Inter-crystal scatter recovery of a light-sharing depth-encoding Prism-PET prototype scanner using a diffusion model.

backgroundInter-crystal scattering (ICS) is a fundamental limitation in ultra-high resolution positron emission tomography (PET), particularly in light-sharing detectors with small crystal elements, where frequent Compton scatter across neighboring crystals obscures the initial interaction position and degrades the spatial resolution, image contrast, and quantitative accuracy. In our previous work, a convolutional neural network, Recovery-Net, was proposed and evaluated to recover the crystal location and depth-of-interaction (DOI) of the initial interaction in a light-sharing depth-encoding Prism-PET detector using Monte Carlo (MC) simulations. PURPOSE: This work aims to develop and experimentally validate a scanner-level ICS recovery framework for a light-sharing depth-encoding Prism-PET prototype scanner using a diffusion model.

methodsWe first developed a novel method to prepare the training dataset by combining scanner-level simulations with experimental measurements. The resulting dataset comprises an 8 × 8 array of silicon photomultiplier (SiPM) signals for each ICS event with associated ground-truth information, including crystal location, DOI, and deposited energy of the initial interaction. Using the SiPM signal patterns as input, we proposed a new diffusion-driven and physics-informed tripartite framework, named D3Recovery, to estimate the 3D position and deposited energy of the initial interaction. The previously proposed Recovery-Net was also implemented and evaluated for comparison. An ultra-micro Derenzo phantom and a 3D Hoffman brain phantom from the Prism-PET prototype scanner experiment were used to evaluate the impact of ICS recovery on image quality.

resultsRecovery-Net and D3Recovery achieved crystal identification accuracies of 64.1% and 70.1% for ICS events, respectively, with D3Recovery further improving DOI estimation to a full width at half maximum (FWHM) of 3.6 mm compared with 5.1 mm for Recovery-Net. Compared with center-of-gravity (CoG) positioning, D3Recovery reduced the mean crystal localization error from 2.18 mm to 0.99 mm and recovered the deposited energy with an error FWHM of 34.0 keV. For the ultra-micro Derenzo phantom images, Recovery-Net and D3Recovery increased peak-to-valley ratios by 13% and 24% for ICS events and by 14% and 17% for all events, respectively. For the Hoffman brain phantom, Recovery-Net and D3Recovery reduced mean absolute error (MAE) by 15.53% and 17.84% for ICS events and by 16.22% and 18.63% for all events, while increasing structural similarity index measure (SSIM) by 14.59% and 15.60% for ICS events and by 4.60% and 4.75% for all events, respectively. In addition, standard deviation (SD) measured in a uniform region decreased by 12.56% and 20.88% for ICS events, and by 5.26% and 8.48% for all events with Recovery-Net and D3Recovery, respectively.

conclusionsExperimental results demonstrate that D3Recovery significantly improves crystal localization, DOI estimation, and reconstructed image quality in high-resolution PET with small-crystal, light-sharing detectors, while providing a practical scanner-level training dataset acquisition strategy for experimental deployment of AI-based ICS recovery. By retaining a large fraction (∼70% in our case) of otherwise discarded coincidences, our results show that we achieve a favorable resolution-sensitivity balance that is most beneficial in count-limited, low-dose imaging.