Experimental Study: Scanner-level Inter-crystal Scatter Recovery for TOF-DOI Prototype Prism-PET Brain Scanner

Inter-crystal scatter (ICS) poses a significant limitation to the quantitative accuracy of reconstructed images in positron emission tomography (PET), as gamma photons scatter among scintillation crystals before complete absorption within the detector. This phenomenon leads to mispositioning of the line of response (LOR), thereby compromising spatial resolution, which is particularly exacerbated in high-resolution PET systems with small crystals. In this study, we introduce an ICS recovery framework that generates synthetic scatter SiPM patterns through simulated photon interactions and experimental photoelectric SiPM patterns. Two convolutional neural networks (CNNs) are trained using synthetic scatter patterns and ground truth DOI and crystal ID data from simulations for crystal ID and DOI recovery. The experimental hot spot data from the Prism-PET prototype brain scanner is utilized to evaluate the CNN models. Comparison of the reconstructed images without employing the ICS recovery model reveals lower background noise and higher spatial resolution.