

Abstract

Compton cameras are being explored as a promising technique for treatment monitoring of targeted alpha therapy,  by imaging the gamma emissions produced during the decay of the alpha-emitting radionuclides. To improve image accuracy, attenuation correction is essential. Here we analyze the effects achieved by incorporating attenuation correction into the list-mode MLEM reconstruction of MACACO III+ synthetic data using ideal simulation. The results demonstrate that, without attenuation correction, contrast recovery coefficients of 97.2%, 84.7%, 73.5%, 47.8% are achieved for rods with diameters of 6, 5, 4 and 3 mm, respectively, using a total activity of 0.135 MBq of 440 keV gammas, a major gamma peak in the decay of ²²⁵Ac and a hot to background ratio 10:1, simulated over 2 hours. Combined energy reconstruction at 218 and 440 keV yields improved image contrast when attenuation correction is applied. This work was carried out within the frame of the Advanced Imaging DEtector for targeted Radionuclide therapy (AIDER) project.

Abstract

Positron emission tomography (PET) is an effective non-invasive modality for in vivo range verification in carbon-ion therapy. This study evaluates an all-digital PET system for in-beam monitoring of carbon-ion treatment plans. Clinical DICOM RT plans were converted into a GATE/Geant4 model of single-field pencil beam scanning (PBS) carbon-ion irradiation in a polymethylmethacrylate (PMMA) phantom. Simulated dose and positron-emitter distributions were compared with treatment planning system (TPS) dose and measured PET data. TPS and GATE showed good dosimetric agreement, with nearly identical planning target volume (PTV) D50% and a 3D gamma passing rate of 97.66% (2%/2 mm). PET validation yielded R50,sim = 124.33 mm and R50,exp = 125.00 mm, corresponding to ΔR50 = 0.67 mm. These results support millimeter-level PET-based range verification for carbon-ion therapy.

Abstract

In-vivo positron emission tomography (PET) imaging of positron emitters produced by the therapeutic proton beam in the patient is a well-studied irradiation verification tool. Adaptive treatment planning benefits from fast feedback, down to the sub-second time scale, requiring an in-beam PET scanner operating during beam delivery for N-12 imaging. Use in clinical routine requires that the performance of such detectors is not significantly affected by radiation exposure over time. We have studied the performance of four state-of-the-art time-of-flight Hamamatsu C13500-4075YC-12 (ToF)-PET detectors in a proton therapy radiation environment. A 150 MeV proton beam (pulsed with intensities of 0.3, 1 and 3 nA at 10% and 1.5% duty cycles) was stopped in PMMA phantoms whilst PET data was acquired. The coincidence energy spectra obtained in-beam (both with the beam on and between beam pulses) show the N-12 511 keV peak. As the experiment progressed, the data output shows cumulative faults from an increasing percentage of corrupt events. Whilst the accumulation of these faults was partially mitigated by system resets during the experiment, comparison of data acquisitions with a combined Na-22 & Ga-68 source before and after the experiment show that the performance degradation is long-lasting, although the detectors show signs of recovery 5 months later.

Abstract

Proton and heavy-ion therapy exploits the Bragg peak to concentrate dose within the tumor while sparing surrounding normal tissues; however, range uncertainties necessitate conservative safety margins and restrict clinical indications. To address this, we are developing an in-beam All-Digital positron emission tomography (PET) platform for dose-guided particle therapy. Two prototype systems have been designed, built, and tested at multiple proton and heavy-ion therapy centers: a 12-module pre-clinical partial-ring system for phantom and small-animal studies, and a 16-module research-clinical open-ring system for human-subject trials. Both employ lutetium-yttrium oxyorthosilicate scintillators read out via the multi-voltage threshold digitization architecture. In-beam PET imaging of proton-induced positron activity has been demonstrated in polymethyl methacrylate phantoms with approximately 1 mm range agreement, and in live rats where biological washout effects on image contrast were observed. A three-phase development roadmap is presented, progressing from online range monitoring to adaptive therapy and ultimately real-time feedforward therapy, aiming to narrow safety margins, broaden treatment indications, and provide the verification infrastructure for FLASH therapy.

Abstract

To overcome the limitations of existing beam range verification methods in terms of detection efficiency, system complexity, and real-time capability, this study proposes a carbon ion beam range verification method based on a pixelated cadmium zinc telluride (CZT) detector and machine learning. The pixelated CZT detector was used to measure the secondary particle response generated during carbon ion irradiation of a polymethyl methacrylate (PMMA) target, and a machine learning regression model was then applied to predict the beam range. The experimental results demonstrate that the pixelated CZT detector can accurately measure secondary particle responses. In addition, the effects of different detection regions, defined by pixel ranges, were investigated at a monitor unit (MU) setting of 6 × 10⁶. Using data from a 16 × 16 pixel region, the proposed method achieved a mean absolute error (MAE) of 0.79 mm. These results indicate that the proposed approach provides a feasible technical strategy for online range verification in carbon ion beam therapy

Abstract

Laser-driven sources of ionizing radiation provide unique possibilities for studying ground-breaking phenomena and for exploring novel applications of compact accelerators. An outstanding example is the controlled irradiation of biological samples with laser-accelerated, high-energy particles. After a pilot experiment at the Spanish Pulsed Laser Center (CLPU), the aim of the present study is an increase of the single-shot dose and the mean dose rate on monolayer cell cultures. Two possible scenarios at the same laboratory have been considered. On the one hand, the previous setup at the VEGA-3 laser line may be upgraded through a set of permanent quadrupole magnets to focalize the proton beam closely behind the laser target. On the other hand, a similar, more compact arrangement inside the VEGA-2 target vessel benefits from shorter flight paths and a 10-fold higher repetition rate. Particle tracking simulations with realistic beam control elements point towards a 5–15 times increase in proton fluence and single-shot dose.

Abstract

Proton therapy is an advanced radiotherapy modality whose clinical advantage arises from the highly localized dose deposition produced at the Bragg peak. However, this feature makes the technique highly sensitive to uncertainties in patient positioning, anatomical changes, and beam delivery. Verifying the dose delivered to patients after treatment session is therefore essential to ensure treatment safety and reducing planning margins. During proton irradiation, nuclear interactions generate positron‑emitting isotopes that can be detected with Positron Emission Tomography (PET). In this work, we focus on the detection of ¹⁸F, whose long half‑life enables offline PET acquisitions 60–150 min after treatment, when short‑lived isotopes have decayed. Because the expected in‑vivo ¹⁸F activity is very low (0.1–1 Bq/mL), we propose a PET system optimized for low‑count imaging, based on BGO scintillation crystals to avoid the intrinsic background of lutetium‑based detectors typically used in commercial scanners. 
The proposed PET prototype (INSPIRE) consists of 16 detector modules arranged in a single ring, each comprising a two‑layer pixelated BGO block (19×19 and 20×20 arrays) coupled to a 12×12 SiPM array, thus enabling Depth of Interaction (DOI) information. The system, which has an inner diameter of 26 cm and an axial length of 5 cm, is intended for brain studies. In this work we present the design of the INSPIRE system and the experimental characterization of the detector modules, including pixel resolvability, layer separation, and energy performance, together with preliminary proton‑irradiation simulations that illustrate the suitability of the ¹⁸F signal for offline PET imaging.

Abstract

Real-time monitoring in hadron therapy is essential to account for anatomical shifts and setup uncertainties that may compromise dose delivery. This work evaluates a software framework designed for online range verification via monitoring of the Prompt Gamma (PG) profile, which is know to be correlated to the Bragg peak position. The architecture employs a decoupled simulation pipeline: PG emission is modeled using MC codes, while the detector response is characterized through a dedicated Python-based tool, PGIPY. This simulated new profile, computed spot by spot, is then compared with measurement inside the software. To ensure spot-by-spot clinical correspondence, the software utilizes machine log files to synchronize simulated data with actual delivery sequences.

Experimental validation with an homogenoeus phantom was performed at the CNAO facility (Pavia, Italy). A high degree of correlation was observed between simulated and experimental profiles, confirming that the framework is free from significant systematic modeling errors. These results demonstrate that the proposed pipeline provides a robust, validated baseline for clinical integration, ensuring that any detected profile deviations during treatment are attributable to physical range shifts.

Abstract

Purpose: To enable personalization dosage of ¹⁷⁷Lu-based Theranostic radiopharmaceutical therapy (RPT), we are developing a Portable Organ Dosimetry Device (PODD) for at-home use to monitor activity in regions of interest (ROI) such as tumors and organ at risk (OAR). To accurately estimate OAR activity, knowledge of the relative positions of probes and ROIs is required. This can be achieved either by consistency of initially measured locations, or by verifying locations for each measurement. Methods: We have constructed a prototype PODD, and we examine the variation of ROI-location estimates of an NEMA IQ phantom in the PODD reference frame by measuring PODD and ROI features from repeat x-ray CT scan with the phantom removed and replaced in the PODD between measurements. Results: The standard deviations of ROI translational location with the phantom in a supine orientation was 1.0mm axially, 0.2mm horizontally, and 0.1mm vertically. Standard deviations of rotational position were 0.7 degree in tilt and less than 0.1 degrees in roll about the PODD and CT scanner axis. Conclusion: Locating uncertainty of a rigid body phantom within our prototype PODD will not significantly affect the accuracy of time-integrated activity estimates using longitudinal measurements with the phantom removed and replaced in the PODD between measurements. We must repeat these measurements for a realistic deformable phantom or with real subjects.

Abstract

Compton cameras are showing promise for targeted radionuclide therapy (TRT). The IRIS group of IFIC (Valencia, Spain) has developed a scintillator-based Compton camera prototype for medical imaging applications which is being assessed in TRT. The prototype MACACO III+ consists of two detector planes, with the second having an active area four times larger than the first. Previous tests performed on phantoms and small animals using ¹³¹I demonstrated the suitability of the system for radiotracer imaging. The system is being employed to obtain initial experimental data with different radionuclides to validate the simulations carried out within the European project AIDER, which will develop a high performance Compton camera. In this work, experimental measurements with MACACO III+ were carried out at La Fe Hospital (Valencia, Spain) to image a vial filled with ²²⁴Ra. GATE simulations were also performed to compare with experimental data. In addition, the prototype MACACO IV, composed of two planes with four detectors each, is under development. Simulation studies also include the configuration of the new prototype to assess influence of detector geometry on the overall system performance.

Abstract

Targeted radionulclide therapy (TRT) is a cancer treatment modality that employs radiopharmaceuticals to selectively deliver radiation to cancerous cells at specific target organs. Its use has increased drastically in the last years, and it has been extended to novel radiopharmaceuticals. The secondary photons emitted by the therapeutic radionuclides can be employed for imaging the compound distribution in the patient’s body for treatment assessment and dosimetry. A precise quantification of the radiation dose delivered to lesions and organs at risk is crucial for improving patient treatment safety and reducing related side effects. Conventional imaging modalities are not optimized for such secondary emission, in particular in the case of alpha-emitting radionuclides, which are gaining prominence due to their exceptional cytotoxicity. For targeted alpha therapy (TAT), image-based dosimetry is an unmet need as currently used gamma cameras are not effective in the case of these high energy and low activity emitters. The European project AIDER aims at developing a high-performance scintillator-based Compton camera to improve imaging in TRT and to enable image-based dosimetry in TAT. The project, coordinated by the IRIS group of IFIC, includes four academic partners, two hospitals, an SME and a patient association. Within the project, the system will be developed and tested with phantoms and patients. Initial work on system design optimization and image reconstruction, as well as detector and electronics development has been carried out.

Abstract

Targeted alpha therapy delivers highly localised dose, so treatment planning and delivery verification depend on whether the radionuclide source remains at the intended deposition site or is redistributed by diffusion or biological transport. This simulation study tests whether decay-chain gamma emissions can provide a detectable source-localisation signal in an AlphaGlue-like geometry. Photon transport was modelled for a thin loaded glue layer on a water phantom surrounded by six detector panels, with each photon entry into a panel counted as one hit. Seven source configurations were analysed: a retained-source baseline, broad redistribution of 10%, 50%, or 90% of the source in water, and localised displacement of 20%, 40%, or 60% of the source to a compact water region. Each configuration used ten repetitions of 10,000 primary decays and was normalised to a 3 μCi reference source.

Total detector-hit yield changed only modestly, but baseline-subtracted spatial hit profiles showed strong source-dependent structure. The central glue-facing signal fell to 0.94, 0.72, and 0.47 of baseline for the broad-redistribution cases, and to 0.84, 0.67, and 0.50 for the localised-displacement cases. Profile-shape features recovered the broad diffusion fraction with an R2 value of 0.998, while a secondary-source metric identified all localised displaced-source cases. The reconstructed displaced-source position was centred at approximately 525–531 mm, compared with the simulated 500 mm source region. Time-window analysis showed that large redistributions reached three-standard-deviation thresholds within 30 minutes to 2 hours; the weakest broad redistribution required about 6 hours, and the 20% localised displacement produced detectable displaced-region excess within about 1 hour. These results indicate that gamma-hit patterns provide source-location information beyond total count rate and could support activity verification and treatment-delivery monitoring.

Abstract

We present a low-cost lensless radiomicroscopy system for cellular-scale imaging of theranostic radiopharmaceutical uptake using a commercial CMOS image sensor. The device enables simultaneous brightfield and β-emission imaging with sub-10 μm spatial resolution, allowing direct correlation between morphology and radionuclide distribution. The system was validated using β⁻-emitting 111Ag and β⁺-emitting 18F-FDG sources, demonstrating quantitative linearity and accurate radioactive decay measurements. Live-cell experiments on MDA-MB-231 breast cancer cells showed detectable intracellular uptake of 111Ag, while imaging of U87 glioblastoma spheroids revealed spatially localized 18F-FDG accumulation in a warm background environment. These results demonstrate the potential of lensless radiomicroscopy as a compact and affordable platform for quantitative in vitro evaluation of theranostic agents at single-cell and multicellular levels.

Abstract

We introduce a novel quantum-inspired method for identifying intercrystal Compton scattering in pixelated scintillation detectors. The approach treats each SiPM readout as a spatial signal and converts it into a discrete eigenfunction representation using a Hamiltonian built from the measured intensity distribution. Event structure is then assessed through quantum localization, quantified by the participation ratio. Single-site events produce highly localized dominant modes, whereas Compton-scattered events produce multiple localized modes associated with distinct interaction sites. A dual criterion based on quantum localization and eigenvalue proximity is used to classify multi-site events without calibration-intensive response modeling or supervised training. The method was evaluated with Geant4 optical photon simulations and measurements from LYSO:Ce detector arrays with thicknesses from 2 to 25 mm. Experimental classification agreed closely with simulation, with differences below 1% for most thicknesses and below 5% in all cases. These results indicate that the proposed representation captures the topology of gamma-ray interactions directly from detector signals.

Abstract

Total-body PET imaging provides unprecedented sensitivity and extended axial coverage but remains affected by partial volume effects (PVE), particularly in small lesions and regions with spatially varying resolution. Conventional deconvolution approaches, such as Richardson–Lucy (RL), can improve spatial resolution but often suffer from noise amplification and inaccuracies in point spread function (PSF) modeling. In this work, we propose a learned unfolded deconvolution framework for robust PVE correction in total-body PET imaging. The method reformulates RL deconvolution as a fixed-depth network in which each iteration explicitly incorporates the system PSF while learnable parameters enable adaptive regularization and noise suppression. Standard-dose [¹⁸F]FDG PET images were used as network inputs to isolate the PVE correction problem from low-count noise effects. The model was trained in a self-supervised manner using RL-generated surrogate references, eliminating the need for voxel-wise ground-truth images. Evaluation was performed on multicenter total-body PET datasets acquired from uEXPLORER and Biograph Vision Quadra systems. Quantitative assessment on an independent test set demonstrated high structural fidelity (PSNR = 47.38 ± 5.71, SSIM = 0.98 ± 0.01) and low reconstruction error (RMSE = 0.005 ± 0.003), while maintaining acceptable SUV bias (12.2% ± 23.1%). Qualitative analysis showed improved lesion delineation and enhanced uptake contrast without visible noise amplification. Bootstrap-based analysis further demonstrated stable performance with narrow confidence intervals across metrics. The proposed framework provides an interpretable and robust solution for PVE correction, improving image quality and quantitative consistency in total-body PET imaging.

Abstract

Many studies have proposed methods to estimate internal dosimetry, enabling patient-specific adjustment of radiopharmaceutical therapy.  We are developing a practical, low-cost multi-pinhole collimated system (PODD) with on the order of 10 sensors at fixed locations around a subject.  Estimation of internal dosimetry to organs at risk is extremely ill-posed with this sparse sensor placement.  In this work, we hypothesize that use of objective functions with an accurate forward model and multiple regularization terms can directly estimate total accumulated activity using simulated PODD acquisitions of realistic distributions of Lu-177 DOTATATE.  Methods: We developed a custom GPU-based forward projector enabling rapid multi-pinhole detection simulation. We propose an objective function for direct estimation of wash-in, wash-out, and total integrated activity (TIA) kinetic parameters from multi-time point, sparse sensor measurements.  This objective function includes regularization based on a) matching sensor measurements to initial time-point SPECT acquisitions and b) promoting kinetic parameters similar to population kinetics. Evaluation was performed by simulating sensor acquisitions from public 4-time point SPECT/CT images (Dewaraja et al, 2021, Lu-177 DOTATATE Anonymized Patient Datasets). Conclusion: Direct estimation of clinically relevant TIA can be performed with on the order of 10 to 20 pinhole-collimated sensors acquiring data at four times post treatment.

Abstract

Highly sensitive long-axial-field-of-view (LAFOV) positron emission tomography (PET) scanners require fast and reliable reconstruction methods, where accurate quantification in short temporal frames is key to kinetic modeling applications. We extended FastPET - a direct neural network reconstruction method based on image loss - with an additional forwardprojection step and data loss penalty, modified the training database to specifically account for short temporal frames and evaluated its PET quantification accuracy in high- and lowcount regimes on the Biograph Vision Quadra. Compared to ordered subset expectation maximation (OSEM, 4i5s, all pass) reconstructions, FastPET based only on image loss resulted in smoothed images and an overshoot in the contrast recovery coefficient (CRC) by up to 20%. A combination of data loss and image loss counteracted this effect but still showed a degraded CRC for frame lengths ≤5s. Including low-count training data in the training stablized the image quality and slightly improved the CRC of small structures in short frames ≤5s.

Abstract

The extension of the axial field-of-view (AFOV) advances in positron emission tomography (PET), offering unprecedented sensitivity and enabling ultra-low-dose 4D total-body imaging. However, long-axial field-of-view (LAFOV) systems introduce significant challenges, including prohibitive computational loads and complex quantitative corrections necessitated by expanded axial acceptance angles. A primary obstacle is the extreme sparsity of acquired data relative to the vast lines-of-response (LOR) space. We present a robust quantitative correction strategy integrated into QuanTOF, a GPU-accelerated, list-mode reconstruction framework. Our primary contribution is the implementation of a modified component-based normalization (CBN) strategy tailored for axial-dependent variations, effectively mitigating noise amplification in sparse data environments. Furthermore, we introduce a simplified calibration protocol utilizing 0.5 m thin-ring and FOV-length line sources to derive geometric and plane efficiencies—an efficient alternative to conventional through-axial rotating rods. To ensure correction accuracy despite data sparsity, a moderate rebinning strategy was employed for scatter and random estimation. The framework was validated via GATE simulations of the 1.5 m uMI Panorama GS system. Reconstruction of a 1.5 m uniform cylinder phantom (8cm radius) demonstrated superior axial uniformity, with intensity fluctuations constrained to less than 2% within 70% of the FOV. Additionally, Derenzo phantom simulations confirmed the framework's high-resolution capabilities, successfully resolving 2.4 mm hot rods. These results show this framework successfully overcomes the inherent challenges of LAFOV data sparsity, securing the quantitative fidelity essential for total-body molecular imaging.

Abstract

This work addresses efficient partial axial field-of-view (aFoV) reconstruction for long-axial PET systems, where reconstructing the full long-axial dataset and subsequently cropping to a target aFoV is computationally expensive. This need may arise when CT coverage is intentionally limited (e.g., organ-focused protocols or pediatric imaging). However, simply limiting the PET aFoV reduces the effective sensitivity of a long-aFoV scanner to that of a shorter-aFoV system. We propose a limited axial rebinning strategy that retains a subset of oblique lines of response that would otherwise be discarded when truncating data to a partial aFoV, thereby leveraging the sensitivity of the long-aFoV system. The approach selects a starting rebin segment and a bounded set of additional oblique segments (up to a maximum rebin segment), and rebins eligible events into target segments while preserving time-of-flight (TOF) geometry. In an FDG human study on a long-aFoV system, the proposed limited rebinning substantially increases retained true counts for a Vision-equivalent aFoV versus truncation without rebinning, while maintaining comparable axial resolution and avoiding the cost of reconstructing the full long-aFoV image.

Abstract

PET/CT radiation exposure arises from both the injected PET radiotracer and the accompanying CT scan used for attenuation and scatter correction. Reducing PET dose while eliminating CT-based correction is therefore an attractive strategy for lower-radiation PET/CT imaging, but this setting substantially degrades image quality. In this study, we propose a deep learning framework for CT-free low-dose [18F]FDG PET restoration based on a Noise-Level-Estimation network and systematically evaluate its generalizability across different PET scanners. The model restores attenuation- and scatter-corrected PET images from reduced-dose non-corrected inputs using a mixed-dose training strategy. It was trained on a main cohort from Ruijin Hospital, acquired with [18F]FDG on a uEXPLORER scanner, comprising 5,731 reconstructed PET volumes derived from 1,163 examinations. Scanner-domain validation was performed on 640 reconstructed PET volumes derived from 320 examinations acquired on three different scanners, including Discovery MI, Vision 450, and uMI 780. Direct pretrained inference, few-shot scanner-domain finetuning, and from-scratch training were compared. Direct application of the uEXPLORER-trained model showed performance degradation under scanner shifts, whereas few-shot finetuning consistently improved restoration fidelity. In the most challenging uMI 780 cohort, median PSNR increased from 35.24 dB to 47.58 dB after finetuning. These results demonstrate the feasibility of scanner-generalizable CT-free low-dose FDG PET restoration.

Abstract

Neural Network (NN) models have emerged as a promising tool for determining 3D photon impact positions in Positron Emission Tomography (PET) detectors based on monolithic scintillators. Nevertheless, NNs typically require long acquisitions of reference data training, becoming a time-consuming process that requires extensive experimental data for each individual detector module. To reduce calibration time, we propose a new methodology based on NNs trained entirely using simulated data to predict the 3D impact position of gamma rays. The feasibility of the proposed approach was verified using a real preclinical PET system composed of two rings of 8 LYSO:Ce monolithic scintillators each.

The simulation of a single detector was implemented in GATE and reference data moving a pencil beam in 1 mm steps was simulated to train two Multilayer Perceptrons (MLPs), MLP_X and MLP_Y, to predict the x- and y- impact positions, respectively. To validate the introduced methodology at reconstructed image level, experimental data was acquired using a ²²Na point source at the center of the preclinical system and processed using both the presented framework and the system’s standard calibration methodology.

Results have demonstrated that similar spatial resolution was obtained with both approaches, thus demonstrating that NNs trained exclusively with Monte Carlo simulations can generate coherent reconstructed images and thus eliminate the need for calibration measurements for each individual detector module.

Abstract

3D medical image segmentation of CT and MRI volumes is essential for clinical diagnosis and treatment planning. While deep learning methods such as 3D U-Net, Attention U-Net, and nnU-Net achieve strong performance, conventional 3D CNNs treat all slices equally along the axial direction, causing insufficient learning of lesion-containing slices. Existing attention mechanisms such as Attention U-Net focus on height×width (H×W) planar spatial attention without explicitly modeling slice-wise importance along the depth (D) axis. This paper proposes two attention modules to address these limitations. First, a Depth Attention Module (DAM) automatically learns the importance of each slice and integrates it into each decoder layer via residual connection. Second, a volume Attention Module (VAM) extends DAM to independently learn importance along H, W, and D axes, combining three independent axis-wise attention maps to achieve disease-adaptive spatial importance learning. DAM achieves Dice improvements of 7.61%, 4.71%, and 3.72% over 3D U-Net on brain hemorrhage CT (HemSeg-200), kidney CT, and ischemic stroke MRI, respectively, and 8.52% over nnU-Net on lung cancer CT (MSD Task06). On brain hemorrhage CT, DAM further improves over Attention U-Net by 4.54% and over nnU-Net by 1.03%. VAM achieves improvements of 7.76% over 3D U-Net on brain hemorrhage CT and 2.83% over nnU-Net on lung cancer CT. Experimental results confirm the effectiveness of the proposed modules across multiple diseases and imaging modalities.

Abstract

PET(O15) is the gold standard for assessing cerebral perfusion in vivo. Assessment of perfusion deficits before surgical interventions in Moyamoya patients with PET(O15) is hampered by poor spatial resolution and low signal-to-noise ratio (SNR). We investigate the use of anatomically-guided reconstruction (AGR) for improving the assessment of baseline perfusion metrics in Moyamoya patients. To evaluate PET(O15) perfusion metrics in Moyamoya patients with and without the use of anatomically guided reconstructions. We assess perfusion metrics across different anatomical brain parcellations to evaluate the differences between the conventional and AGR-enhanced PET(O15) images. Regions that overlap with clinically relevant regions are significantly different in signal intensity when comparing the use of the reconstruction method. Many of which overlap with those reported in the Moyamoya literature. Implementation of the anatomy-guided PET reconstruction(AGR) algorithm can provide clinicians with a better understanding of the disease progression, especially if data integrity is limited, and of low resolution. Particularly in cases of radioisotopes with short decay half-life, such as O15, or when studying diseases of tissue perfusion and oxygen metabolism.

Abstract

Spatial normalization enables standardized volume-of-interest quantification in brain perfusion single-photon emission computed tomography, but accurate alignment is difficult because of low resolution, noise, partial volume effects, and uptake-dependent variation. We propose a MRI-informed, SPECT-only framework that distills anatomical deformation fields and feature guidance from a magnetic resonance imaging registration teacher. The student disentangles anatomical and uptake-related representations to suppress functional-uptake interference while requiring only single-photon emission computed tomography at inference. In 173 paired imaging datasets, the method improved standardized uptake value ratio consistency between subject and template spaces, especially for fine-grained volumes of interest, supporting MRI-free, single-pass template-based quantification.

Abstract

Dynamic ^99mTc-DTPA renography provides both spatiotemporal image information and clinically interpretable time-activity curve (TAC) and region-of-interest (ROI) descriptors for renal functional assessment. However, how these heterogeneous information sources should be integrated for automated Gates-derived total glomerular filtration rate (GFR) estimation remains unclear. This study developed a residual multimodal learning framework in which a 3D CNN image branch first estimated Gates-derived total GFR from multi-frame renography sequences, and selected TAC/ROI descriptors were then used to predict a residual correction term rather than being directly concatenated with image features. A total of 471 complete paired examinations were retrospectively included, with 376 cases for training and 95 cases for held-out testing. Representative image-only, TAC-only, direct-fusion, ROI/TAC tabular, and residual-correction models were compared under the same test split. The dynamic image-only model outperformed perfusion TAC-only, function TAC-only, and pure ROI/TAC tabular models, indicating that the multi-frame renography sequence was the dominant information source. Direct fusion of image and TAC inputs did not improve image-only prediction. In contrast, the proposed residual-correction model achieved the best overall performance, with the lowest mean absolute error of 8.21 and root mean square error of 10.93. These findings suggest that TAC/ROI descriptors are more useful as clinically interpretable residual correction cues than as standalone predictors or directly fused inputs. The proposed framework provides a practical strategy for approximating total GFR from routine dynamic ^99mTc-DTPA renography while clarifying the complementary role of TAC/ROI descriptors.

Abstract

Lumbar spine MRI segmentation is an important prerequisite for quantitative spine image analysis; however, it remains challenging due to complex anatomy, low contrast, and high inter-patient variability. Recent diffusion-based approaches have shown potential for medical image segmentation; nevertheless, preserving anatomically consistent boundaries remains challenging, particularly in low-contrast regions. We propose a conditional diffusion-based framework for lumbar spine MRI segmentation with uncertainty-guided localized refinement. The model integrates a UNet-based denoising backbone with multi-scale feature encoding to preserve anatomical structures during the diffusion process. Furthermore, an entropy-driven uncertainty mechanism is introduced to identify ambiguous regions, enabling localized refinement and selective diffusion re-sampling in low-confidence regions to support boundary delineation. The framework is evaluated using 5-fold cross-validation on multi-contrast (T1- and T2-weighted) MRI data, achieving Dice scores of 0.87-0.89 with low surface errors (ASSD ≈ 1-2 mm, HD95 < 10 mm). The current results demonstrate consistent segmentation performance across spinal canal, vertebral bodies, and intervertebral discs, indicating the feasibility of automated multi-structure lumbar MRI segmentation. These results suggest that the proposed framework can serve as a technical foundation for automated quantitative spine MRI analysis. In future work, this framework will be extended to vertebral body and corner-region analysis for edema detection, assessment of corner changes, and quantitative MRI evaluation of axial spondyloarthritis.

Abstract

Total-body positron emission tomography (PET) increases system sensitivity and enables whole-body dynamic imaging, but its extended axial field of view makes accurate depth-of-interaction (DOI) calibration essential for maintaining uniform spatial resolution and reducing parallax error. Conventional DOI calibration requires mechanically scanned collimated sources or detector-specific reference measurements, which become impractical as the number of detector modules increases. PhysiCal-PET introduces a physics-based neural-network framework for autonomous in-system DOI calibration of PET scanners. The method trains a network to map raw detector signals to DOI using coincidence data acquired with an uncollimated positron source at known positions. For each event, the loss function enforces annihilation geometry by requiring the predicted line of response to be consistent with the source position, thereby providing supervision without measured DOI labels. The framework was tested on realistic PET simulations and applied to data from a real preclinical PET scanner. Using this procedure, the DOI response of all 144 detector modules was calibrated in less than 2 hours of point-source measurement time. These results demonstrate that the annihilation physics can replace labor-intensive DOI calibration hardware and enable rapid, routine calibration of high-resolution PET systems. PhysiCal-PET therefore provides a scalable path toward autonomous calibration for total-body PET and other systems that rely on accurate DOI positioning.

Abstract

This paper presents a Nonsubsampled Contourlet Transform-Based Generative Adversarial Network (NSCT-GAN) for low-dose computed tomography (LDCT) denoising. LDCT reduces radiation exposure but introduces quantum Poisson and electronic Gaussian noise, which blurs anatomical structures. Conventional deep learning methods using mean-squared error loss produce over-smoothed images, while existing image-domain generative adversarial networks (GANs) struggle with high-frequency CT noise. To address this, the proposed NSCT-GAN extracts multi-scale, multi-directional, and shift-invariant high-frequency features using the NSCT. The network trains a discriminator on these high-frequency images to distinguish noise from anatomical edges. The generator utilizes a U-Net architecture to maintain spatial consistency. The objective function combines a Wasserstein GAN with gradient penalty adversarial loss for high-frequency reconstruction, an L1 loss for accurate CT number recovery, and a multi-scale structural similarity (MS-SSIM) loss for global and local structural consistency. The method was evaluated using the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge dataset and compared against five deep learning-based denoising methods: WavResNet, SMGAN, CPCE, DU-GAN, and WDLP. Quantitative results demonstrate that NSCT-GAN achieves a peak signal-to-noise ratio (PSNR) of 27.1850 dB, a structural similarity index (SSIM) of 0.8215, and a feature similarity index (FSIM) of 0.9698, yielding the highest SSIM and FSIM values among the tested models. Qualitative evaluations show the method restores continuous vascular boundaries in the lungs and maintains heterogeneous parenchymal patterns in the liver without introducing over-smoothing or artifacts.

Abstract

Positron range is an inherent physical factor that degrades the spatial resolution of positron emission tomography (PET), particularly for radionuclides with high positron energies. This study investigated the feasibility of deep learning-based positron range correction for PET brain imaging. A total of 304 brain datasets acquired using a Siemens Biograph Vision PET/CT scanner at the First Affiliated Hospital of the University of Science and Technology of China were voxelized to generate brain phantoms. These phantoms were used as activity maps for GATE simulations with three emission settings: ¹⁸F, ⁶⁸Ga, and back-to-back sources. The ¹⁸F and ⁶⁸Ga simulations represented mild and severe positron range effects, respectively, whereas the back-to-back simulation served as the range-free reference. Two deep learning-based correction methods were evaluated, including a conditional denoising diffusion probabilistic model (C-DDPM) and a residual U-Net (Res-UNet). Visual results showed that C-DDPM better restored fine sulcal and gyral structures and achieved closer agreement with the back-to-back reference, especially for the more severe ⁶⁸Ga correction task. Although Res-UNet achieved superior PSNR, SSIM, and NRMSE values, its restored images exhibited noticeable oversmoothing and partial loss of subtle anatomical details. These findings suggest that diffusion-based generative restoration is a promising strategy for positron range correction, with the potential to reduce positron range induced blurring while preserving fine anatomical structures.

Abstract

Low-dose positron emission tomography (PET) reduces injected activity and acquisition time, but low-count statistics introduce strong noise, impaired lesion conspicuity, and unstable standardized uptake value (SUV) quantification in whole-body imaging. Recent 3D diffusion models improve PET denoising, and wavelet-conditioned diffusion further shows that the Haar low-frequency LLL subband can serve as a robust structural prior. However, diffusion sampling is computationally expensive for large 3D PET volumes. We propose Wavelet-Conditioned Rectified Flow (WCRF), a deterministic whole-body PET restoration framework that integrates wavelet-domain structural guidance with rectified-flow matching.

Abstract

A spatio-temporal denoising training strategy for low-dose fluoroscopy is proposed, based on a spatially-adaptive temporal consistency metric. This approach acts on a small set of consecutive frames coupling a conventional denoising operator with an adaptive motion-gradient attention gate to selectively modulate the temporal consistency terms within the loss function, ensuring effective noise suppression and flickering reduction while maintaining high anatomical fidelity in dynamic regions.
The proposed method was used to train a denoising operator in which a three-frame temporal window is used as input. The model was trained across multiple dose reduction factors (qF), and demonstrated superior structural recovery and enhanced temporal stability in the processed image sequences, compared to single-frame models. Quantitative and qualitative evaluation of image quality, along with stability maps—defined as the Normalized Cross-Correlation (NCC) of temporal Coefficient of Variation (CV) maps—confirms high fidelity relative to high-dose reference acquisitions, effectively mitigating stochastic fluctuations without introducing motion artifacts.

Abstract

Carbon-ion radiotherapy (CIRT) combines high biological effectiveness with outstanding treatment precision. Since treatment plans are based on the most recent Computed Tomography (CT), anatomical changes during multi-week treatments may compromise dose delivery accuracy. To address the increasing uncertainty about the patient’s anatomical state with time elapsed since the planning CT, our group developed an in vivo monitoring method using secondary charged particles emitted during therapy. This non-invasive approach requires no additional dose and reduces uncertainty about patient anatomy and thus, loss of treatment accuracy. It uses a custom detection system to detect nuclear fragments emmited from the patient and allows approximating their origin. To reconstruct these changes in 3D, we utilize our recently published method for detecting the depth (z) of changes along the beam axis and apply it to a sliding window (x,y) along the transverse plane. However, splitting reconstructed nuclear fragment origins into smaller subsets reduces the statistical stability of each subset distribution. In this study, we present an approach to robustly distinguish whether a windowed subset belongs to a region with or without an anatomical change based on correlation matrices generated by our depth detection method. These images contain location-wise correlations of joined frequency band variations between projections of nuclear fragment measurements from different treatment days. To classify regions of change and approximate change depth, we train an ensemble of deep learning-based image classifiers with different architectures (i.e. ResNet, RegNet and ConvNeXt) and combine their outputs into an ensemble decision. The approach was validated using dedicated irradiation measurements with coin-sized cavities inserted into PMMA head phantoms. Despite limited training data, we achieved robust improvements in change localization using an ensemble of different deep learning architectures.

Abstract

Despite its widespread use for compensating attenuation in positron emission tomography (PET), computed tomography (CT)-based attenuation correction has limitations which can create imaging artifacts associated with breathing motion and metallic implants. Using deep learning networks to create synthetic attenuation correction maps represents a data-driven, non-ionizing alternative with the potential to reduce the frequency and/or intensity of such errors. In this study, the authors aimed to demonstrate that using synthetic maps would reduce breathing motion and metallic implant artifacts. To achieve this, ten anonymized whole-body F¹⁸-Fluorodeoxyglucose exams were assessed. Scans were divided into two Cohorts; Cohort 1 comprised exams with breathing motion artifacts, whilst Cohort 2 contained scans of patients with hip replacements. Each scan was corrected separately with CT-based attenuation correction and synthetic maps prior to identical reconstruction. Quantitative assessments of standardized uptake value (SUV) parameters were performed, and differences were tested for statistical significance. Additionally, each image’s lesion detectability, diagnostic confidence and overall image quality were scored on a Likert scale by an experienced radiologist. After assessment, no statistically significant differences in SUV quantification were detected (p > 0.05). Images corrected with synthetic maps were preferred by the radiologist in Cohort 1, but no preference was expressed in Cohort 2. Moreover, a single Cohort 1 case was identified in which breathing motion artifacts impacted clinical diagnosis only for the image treated with CT-based attenuation correction. Thus, it was concluded that synthetic maps reduce breathing motion artifacts to an extent which may support clinical diagnoses, although their effects on metallic implant artifacts require further investigation.

Abstract

Full-volume 3-D medical image generation at clinical resolutions remains a significant challenge due to GPU memory constraints. Existing voxel-space generative methods are difficult to scale to large clinical volumes, as memory and compute costs grow rapidly with spatial resolution. Latent diffusion approaches offer a more scalable alternative [1], but they typically require training a dedicated medical autoencoder for each modality or dataset before generative modeling can begin, making the compression stage itself a major bottleneck [2], [3], [4]. A lighter alternative is to encode axial slices independently with a 2-D autoencoder and stack the resulting latent maps into a 3-D tensor [5]; however, this still requires training a modality-specific encoder from scratch, limiting generality across modalities and resolutions.
We propose a scalable latent flow framework that eliminates this bottleneck by using a frozen pretrained Stable Diffusion VAE as a ready-to-use slice-wise encoder for medical volumes, requiring no modality-specific retraining or adaptation. Each axial slice is resampled to a common resolution, intensity-normalized, and encoded independently by the frozen encoder, yielding a high in-plane compression factor compared to dedicated medical autoencoders. The resulting latent maps are stacked into a compact 3-D tensor, on which a full-volume 3-D U-Net flow model is trained directly. The objective predicts clean latents and supervises in velocity space [6], ensuring stable optimization and sharp generations in high-dimensional latent spaces, avoiding the blurriness and instability observed with standard diffusion objectives.
We validate our framework on CT, PET, and MRI volumes at clinical resolutions, demonstrating strong encoding fidelity and coherent full-volume generation across axial, coronal, and sagittal planes on a single consumer-grade GPU, without any modality-specific autoencoder training.

Abstract

Dedicated standalone brain positron emission tomography (PET) scanners have emerged in recent years as a promising technology for neuroimaging, offering high spatial resolution, increased sensitivity, reduced injected dose and shorter acquisition times. However, unlike hybrid PET/computed tomography (PET/CT) or PET/magnetic resonance imaging (PET/MRI) systems, these scanners do not provide an anatomical image from which attenuation information can be derived. We present a SynthSR-based framework to predict CT-derived attenuation maps from non-attenuation-corrected (NAC) PET images, aiming to enable attenuation correction in standalone brain PET. Building on SynthSR, a large and widely variable set of synthetic PET volumes is generated to sample a space larger than the potential clinically viable PET images, and thus increasing the robustness and clinical applicability of the model. The synthetic PET volumes are generated on-the-fly from anatomical labels using label-conditioned Gaussian-mixture intensity sampling, spatial/intensity augmentation and PET-specific priors estimated from real normalized PET data. A PET appearance model introduces count-based noise, correlated texture, smoothing, and controlled intensity variability. Finally, spatial transformations are used to increase sample diversity while preserving voxel-wise alignment. A three-dimensional U-Net was trained to predict normalized attenuation maps from synthetic PET, while validation and testing were performed on never-seen real PET/CT data. The model was optimized with 24,000 on-the-fly synthetic samples and evaluated on held-out real PET/CT data. The best validation checkpoint achieved MAE=0.03 (3%), PSNR=25.74 dB, and Pearson correlation r=0.859. On the independent test subject, the model achieved MAE=0.027 (2.7%), PSNR=26.35 dB, and r=0.854. These results support the feasibility of CT/MRI naive pseudo-CT generation for PET attenuation correction in novel standalone dedicated brain PET scanners.

Abstract

Targeted Alpha Therapy with actinium-225 induces double-strand DNA breaks through individual alpha decays, motivating quantitative imaging of single-event distributions for dose assessment. We address point detection in a cerium-doped gadolinium aluminum gallium garnet scintillator coupled to an electron-multiplying camera, in which each alpha produces a point-spread function of approximately 2-micrometer diameter at 10 to 150 events per 100-millisecond frame. We compare three method groups on Monte Carlo simulated images: 20 classical preprocessing-and-clustering pipelines, two deep learning detectors based on the U-Net and YOLOv8-nano architectures, and Alpha-CenterNet, a heatmap-based detector that adapts CenterNet with a feature pyramid network and a modified focal loss. The highest-F1 classical pipeline attains 0.641 with 447-nanometer mean localization error. U-Net reaches F1 = 0.978 with 302-nanometer error; YOLOv8-nano reaches 185-nanometer error with F1 = 0.938; the two metrics produce divergent rankings. Alpha-CenterNet (3.78 million parameters) achieves F1 = 0.965 and 244-nanometer error, reducing localization error by 19 percent relative to U-Net at 26 percent of its parameter count, and reaching the validation F1 plateau of 0.98 in 29 versus 86 epochs—a 2.97-fold convergence speedup within the same ablation. The modified focal loss decouples F1 from localization error: F1 changes by 0.0004 while the mean error decreases by 76 nanometers, and this effect persists across all tested densities. The method generalizes to clustered and edge source distributions at unseen densities, supporting point detection for Targeted Alpha Therapy monitoring.

Abstract

Artificial-intelligence-assisted PCB layout is an emerging capability that promises to reduce design iteration time and enable rapid exploration of placement and routing alternatives, which is particularly valuable for complex, multi-channel analog front-end boards where manual layout is labor-intensive and experience-dependent. In this work, we explore the application of an AI-assisted layout workflow to a timing-sensitive silicon photomultiplier (SiPM) front-end readout board for time-of-flight positron emission tomography (TOF-PET) detectors. A four-channel RF front-end readout board, extending a validated single-channel timing topology, underwent automated placement and routing via AI layout tools. To establish a performance baseline, a single-channel prototype of the same RF topology was fabricated and characterized using a coincidence measurement setup with LYSO scintillators, achieving a coincidence time resolution of 100.78 ps and confirming the near-100-ps timing capability of the front-end design. Preliminary evaluation of the AI-generated four-channel layout shows that the tool achieved high routing completion and produced a viable candidate layout. Based on these encouraging results, we plan to fabricate the AI-generated board and systematically compare its timing performance, energy response, and channel consistency against our manually designed baseline, providing a direct, measurement-level assessment of AI-assisted layout quality for timing-sensitive detector electronics.

Abstract

Simultaneous dual tracer PET provides complementary physiological information compared with single tracer imaging. For example, combining [⁸⁹Zr]labeled monoclonal antibodies (mAbs) with [¹⁸F]fluorodeoxyglucose (FDG) enables the assessment of both targeted uptake and glucose metabolism, potentially improving patient selection [1]. Existing dual tracer separation methods exploiting differences in tracer kinetics are typically applied post-reconstruction in the image domain. However, these approaches may be sensitive to noise propagation. In this work, we evaluate a sinogram-domain iterative separation method that is incorporated directly into the reconstruction process, enabling direct generation of single tracer images.

Abstract

Proton radiography and tomography often use residual energy or water equivalent path length (WEPL) to estimate relative stopping power (RSP). Since the maximum likelihood path (MLP) formalism of Schulte et al. (2008), proton trajectories in tissue have been modeled as curved paths constrained by entry and exit tracking. In full proton CT, this path estimation is coupled to energy dependent multiple Coulomb scattering and longitudinal energy loss, which becomes computationally and physically complex at low residual energies. Most MLP based reconstructions use the estimated path mainly to assign WEPL or RSP contributions to voxels. Here, we examine a thin object regime in which scattering itself provides imaging contrast. For high energy protons traversing a human hand, the integrated energy loss is small, allowing the scattering relevant βcp factor to be approximated as constant along the path. Under this constant energy approximation, we reconstruct a volumetric scattering density map using an M-value calculated from each proton scattering angle and kinetic energy, together with the μTRec Bayesian voxel attribution algorithm developed for muon scattering tomography. The method is demonstrated in Geant4 using an anthropomorphic hand phantom containing bones, muscles, tendons, and skin. A single orientation 200 MeV parallel proton beam with 10^7 protons is simulated and reconstructed on a 0.3 mm voxel grid. The resulting proton radiograph resolves carpals, metacarpals, and phalanges, and detects a submillimetre osteochondral gap at the thumb interphalangeal joint. These results suggest that constant energy, MLP inspired scattering radiography can provide a computationally lightweight alternative or complement to energy loss based proton imaging for thin anatomical sites.

Abstract

The growing importance of theranostics and research with long‑lived isotopes demands imaging systems far more sensitive than standard PET or SPECT. Many emerging isotopes lie outside current detector capabilities, resulting in low sensitivity, high background, and poor quantification, as seen with ⁹⁰Y verification and ¹²⁴I or ⁸⁹Sr imaging. The Multiple Emission Tomography (MET) project introduces a novel imaging device based on inorganic scintillators, capable of operating in dual modes: PET-style coincidence detection of 511 keV photon pairs and single-photon Compton-scattering detection in a single-layer configuration. The technical configuration of the MET prototype utilizes a DOI-capable light-sharing scheme. Each module consists of a 16x16 matrix of 1.53x1.53x20 mm³ depolished LYSO crystals coupled to an 8x8 array of 3x3 mm² Hamamatsu Silicon Photomultipliers (SiPMs). The detector's readout electronics are based on the ALCOR ASIC, a low-power, mixed-signal device featuring single-photon sensitivity, a high bandwidth of up to 2 Mevents/s per channel, and a fine 25 ps time binning. Supported by a combination of analytic algorithms and a machine learning pipeline, the MET detector can successfully solve the temporal sequence of inter-crystal scatters, and determine the points of interaction and the energies deposited. To fully leverage these recovered events, a dedicated Compton image reconstruction algorithm has been developed, employing an event-by-event statistical model with ad-hoc projector functions. This technology is expected to significantly impact theranostics, collimator-less SPECT, and in-vivo range verification for CPT.

Abstract

Gold and gadolinium-based nanoparticles have been investigated as theranostic agents whose characteristic X-ray fluorescence (XRF) emissions enable non-invasive, element-specific biodistribution imaging. This work presents the design, assembly, and characterization of a scalable NaI(Tl)-SiPM scintillation detector module developed as a building block for a 16-element detector array for the Confocal Orthovoltage X-Ray--Induced Radiation System (COXIRIS) Project. Each module comprises a 25.4x25.4x4 mm³ monolithic NaI(Tl) crystal optically coupled to a custom SiPM array and read out by the TOFPET2 ASIC.
The architecture targets the K-shell fluorescence lines of gold (Kα = 68 keV) and gadolinium (Kα = 43 keV), while providing a large active area at low cost.
Four modules were characterized with a 2.0 μCi ²⁴¹Am source, whose 59.5 keV emission lies within the 40--80 keV target interval. The peak was clearly identified in all modules, with energy resolutions ranging from 47% to 65%. Per-pixel gain calibration produced an additional relative improvement of approximately 2% for each module. These results support the feasibility of the proposed assembly approach as a detector building block for XRF imaging applications.

Abstract

Positron emission tomography (PET) is traditionally limited to scanning one tracer per session. Positron and prompt gamma emitters may be used to form triple coincidences that allow for signal separation based on energy. We present a proof-of-concept, preclinical PET system designed for detecting and distinguishing prompt gamma and annihilation photons in the same detector. We describe the data processing pipeline with advances for estimating randoms, scatter, and sensitivity of the prompt gamma photon. System performance is characterized with NEMA and Derenzo phantom studies. Next, we present multi-isotope studies using ⁵²Mn and ¹⁸F in a custom image quality phantom. We compare two reconstruction methods and find alternating updates yield faster convergence and higher image quality. After correcting for injected activity and decay, the mean voxel intensity ratio between ¹⁸F and ⁵²Mn across background ROIs is 0.95 ± 0.06, where 1 represents perfect quantitative agreement. The uncertainty reflects variability across ROIs and dose calibrator uncertainty (±2%), suggesting unmixing accuracy is within dose calibrator tolerance. Further, we scan the phantom after ¹⁸F decay to determine the relative image quality between single (⁵²Mn only) and dual (⁵²Mn with ¹⁸F) acquisition. We find the ⁵²Mn image has similar contrast recovery (0.65/0.62) and crosstalk ratio (0.24/0.24) between the single and dual acquisition. Due to low triples sensitivity, simultaneous acquisition of both radionuclides reduces the contrast to noise from 10.2 to 4.7. Despite limitations on axial field of view, data rate, and the high background from ⁵²Mn’s 3 gamma cascade, we achieve high quantitative accuracy. Future studies will scale up the system, use more favorable tracers, add regularization to mitigate noise propagation and include preclinical animal scans.

Abstract

X-ray scattering tensor tomography retrieves directional dark-field projections from the focused, structured illumination produced by 2D diffractive optics, and is increasingly attractive as a route to anisotropic micro-structural imaging on laboratory sources. Quantitative tensor reconstruction requires the focused reference field to cancel out exactly in the Fourier-domain retrieval, a condition met only under monochromatic illumination. We show that beam hardening through an absorbing sample reshapes this focused reference field non-uniformly across its bright core and the surrounding plateau, so the retrieval assigns a non-zero scattering tensor to a purely absorbing object—a pseudo dark-field signal. We derive a closed-form expression for the pseudo dark-field signal, prove its exact isotropy arising from the circular symmetry of the focal spot, and identify the retrieved observables that are biased by beam hardening versus those that are not. The standard angular-consistency metric is blind to this bias, so a dedicated calibration is required. We describe a step-wedge calibration protocol that compresses the family of pseudo dark-field signals into a single thickness-dependent function suitable for per-pixel correction. Removing this artefact is a prerequisite for quantitative X-ray scattering tensor tomography on laboratory sources.

Abstract

Simultaneous dual-radionuclide imaging offers complementary information for diagnosis, stratification, treatment planning, and therapy evaluation. However, the fundamental differences between PET and SPECT present significant challenges for integrated imaging. We propose a scintillator-as-collimator detector enabling both PET and SPECT. It consists of three scintillator layers separated by two light guides. Parallel holes in the upper portion function as a SPECT collimator, referred to as the scintillator-as-collimator volume (SCV). The lower portion serves as the SPECT detector (SD). In PET mode, the entire detector acts as a multi-resolution TOF-DOI PET detector. Thus, it serves as a self-collimated SPECT or TOF-DOI PET detector. A dual-panel prototype was developed with rotational acquisitions for PET, SPECT, and simultaneous imaging. SPECT events are selected using flood image mask, maximum SiPM channel energy (Pmax), and DOI-integrated charge (QDC). SPECT resolution reaches ~5 mm. PET performance is preserved, with best SD–SD CTR of 352 ps and two-layer DOI. The introduction of holes has minimal impact on PET detection, enabling high-quality PET images. In simultaneous imaging, a unified timing framework rejects 511 keV down-scatter, and residual artefacts were further mitigated using a 200–240 keV energy window for scatter correction. Phantom experiments validate tri-mode imaging. This detector eliminates external collimators, supports multi-tracer imaging, and offers a cost-effective single-detector-set PET&SPECT system.

Abstract

Diffusion models have shown strong potential for improving PET image quality. However, existing methods based on unconditional pretrained diffusion priors often struggle to balance realism and fidelity, particularly in preserving high-resolution details. To address this issue, we propose MI-ResShift, a zero-shot mutual information-guided diffusion model with residual shifting for PET image super-resolution. MI-ResShift introduces a differentiable local-wise mutual information layer to guide the denoising process using structural information from CT or MRI, enabling robust cross-modal feature alignment in the statistical domain. By combining mutual information guidance with residual shifting, the proposed method enhances PET detail fidelity and robustness while reducing diffusion-based inference to only four steps. Experiments on PET/MR brain datasets demonstrate that MI-ResShift improves image quality over classical interpolation methods and provides an efficient solution for cross-modal PET super-resolution.

Abstract

Coincidence-based imaging with the back-to-back 511 keV photon pairs emitted during positron annihilation enables transmission imaging with reduced quantum noise. Conventional transmission imaging recovers only a single attenuation-related quantity per voxel or pixel. In this work, we extend coincidence-based imaging to simultaneously recover both the total and the Compton-scatter linear attenuation coefficients from a single acquisition, providing a second axis for material identification. Forward-scattered photons collected in coincidence within an analytically defined polar window, obtained by integrating the Klein-Nishina differential cross section over a 20-to-30-degree band, enable direct estimation of the Compton macroscopic cross section, while the idler channel provides the incident count for the total coefficient. We evaluate the approach using Monte Carlo simulations of opposing cadmium zinc telluride panels viewing a water phantom of thickness 2 to 10 millimeters. Idler-normalized transmission was recovered to within 0.1 percent across all thicknesses, while the extracted total and Compton coefficients exhibited 1 to 4 percent absolute error, with the Compton coefficient transitioning from over- to underestimation near 6 millimeters. These deviations arise from logarithmic amplification of small counting errors and from competing air-scatter contamination at thin phantoms and self-attenuation of scattered photons at thicker ones. The results establish that scatter data from coincidence-based annihilation imaging can recover a second material descriptor, motivating further development toward dose-efficient material identification.

Abstract

The development of lightweight radiation shielding materials is of high interest for a broad range of defense, space, and medical applications. In particular, mass-savings are desired in vehicles (ground, naval, and/or aerospace) and as wearable fabrics in high dosage environments such as those encountered in the medical setting. Here we investigate the potential of enhanced shielding via radiation attenuation of nanostructured materials which take advantage of interfacial and nanocrystalline loss processes increase the multiplicity of energy loss mechanisms within the shielding solid. In this work, colloidal metallic nanoparticles grown in an aqueous environment (such as palladium (Pd) and silver (Ag)) with sizes in the 5 - 50 nm range are imbedded in a polymeric matrix material (epoxy or aramid nanofiber (ANF)) to demonstrate photon and electron shielding while maintaining mechanical properties such as strength, durability, and form-factor flexibility. The mechanisms of energy loss at the interfaces of the nanomaterials and overall in the solid are quantified through transport modeling as well as the unorthodox usage of the Transmission Electron Microscope (TEM), in which a focused electron beam and state-of-the-art detector system are utilized to measure the effects of particle interactions within our material system to a near single particle resolution. Therefore, this work encompasses both an investigation of the potential enhancement of radiation attenuation by nanostructured materials while also introducing an innovative nanocomposite material system with manufacturing versatility enabling lightweight, flexible shielding.

Abstract

The Sensing Collimation Imager (SCI) is an ultrasensitive gamma detector that integrates multi-pinhole, coded aperture, and self-collimation imaging into a single compact device, achieving substantially higher sensitivity than conventional gamma cameras. Maximizing the spatial encoding capability of the SCI module requires optimizing the crystal arrangement over a high-dimensional combinatorial design space, which is too large for exhaustive search.

We propose an information-theoretic detector geometry optimization framework that uses the Fisher Information Matrix (FIM) spatial correlation map as a computationally efficient performance metric. The metric requires only computing the FWHM of the FIM column vector (O(n)), avoiding the O(n³) cost of inverting the FIM for the Cramér–Rao bound, enabling broad geometry searches. A genetic algorithm (GA) guided by this metric searches over fixed-cardinality crystal arrangements to identify high-performing designs.

Applied to the SCI module, the GA converged within 50 generations and identified an optimized arrangement producing narrower FIM correlation peaks throughout the field of view than a heuristic design. GATE Monte Carlo simulations of a Derenzo phantom confirmed improved recovery of the 8 mm sphere group with clearer feature separation, validating the optimization framework.

Abstract

Whole gamma imaging (WGI) is a novel imaging technique that combines PET and Compton imaging by inserting a scatterer detector ring inside a PET detector ring. One of the suitable nuclides for WGI is ⁸⁹Zr. ⁸⁹Zr emits a 909 keV single gamma ray after undergoing positron emission or electron capture to form ^89mY. Because these 909 keV gamma rays are emitted approximately four times more frequently than positrons, hybrid imaging of WGI combining both modalities is expected to improve image quality compared to conventional ⁸⁹Zr-PET which uses only annihilation gamma rays. In this study, we aimed at finalizing the design of a clinical whole-body WGI system. The hybrid imaging method was applied to Geant4-simulated data of a whole-body WGI system. We modeled a system employing HR-GAGG crystals as the scatterer and LGSO crystals as the absorber. Measurement of an ⁸⁹Zr-filled NEMA IEC body phantom placed at the center of the field of view was simulated. Images were reconstructed using the hybrid reconstruction method combining PET and Compton events and compared with PET images reconstructed by the list-mode OSEM algorithm using only PET events. We calculated the percent contrast (Q_H,j) and percent background variability (N_j) for each sphere size j (10-37 mm). The results showed that, for large spheres with diameters of 22 mm or more, hybrid imaging reduced N_j, but the maximum achievable Q_H,j was  lower than that of PET. In contrast, for small spheres with diameters of 17 mm or less, hybrid imaging demonstrated a specific advantage by suppressing noise amplification while accelerating contrast recovery in the early iterations. In conclusion, although the overall image quality improvement was limited in certain aspects, the potential to enable novel applications such as multi-isotope imaging and positronium lifetime imaging using appropriate position emitters was suggested.