Online program Overview Session: MIC-11

Abstract

Accurate positron emission tomography (PET) depends on calibration of detector alignment, gamma interaction positioning, and coincidence timing, which jointly define line-of-response (LOR) and time-of-flight (TOF) information. Conventional workflows calibrate these quantities in separate stages using dedicated acquisitions, hardware, or phantoms. This is particularly burdensome for light-sharing detectors, where these quantities are coupled through the detector response but are typically treated as separate calibration tasks. We introduce a differentiable in-system calibration framework that enables rapid scanner-level calibration by replacing staged workflows with a single coincidence-level optimization. A shared neural network predicts local interaction positions and timing corrections from single-event readouts, while learnable detector poses map predictions into scanner coordinates. Singles are processed independently, and coincidence information enters only via physical LOR and TOF construction. This prevents source-label access at the network input while enforcing scanner-level consistency. The framework was trained on 28M coincidences from 450 randomized 22Na point-source positions acquired on a semi-monolithic preclinical PET system. In hot-rod reconstructions, the proposed calibration achieved image quality on par with a conventional blueprint and fan-beam workflow, as assessed by peak signal-to-noise ratio, structural similarity, coefficient of variation, recovery coefficients, and peak-to-valley ratios. Differentiable in-system calibration therefore compresses labor-intensive staged workflows into a unified physics-constrained optimization completed in hours, reducing effort and mechanical constraints without sacrificing reconstruction performance.

Abstract

PET is a cornerstone of preclinical imaging, where achieving ultra-high spatial resolution is paramount for the precise quantification of sub-regional metabolism. While commercial scanners typically yield 1.0–1.5 mm resolution, advanced research prototypes target sub-millimeter precision. This study introduces a compact preclinical PET scanner achieving NEMA-standard sub-0.5 mm spatial resolution via the analytic filtered back-projection (FBP) algorithm. To the best of our knowledge, this is the first reported system demonstrating such resolution through FBP. The dual-ring scanner features a 64 mm transaxial and 53.2 mm axial field of view, integrating 16 modules of 44 × 44 LYSO crystal arrays (0.497 × 0.497 × 5 mm³ elements). These arrays are coupled to Hamamatsu Multi-Pixel Photon Counters (MPPCs, S14161-3050HS-08) and managed by a PETsys TOFPET2 ASIC system. System performance was characterized following NEMA NU 4-2008 protocols. The scanner demonstrated an overall energy resolution of 13.43% and an absolute sensitivity of 0.87% (350–650 keV). Peak true and noise equivalent count rates (NECR) were 7.61 kcps and 6.02 kcps, respectively, with a scatter fraction of 8.92%. FBP reconstruction of a point source demonstrated the system achieved average FWHM resolutions of 0.90 mm, 0.86 mm, and 0.69 mm in the radial, tangential, and axial directions, respectively, reaching a peak resolution of 0.48 mm. Imaging of a Micro-Derenzo phantom resolved 0.4 mm rods via iterative reconstruction. In vivo studies using ¹⁸F-FDG and ¹⁸F-Florbetazine delineated precise tracer distributions within sub-regional structures. While the 5-mm scintillator thickness involves a trade-off in sensitivity, the system establishes a benchmark for spatial fidelity in preclinical instrumentation. Future work will integrate thicker crystals with depth-of-interaction capabilities to enhance sensitivity without compromising spatial precision.

Abstract

Current positron emission tomography (PET) imaging is fundamentally limited by physical background noise, such as scatter and random coincidences, which significantly degrade image quality and quantitative accuracy. Quantum-entangled PET (QEPET) is a promising solution that utilizes the inherent orthogonal polarization correlation of annihilation photon pairs to physically reject noise. However, its practical application to in-vivo imaging has been a major challenge due to the severe reduction in usable event statistics caused by strict event selection. In this study, we developed a triple-ring PET scanner consisting of 24 Ce:GAGG modules (8×8 arrays of 2.5×2.5×9.0 mm³ pixels at 3.2 mm pitch; 1536 pixels in total) arranged in three axial rings with an inner radius of 36.7 mm, read out by SiPMs via dynamic time-over-threshold circuits, achieving an energy resolution of 9.5% at 511 keV and a coincidence timing resolution of 16.8 ns. By selecting events within a scattering-angle range of 70°–121° and an azimuthal-angle difference of 80°–100°, we obtained a modulation contrast ratio of 1.91, which clearly exceeds the classical limit of 1.63, confirming the entanglement of annihilation photon pairs. Furthermore, to overcome the limited counting statistics of conventional QEPET, we developed a novel reconstruction algorithm that seamlessly integrates the high-statistics structural information of standard PET into the iterative MLEM process. In phantom experiments, this approach improved the SBR to ~1.25× that of raw QEPET while preserving an SNR ~1.27× that of standard PET, achieving the highest CNR among the compared methods. Finally, we performed in-vivo QEPET imaging of a tumor-bearing mouse. The developed method effectively suppressed the granular noise characteristic of conventional QEPET while demonstrating highly localized tracer accumulation, thereby verifying its superiority over both standard PET and conventional QEPET.

Abstract

Positronium lifetime tomography (PLT) is an emerging imaging modality for imaging positronium lifetime in vivo. Previously, we developed SIMPLE (Statistical IMage reconstruction of Positron Lifetime via time-wEighting), a reconstruction framework for estimating average positron lifetime images, and demonstrated high-resolution imaging on a phantom at clinical activity levels. However, the reconstruction problem remains highly ill-posed, particularly under low-count and high-random conditions. In this work, we propose SIMPLE-DIP, a deep image prior (DIP)-based reconstruction method that regularizes SIMPLE without requiring external training data. The proposed approach employs a single patient-specific U-Net that takes a prior anatomical image as input and jointly constrains both the activity and lifetime-weighted images during reconstruction. Computer simulations using a brain phantom show that SIMPLE-DIP effectively suppresses image noise while preserving lifetime contrast and improving lesion recovery, outperforming the compared SIMPLE formulations regularized by the quadratic penalty or independent DIPs. These results suggest that the proposed SIMPLE-DIP is a promising method for high-resolution positronium lifetime imaging.

Abstract

Personalized 177Lu-PSMA-617 radiopharmaceutical therapy (RPT) for prostate cancer requires patient-specific dosimetry, but post-therapy SPECT imaging for dose computation is not universally performed. In contrast, pre-therapy PSMA PET is widely available and often mandatory for patient selection. PET-based predictive dosimetry therefore has the potential to enable personalized treatment planning before RPT administration. We developed a machine learning framework for organ-level predictive dosimetry from pre-therapy PET, where PET-derived organ features are used to predict kinetic parameters of post-therapy SPECT time-activity curves (TACs). Organ masks for the liver, spleen, and bilateral kidneys were used to extract PET features. Serial 177Lu-PSMA SPECT images at multiple post-injection time points provided reference organ activity values, from which mono-exponential fitting yielded organ TACs and time-integrated activities (TIAs). Absorbed dose was computed from TIAs using the MIRD formalism with organ-specific S values. Our deep learning model predicting kinetic parameters outperformed a comparable deep learning model predicting TIA directly, as well as three classical machine learning models predicting kinetic parameters. This workflow demonstrates the feasibility of forecasting organ absorbed dose from pre-therapy PET, supporting individualized therapeutic regimen planning.

Abstract

Although targeted alpha therapy (TAT) using ²²⁵Ac has demonstrated high therapeutic efficacy, clinical translation remains limited due to redistribution of daughter isotopes, leading to off-target dose. Therefore, discriminating localized emissions from redistributed emitting progeny is significant for small-scale dosimetry and radiopharmaceutical evaluation. Here, we propose a low-cost CMOS-based lensless radiomicroscopy system for direct / discrimination using cluster morphology. Event level features extracted from ²¹⁰Po and ⁹⁰Sr measurements were used for classification applying both a physics-driven probability density function (PDF) approach and a data-driven multilayer perceptron (MLP). Both methods achieved > 98% accuracy with AUC > 0.99. Mixed-source experiments mimicking ²²⁵Ac decay conditions demonstrated robust event-level discrimination. These results indicate that cluster morphology encodes sufficient physical information to monitor daughter isotope migration, enabling pathway toward quantifying parent and progeny distributions in radiopharmaceutical imaging.

Abstract

Actinium-225 (Ac-225) is a promising alpha-emitting radionuclide for targeted alpha therapy, and Compton cameras have emerged as a potential imaging solution for this radionuclide. We developed a physics-guided implicit neural representation (PG-INR) framework for Ac-225 Compton image reconstruction that does not require any ground truth data. Unlike conventional deep learning approaches, the proposed method incorporates physical laws directly into the training process by embedding physics-based loss terms, including maximum-likelihood estimation and Klein-Nishina differential cross section. This enables unsupervised learning solely guided by the measurement results. Furthermore, by representing the reconstructed image as a continuous implicit neural function, the framework achieves resolution enhancement beyond the native voxel grid. The proposed method was evaluated against a conventional ordered subsets expectation maximization (OSEM) algorithm using measured Compton imaging data from a mouse phantom and a Monte Carlo-simulated resolution phantom (from 1.0 to 2.5 mm). For the mouse phantom experiments, PG-INR achieved up to 18-fold faster reconstruction compared to OSEM at comparable resolution. For the resolution phantom simulation, PG-INR successfully resolved 1.0 mm diameter rods with up to 4.8 dB peak signal-to-noise ratio improvement and 0.25 structural similarity index measure enhancement over OSEM. These results suggest that the PG-INR performs significantly better than standard OSEM for Ac-225 Compton imaging.

Abstract

Attenuation correction in PET/CT imaging typically uses a CT-derived attenuation map (μ-map), prone to misalignment artifacts, CT-induced artifacts, and additional radiation dose. CT-free alternatives are therefore clinically attractive, and deep learning methods are a promising solution. As clinical PET centres increasingly adopt diverse tracer protocols beyond FDG, a single attenuation correction model that generalizes across tracers without retraining would further reduce deployment complexity. In this work, we investigate the feasibility of supervised μ-map estimation from non-attenuation-corrected (NAC) brain PET images using a 3D patch-based U-Net, trained and evaluated on patient data from the NeuroEXPLORER, a brain- dedicated PET/CT scanner. A dataset of 353 scans across 12 different tracers was used to study the effect of tracer composition on model generalizability. To our knowledge, this is the first study to systematically assess how training tracer composition affects cross-tracer generalization in deep learning-based CT- free attenuation correction, using a large multi-tracer cohort. The tuned U-Net outperformed Attention U-Net, HighResNet, and Swin UNETR under matched conditions. FDG-only training showed poor generalization to non-FDG tracers, with Normalized Root Mean Squared Error (NRMSE) on average 5× higher in the brain for non-FDG tracers than multi-tracer models across different dataset sizes. Notably, a model trained without any FDG data produced FDG μ-maps as well as a FDG-only model, suggesting the network learns anatomy-based priors that generalize across tracer distributions. Tracer diversity during training did not degrade FDG performance and improved generalization. Weighted resampling to correct FDG/non-FDG imbalance worsened performance when the non-FDG patient pool was small. These results indicate that training set diversity, rather than size or tracer-specific data, is an important driver of generalization.