Online program Overview Session: M-LT-01

Abstract

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

Abstract

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

Abstract

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

Abstract

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

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

Abstract

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

Abstract

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

Abstract

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

Abstract

Joint Maximum Likelihood Activity and Attenuation (MLAA) provides a promising CT-less attenuation correction framework but suffers from global scaling ambiguity. Intrinsic Lu-176 background radiation provides patient-specific transmission information to reduce this issue, while introducing a highly nonlinear and ill-conditioned joint likelihood. We propose the Separable Cubic Surrogate with Trust Region (SCS-TR) algorithm, utilizing an analytical Lipschitz constant and a vectorized solver to ensure strictly monotonic convergence and avoid local optima. The reconstruction is initialized with Maximum Likelihood Transmission Reconstruction (MLTR) results. GATE simulations studies demonstrate that SCS-TR effectively recovers clear bone structures and smooth uniform regions without explicit regularization. The proposed method recovered clear bony structures and uniform soft-tissue regions without explicit regularization, and maintained structural fidelity under 20% count statistics.

Abstract

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

Abstract

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