Online program Overview Session: MIC-04

Abstract

Accurate attenuation correction (AC) is essential for quantitative PET imaging, but CT-free AC across diverse radiotracers remains challenging. MLAA-derived μ-maps are useful inputs for deep learning-based AC. However, tracer-dependent biodistribution introduces domain shifts that limit the generalizability of models trained only on ¹⁸F-FDG. We propose a single-source cross-tracer CT-free AC framework that combines tracer-robust U-Net regression with anchored flow matching. Local Gamma Field augmentation is used during U-Net training to mimic anatomy-preserving regional intensity and contrast variations in MLAA μ-maps, encouraging tracer-independent MLAA-to-CT mapping. The U-Net output is then used as an anchor for an unconditional flow-matching model trained on CT μ-maps, enabling generative refinement from a reliable deterministic estimate. The proposed method achieved the highest PSNR across all evaluated tracers and reduced MLAA noise, tracer-specific artifacts, and U-Net over-smoothing. AC PET analysis showed strong agreement with CT-based AC PET, with regional uptake correlations closest to the identity line. These results demonstrate that the proposed framework enables accurate and structurally faithful μ-map estimation across multiple tracers using only FDG training data.

Abstract

Multi-aperture single-photon emission computed tomography (SPECT) at ¹³¹I (364&nbsp;keV) exhibits a detector-wide plate-penetration halo: a 1.8&nbsp;cm tungsten plate transmits ~4 &times; 10^&minus;4 at &mu;_W = 4.4&nbsp;cm^&minus;1. Prior corrections perform the aperture/plate split at the photon level via effective-radius point-spread function enlargement, or at the system-matrix level via per-ray full-physics tracing; both merge the two components at storage time, inflating memory by one to two orders of magnitude. We propose a <em>storage-agnostic</em> per-ray partition: a topological rule on the source-to-detector intersection list separates the response into a sparse, peaked aperture component <em>A</em> and a slowly varying plate halo <em>B</em>, upstream of any representation choice. The rule is mutually exclusive and jointly exhaustive, so <em>A</em>_ij + <em>B</em>_ij = <em>C</em>_ij holds bit-exactly against the unsplit baseline for any per-ray contribution functional; the plate path length is a closed-form slab chord minus an analytic aperture chord at O(<em>N</em>_ap) per ray. On a 34-module multi-aperture SPECT ring (small-aperture configuration), the identity holds bit-exactly across three projection modes (collimator-path, sensitivity, counts) for a centered single-photon source and across every pixel of 34 Jaszczak hot-rod ¹³¹I views, with per-pixel plate intensity ~10^&minus;4&ndash;10^&minus;3 of the aperture peak. OSEM correction with a lagged plate estimate will be presented at the conference.

Abstract

Lutetium background radiation in modern PET scanners enables activity- and tracer-independent transmission measurement for attenuation correction (AC) without CT scans. However, quantitative accuracy of attenuation maps (µ-maps) derived from lutetium transmission (Lu-Tx) is limited by low transmission statistics and noisy randoms estimation from delayed coincidences (DE). In this work, we investigate the use of singles-derived randoms (RS) to improve Lu-Tx-based AC for CT-less PET. Emission and Lu-Tx data were acquired simultaneously on the PennPET Explorer, a 142-cm long axial field-of-view (LAFOV) scanner. µ-maps were reconstructed using maximum-likelihood transmission reconstruction with randoms estimated from either delayed coincidences or singles data, then used for attenuation and scatter correction during PET image reconstruction. The methods were evaluated using physical phantoms, including uniform cylinders, a 190-cm pipe phantom, cylinders with non-uniform axial activity distributions, and an image quality (IQ) phantom. Data were subsampled to simulate low-count short clinical scans. RS produced substantially smoother sinograms while maintaining agreement with DE in total randoms counts. Incorporation of RS improved axial uniformity, reduced noise, and mitigated emission contamination-induced bias in µ-maps, particularly under low-count conditions and non-uniform activity distributions. In IQ phantom studies, AC PET images reconstructed using RS demonstrated contrast comparable to CT-AC PET, whereas DE led to reduced contrast and increased bias, especially for short scans. These results demonstrate that RS improves the quantitative accuracy of Lu-Tx-based AC and support the proposed method as a practical pathway towards CT-less PET for both long dynamic imaging and short clinical scans.

Abstract

Correction of random coincidences is required in clinical PET imaging. Several methods have been proposed to estimate their distributions in a PET acquisition and they mostly use singles detection and/or a delayed coincidence window (DCW). We propose a new approach to achieve this task, which uses the differences in arrival times between consecutive single events that are greater than the coincidence window. This approach is thus less susceptible to bias induced by true coincidences and has more statistics than a DCW. As a proof of concept, a first implementation of this approach was compared to the classic methods using simulations from a small-sized PET scanner. The acquisition of four phantoms, each with different challenges in terms of randoms estimation, were simulated over a range of activities. Overall, the proposed method was less biased than the classical methods while also being more precise than DCW.

Abstract

The Alpha-SPECT Mini system integrates SPECT imaging with CT-based attenuation correction, providing a platform for quantifying in vivo distributions of alpha-emitting radiopharmaceuticals and daughter radionuclides. In this study, three Ac-225 filled syringes with different volumes, 0.5, 4.2, and 18.0 mL, respectively, were scanned with Alpha-SPECT Mini. The ground truth activities were[YD1] 26.12 µCi, 21.62 µCi, and 18.80 µCi from the smallest to the largest syringe, respectively. SPECT listmode acquisition was performed for 20 minutes for each syringe, together with an 80 kV CT scan for attenuation correction. SPECT images were reconstructed for the following photopeaks: 100 keV for Ac-225, 218 keV for Fr-221, 440 keV for Bi-213, and 117 keV for Tl-209. Activities of Ac-225 and these daughter radionuclides, were then estimated from SPECT images using CT defined ROIs and compared with ground truth.

The results showed that quantification accuracy varied with the radionuclides and syringe sizes. Fr-221 and Bi-213 showed relatively lower errors (<5%), while Ac-225 and Tl-209 exhibited larger deviations. The higher errors likely due to the low abundance of low energy (100 keV for Ac-225 and 117 keV for Tl-209) photopeaks and interference from downscatter and collimator penetration from high energy photons. Tl-209 also exhibited larger deviations with a size-dependent trend, suggesting sensitivity to source geometry and scatter correction. Overall, the Alpha-SPECT Mini system showed potential for quantitative imaging of Ac-225 and daughter radionuclides. Future work includes further calibration and implementation of scatter correction.

Abstract

Radiomics have demonstrated potential in numerous clinical applications. However, feature variability resulting from different scanner hardware and image processing degrades predictive ability and reduces model generalizability. We present a novel physics-driven framework for radiomics standardization. Assuming a linear, shift-invariant image formation model, we developed an analytical method to remove the dependency of radiomics features classes on known spatial resolution and noise properties. The method consists of two deconvolution operations: deconvolving the system blur from the reconstructed image, then deconvolving the noise probability density function from radiomics feature distributions. This approach is applicable to multiple feature classes. We further developed a workflow to implement this framework on a clinical CT scanner. The method was evaluated on an anthropomorphic lung phantom imaged at five dose levels and reconstructed using eight kernels. Results demonstrate that the standardization method effectively improves the consistency of radiomics features value with those at reference imaging conditions. As expected, standardizing from smoother kernels to sharp kernels is more challenging than the reverse. The proposed standardization method is a promising tool for improving the robustness and generalizability of radiomics models.

Abstract

¹²³I-DaTSCAN is used to aid diagnosis of Parkinsonism. Visual analysis of ¹²³I-DaTSCAN is used in conjunction with patient clinical history and semi-quantification. Specific Binding Ratios (SBR) are commonly estimated for the right and left putamen and caudate using commercially available software. The estimation of uncertainties is not common for this clinical study; however, the benefits of estimating uncertainties have been recognised for other clinical applications. In this work, a method to evaluate the uncertainty associated with SBR is developed using phantom data and then applied to estimate uncertainties for imaging data from a cohort of patients. List mode data was acquired and bootstrapped to create equivalent sinograms. The sinograms were reconstructed and processed using BRASS software to obtain the distribution of SBR and respective standard deviations were estimated. The impact of estimating uncertainties in patient classification is also considered.
For the patient data, the estimated standard uncertainties varied between 2.8 % and 19 %. The correlation of SBR uncertainty with patient classification within the cohort is presented. Initial analysis suggests the patient classification does not change when considering uncertainty, but does provide confidence in the results. 
This is the first time that uncertainties on SBR values are used to evaluate impact on patient classification. Future work will include analysing additional patient data to further investigate the impact of uncertainties in patient classification. Analysis to understand if the uncertainties can be reduced is also planned.

Abstract

Absolute resolution characterization of OSEM-reconstructed images is challenging because the OSEM algorithm is nonlinear and its resolution strongly depends on the activity distribution and the number of iterations. In contrast, the linear FBP algorithm provides a more consistent resolution, albeit with higher noise. In this work, we propose using FBP-reconstructed images as a bridge to estimate the absolute resolution of OSEM images without requiring ground truth. Two FBP-bridge approaches are investigated: an image-domain Gaussian fitting method and a frequency-domain modulation transfer function (MTF) ratio method. Hoffman phantom experimental results show that the Gaussian fitting approach has better agreement with the reference FWHM obtained using the known phantom ground truth. Both methods are further applied to human brain PET scans. This method also allows pixel-wise visualization of spatially variant image resolution. Future work will investigate the effects of noise and FBP filter selection.