Scattered-photon imaging in PET — An initial estimate for improved MLEM reconstruction

Abstract We proposed a novel PET image reconstruction strategy that utilizes scattered photon timing information, as a prior to improve image quality in low-count scenarios. Single-scattered coincidence data is typically discarded in conventional reconstructions. But using GATE simulations of a cylindrical PET scanner, we extracted single-scattered (SS) coincidence events and created a Scattered Activity Map (SAM) without relying on energy information. These SAMs serve as informed initial estimates for Maximum Likelihood Expectation Maximization (MLEM) reconstruction. To mitigate the blurring and noise inherent in SAMs, we applied Lucy-Richardson (LR) deconvolution using a system-derived point spread function (PSF), providing a deconvolved SAM (DSAM). Reconstruction of SS events were performed in the Customizable and Advanced Software for Tomographic Reconstruction (CASToR) framework across three scanner configurations (with different timing resolution) — Ideal (1 ps), Future (100 ps), and State-of-the-art (200 ps). We demonstrated that both SAM and DSAM initializations accelerate convergence and improve contrast recovery coefficient (CRC), contrast-to-noise ratio (CNR) over standard uniform initialization. The benefits are most significant in low-statistics datasets, with DSAM offering consistent performance advantage. This study establishes the feasibility of incorporating SS-derived scatter maps into the reconstruction process for TOF-PET systems with limited or no energy discrimination, offering a pipeline for enhancing image quality in low-dose or time-constrained PET imaging.