Experimental evaluation of convolutional neural network-based inter-crystal scattering recovery for high-resolution PET detectors.

. One major limiting factor for achieving high resolution of positron emission tomography (PET) is a Compton scattering of the photon within the crystal, also known as inter-crystal scattering (ICS). We proposed and evaluated a convolutional neural network (CNN) named ICS-Net to recover ICS in light-sharing detectors for real implementations preceded by simulations. ICS-Net was designed to estimate the first-interacted row or column individually from the 8 × 8 photosensor amplitudes.. We tested 8 × 8, 12 × 12, and 21 × 21 LuSiOarrays with pitches of 3.2, 2.1, and 1.2 mm, respectively. We first performed simulations to measure the accuracies and error distances, comparing the results to previously studied pencil-beam-based CNN to investigate the rationality of implementing fan-beam-based ICS-Net. For experimental implementation, the training dataset was prepared by obtaining coincidences between the targeted row or column of the detector and a slab crystal on a reference detector. ICS-Net was applied to the detector pair measurements with moving a point source from the edge to center using automated stage to evaluate their intrinsic resolutions. We finally assessed the spatial resolution of the PET ring.. The simulation results showed that ICS-Net improved the accuracy compared with the case without recovery, reducing the error distance. ICS-Net outperformed a pencil-beam CNN, which provided a rationale to implement a simplified fan-beam irradiation. With the experimentally trained ICS-Net, the degree of improvements in intrinsic resolutions were 20%, 31%, and 62% for the 8 × 8, 12 × 12, and 21 × 21 arrays, respectively. The impact was also shown in the ring acquisitions, achieving improvements of 11%-46%, 33%-50%, and 47%-64% (values differed from the radial offset) in volume resolutions of 8 × 8, 12 × 12, and 21 × 21 arrays, respectively.. The experimental results demonstrate that ICS-Net can effectively improve the image quality of high-resolution PET using a small crystal pitch, requiring a simplified setup for training dataset acquisition.