Online program Overview Session: MIC-01

Abstract

Ultra-low-count SPECT imaging in alpha-particle radiopharmaceutical therapy (α-RPT) remains challenged by low administered activities and severe noise amplification associated with conventional reconstruction. We propose a fully three-dimensional diffusion posterior sampling (3D-DPS) framework for physics-informed SPECT reconstruction integrating learned generative priors with a Poisson-consistent likelihood formulation for emission tomography. A 3D unconditional diffusion model was trained on PSMA PET volumes to learn tracer uptake distributions. To further constrain the inverse problem, patient-specific prior adaptation was introduced through synthetic subject-specific uptake realizations designed to model physiologically plausible tracer variability during training. During reconstruction, the learned priors are coupled with a physics-based SPECT forward model through a Poisson log-likelihood formulation consistent with MLEM reconstruction. Validation was performed using realistic ultra-low-count ²²⁵Ac-PSMA simulations, as well as two clinical ¹⁷⁷Lu-PSMA SPECT acquisitions retrospectively degraded to mimic ²²⁵Ac imaging conditions using Poisson thinning (~10⁵ detected photons) and angular downsampling (120 to 30 projections). Compared with MLEM and regularized MLEM, the proposed framework reduced reconstruction error (NMAE = 0.34 ± 0.04 versus 0.53 ± 0.02 and 0.41 ± 0.05, respectively) and improved structural similarity (SSIM = 0.90 ± 0.05 versus 0.77 ± 0.13 and 0.85 ± 0.08, respectively) while preserving anatomical detail and projection-domain consistency. Reconstructions from degraded clinical acquisitions demonstrated substantially improved noise suppression and lesion preservation compared with conventional reconstruction. These results demonstrate the potential of diffusion-based reconstruction for quantitative ultra-low-count SPECT imaging and α-RPT dosimetry.

Abstract

Respiratory motion and attenuation mismatch degrade positron emission tomography (PET) reconstruction. Furthermore, the additional scan required for attenuation correction (AC) may complexify the imaging workflow and increase radiation exposure. Joint reconstruction of activity and attenuation with motion estimation (JRAAM) methods address these issues by jointly estimating activity, attenuation, and motion solely from emission data. We propose JRAAM-DPS, a diffusion model (DM)-based framework that leverages a diffusion prior on activity–attenuation image pairs to recover motion-compensated and attenuation-corrected PET images directly from respiratory-gated PET data. We evaluated the proposed method on simulated data, and demonstrated improved image quality compared to maximum likelihood activity and attenuation (MLAA) and the non-motion-corrected JRAA-DPS

Abstract

Flow-based generative models can be used as pre-trained priors for inverse problems; however, their use in PET image reconstruction remains largely unexplored. We present a proof-of-principle study using flow priors for 2D PET reconstruction. A U-Net was trained by flow matching on simple brain phantoms and used as an unconditional image prior. We used the MPC-Flow (Model Predictive Control) framework to condition the pre-trained flow via direct control of the generative dynamics, and adapted it to the PET setting by incorporating a Poisson likelihood and an explicit intensity scaling to address the dynamic range mismatch between the generative model and PET data. This method was used to reconstruct images from simulated low-count PET sinogram data. On held-out phantoms, MPC-Flow achieved lower error than diffusion-based reconstruction and conventional iterative baselines, likely due to the explicit control objective, which balances likelihood maximization with a penalty on deviations from the learned flow trajectory. These results demonstrate that flow priors can be coupled with PET likelihoods in a principled framework. Evaluation on more realistic anatomical phantoms and clinical data, where the learned manifold may be less accurate, will be a necessary next step.

Abstract

Generative models, especially diffusion models, have recently demonstrated strong potential for positron emission tomography (PET) image reconstruction. However, diffusion models often require many sampling steps and incur a high computational cost, which may limit their practical use in PET image reconstruction. Flow matching is an attractive alternative, enabling continuous generative modeling with a simpler sampling procedure and a stable, efficient training framework. In this study, we present flow matching-based PET image reconstruction methods. We first establish PET-FlowDPS as a baseline by incorporating Poisson likelihood guidance into the FlowDPS framework. We then propose a model-based PET image reconstruction method that uses a pretrained flow matching model as a prior. The proposed reconstruction method iteratively performs three steps: (1) flow-based destination estimation to obtain a clean image estimate, (2) penalized PET image reconstruction refinement to enforce data consistency, and (3) flow-based propagation to update the image to the next time step. The proposed reconstruction method was evaluated using 10 low-dose clinical brain [¹⁸F]FDG datasets. Experimental results showed that the flow matching-based methods reduced image noise while preserving cortical structures compared with the maximum a posteriori expectation-maximization (MAPEM) algorithm. The proposed method further improved image quality compared with PET-FlowDPS in terms of PSNR. In addition, the computation time of the proposed method was less than 2 minutes. These results demonstrate that flow matching is a promising generative prior for PET image reconstruction.

Abstract

Dynamic positron emission tomography (PET) reconstruction remains challenging due to low count statistics in individual frames and the ill-posed nature of the inverse problem. Kernelized expectation maximization (KEM) methods have been developed to incorporate prior information through a kernel matrix; however, their performance strongly depends on the design of the kernel and the prior information, as well as the associated hyperparameters. In this work, we propose a data-driven framework for kernel construction based on attention-guided learning from dynamic PET images.

A neural network is used to learn spatial-temporal relationships from dynamic image frames, producing attention maps that encode pairwise similarities. These features are further analyzed via dimensionality reduction and clustering to identify regions of interest (ROIs) with similar spatial-temporal characteristics. The resulting ROI masks are used to define adaptive neighborhoods and compute kernel weights, producing a data-driven kernel that is subsequently integrated into the KEM reconstruction framework.

Experiments on simulated data demonstrate that the proposed method not only improves image quality but also enables accurate estimation of time-activity curves (TACs) across different ROIs. In real PET data, the method achieves enhanced reconstruction quality, while exhibiting reduced sensitivity to the choice of clustering parameters compared with the baseline approach that performs clustering directly on the dynamic PET images without attention-guided learning. In particular, stable reconstruction performance is maintained across a wide range of cluster numbers, whereas baseline approaches show strong parameter dependence and introduce noticeable noise at larger cluster sizes.

These results demonstrate that attention-based kernel learning provides a robust and effective strategy for improving dynamic PET reconstruction quality and quantitative accuracy.