Online program Overview Session: MIC-08

Abstract

Positron emission tomography (PET) requires dense detector crystals and associated electronics, resulting in high manufacturing costs. Sparse detector configurations can reduce hardware complexity but may degrade image quality due to reduced coincidence events, incomplete line-of-response coverage, and geometry-dependent artifacts. In this study, different sparse detector configurations were investigated for clinical PET imaging, and a sensitivity-informed network, termed SIU-Net, was proposed to recover image quality from sparse ring reconstructions. The residual sensitivity image was used as structural guidance to improve the restoration of degradation caused by detector sparsity. Experiments on 303 clinical brain datasets showed that the 1/2-TS-AR configuration achieved the best image quality among the evaluated sparse designs. SIU-Net consistently outperformed the baseline models across seven degraded image settings. These results demonstrate the potential of combining optimized sparse detector geometry with sensitivity-guided deep learning restoration for cost-effective PET imaging.

Abstract

The intrinsically limited spatial resolution of positron emission tomography (PET) confounds image quantitation. Post-reconstruction deblurring is an ill-posed inverse problem in which the blur kernel is often unknown, and noise renders direct inversion infeasible. We present a principled framework for self-supervised blind PET image deblurring that incorporates high-resolution MRI anatomical guidance in a data-consistent way. The central contribution is an observation-dependent spectral decomposition of the PET image into a practical range space, i.e., the set of spatial frequencies recoverable from the blurry PET observation at a specified signal-to-noise ratio (SNR) threshold, and a complementary practical null space whose content cannot be estimated from the PET data alone, regardless of the estimator. A self-supervised deep learning framework, Anatomy-Guided Null-Space Learning (AGNS), instantiates this decomposition with three jointly optimized components: a blind 3D kernel estimation network, a range-space Wiener deconvolution, and an MRI-guided null-space imputation network. Joint optimization proceeds via the alternating direction method of multipliers (ADMM). Methodological evaluation is performed through clinical paired-resolution validation using images from the HRRT and NeuroEXPLORER (NX) scanners, where the NX images are only used as the validation ground-truth. AGNS produces strong qualitative and quantitative benchmarks relative to competing unsupervised and non-AI approaches.

Abstract

PET imaging enables quantitative assessment of disease burden and treatment response, but performance is constrained by a trade-off between radiation dose and image quality. Reducing injected adiotracer activity decreases patient exposure, yet increases image noise, degrading lesion detectability and quantitative accuracy. Total-body PET systems, with substantially increased sensitivity, enable significant dose reduction. However, ultra-low-dose acquisitions remain highly noisy, limiting their clinical utility. In this work, we investigate Brownian Bridge Diffusion Models (BBDM) as a novel framework for ultra-low-dose PET denoising. BBDM formulates dose-level image translation as a stochastic bridge between paired standard- and low-dose PET images, with the diffusion process constrained at both endpoints. This formulation is well-suited for dose translation. The proposed method was evaluated on a multi-center, multi-scanner total-body 18F-FDG PET/CT dataset of 80 patients. Paired full-dose and simulated 1/10-dose datasets were generated through statistical count reduction in the list-mode data. The model was trained with T = 1000 diffusion steps, and inference was accelerated using S = 200 sampling steps. On the held-out test cohort, BBDM achieved substantial gains across both global image-quality and lesion-level quantitative metrics. Compared with low-dose inputs, BBDM increased PSNR by approximately 26.6% and SSIM by 8.9%, while reducing RMSE by 60.0%. Across evaluated lesions, SUVmean and SUVmax biases decreased by 54.5% and 70.9%, respectively. Lesion noise was also reduced by 45.8%. The approach effectively preserves both image quality and lesion quantification at substantially reduced dose levels, highlighting its potential for ultra-low-dose total-body PET imaging.

Abstract

Low-dose or fast-scan single-photon emission computed tomography/computed tomography (SPECT/CT) suffers from Poisson noise, but clinical archives usually provide only one noisy standard-dose projection per patient and no noise-free target. A common single-scan workaround synthesizes a low-dose input from the standard-dose scan and uses the original projection as the supervised target. This creates a nested pair because the target contains the input photons, so the objective can learn a noisy full-dose image with input-noise leakage. We propose a self-supervised projection-domain framework based on mutually exclusive Poisson thinning. Each standard-dose projection is partitioned online into K∈{2,3,4,5} disjoint equal-dose subsets by multinomial assignment, and two subsets are randomly paired for Noise2Noise training. These paired subsets contain disjoint photons and are conditionally independent given the underlying count rate while preserving the same dose; mean-square or Poisson-likelihood training therefore targets the clean same-dose conditional mean rather than a noisy full-dose target. A three-dimensional residual U-Net is trained only from thinned noisy pairs and applied before OSEM reconstruction. Training used 225 clinical bone SPECT/CT cases. On 31 held-out cases evaluated by patient-specific controlled Poisson resampling over 1×--5× fast-scan factors, the method achieved 45.72 dB PSNR and 0.9862 SSIM in the projection domain, outperforming an architecture-matched low-to-standard-dose supervised baseline by 2.68 dB; the paired gain increased from 1.82 dB at 1× to 3.34 dB at 5×. The framework enables low-dose bone SPECT/CT denoising from routinely stored projections without clean ground truth or repeated acquisitions.

Abstract

Accurate attenuation and scatter correction is important for positron emission tomography (PET) to improve diagnosis. Conventional PET correction pipelines rely heavily on CT-derived attenuation information, which inevitably increases radiation exposure during clinical examinations, especially in total-body PET/CT imaging. In this study, we introduce a vision-language-model-guided conditional flow framework for CT-free PET attenuation and scatter correction. Specifically, non-corrected PET slices are analyzed to generate organ-aware textual prompts, which are subsequently encoded as semantic representations and integrated into a latent-space flow generation process. Then, a cross-modal attention mechanism is further designed to establish interactions between semantic prompts and PET image features under multiple scales during conditional flow matching. Experiments were performed on 264 clinical subjects under total-body PET/CT scans. Comparison methods consist of several representative deep learning approaches, including UNet, Uformer, CycleGAN, latent diffusion models, and VM-UNet. Quantitative evaluation demonstrated that the proposed framework achieved the highest metric index, reaching 38.66 dB PSNR and 0.9754 SSIM on the testing dataset. Visual assessment and Correlation Analysis additionally confirmed improved anatomical continuity and more accurate tracer uptake recovery. The study suggests that semantic-aware flow modeling offers an effective strategy for CT-free PET correction.

Abstract

Respiratory motion during acquisition degrades image quality in positron emission tomography (PET). Data-driven gating (DDG) extracts a respiratory signal directly from the measurement data, but classical methods such as axial center-of-mass and principal component analysis (PCA) degrade under irregular breathing and at low count rates. We propose a self-supervised framework based on implicit neural representations (INRs) that jointly estimates a respiratory gating signal and a motion-corrected reconstruction. Two coordinate-based networks parametrize a motion-free canonical activity image and a deformation field conditioned on learnable per-frame latent codes that encode the respiratory state. We evaluate the framework on XCAT thoracic phantom simulations of an 18F-FDG acquisition at clinically realistic count levels under regular and irregular breathing. The recovered gating signal achieves Spearman correlations of 0.87 (regular) and 0.85 (irregular) with the ground-truth respiratory signal, exceeding center-of-mass (0.62; 0.30) and PCA-based (0.74; 0.34) DDG baselines, and the framework simultaneously produces a motion-resolved reconstruction with PSNR comparable to maximum likelihood expectation maximization (MLEM) reconstruction.

Abstract

Long axial field-of-view (LAFOV) PET enables imaging at very late timepoints but requires repeated CT acquisitions for attenuation correction (AC) in multi-time-point protocols, increasing radiation exposure. This study evaluates two CT-free AC strategies for [¹⁸F]PSMA-1007 LAFOV PET at delayed imaging timepoints: synthetic μ-maps derived from non-attenuation-corrected PET imaging using a ResNet-based model (Synth-AC) and elastic co-registration of CT-based μ-maps from the initial scan (CoReg-AC). Ten patients were scanned on a Siemens Biograph Vision Quadra LAFOV system at 90 min, 3 h, and 5 h p.i. Quantitative evaluation was performed based on relative differences in SUV_mean across seven major organs. Lesion quantification was assessed for SUV_mean (28 lesions at 3 h p.i., 20 lesions at 5 h p.i.) using Bland–Altman analysis and linear regression. Visually, synthetic μ-maps appeared smoother, while co-registration preserved more anatomical detail but showed localized inconsistencies. In PET images, Synth-AC exhibited smaller and more homogeneous deviations (±5%), whereas CoReg-AC showed more pronounced peripheral differences. At organ level, reduced variability was observed at 5 h for both methods. CoReg-AC showed a positive bias in high-uptake organs (liver +1.7% to +4.3%, spleen +2.7% to +4.3%), while Synth-AC shifted from near agreement at 3 h to moderate negative bias at 5 h (liver −2.1%, heart −13.5%). Both methods demonstrated a pronounced positive bias in lung regions. Linear regression demonstrated strong agreement with CT-based AC (R² = 0.993–0.998), and Bland–Altman analysis at 3 h p.i. showed a mean bias of −0.97% with limits of agreement [−12.11%; +10.17%] for Synth-AC and +1.40% [−8.71%; +11.52%] for CoReg-AC. Both CT-free approaches enable robust quantitative imaging in delayed LAFOV PET. Observed differences are primarily driven by organ-level bias patterns, while lesion quantification remains stable.

Abstract

Full-volume 3-D medical image generation at clinical resolutions remains a significant challenge due to GPU memory constraints. Direct voxel-space generative modeling becomes increasingly difficult once spatial resolution exceeds 256³, especially on single-GPU hardware. Latent diffusion approaches reduce this burden by performing generation in a compact representation [1], but decoding onto a dense fixed grid remains costly at high resolution. Implicit neural representations offer an attractive alternative, as they reconstruct images by querying continuous spatial coordinates rather than materializing a full voxel grid [2], making them naturally suited to memory-efficient volumetric generation.
We propose a memory-efficient 3-D latent implicit diffusion framework that combines a lightweight 3-D encoder with a coordinate-conditioned local implicit decoder, inspired by [3]. The encoder compresses the input volume into a compact latent grid, while the implicit decoder reconstructs the volume by querying continuous spatial coordinates from learned latent features, avoiding the need to materialize the full voxel grid during training. A full-volume 3-D diffusion model is then trained directly in this compact latent space. At inference, new volumes are generated by sampling from the diffusion model and rendering the full grid through the implicit decoder.
We validate our framework on CT and MRI volumes at resolutions up to 512³ on a single consumer-grade GPU, demonstrating coherent full-volume generation across axial, coronal, and sagittal planes without materializing the full voxel grid at any stage of training.