Deep Learning-Based Limited Angle Tomography for a 1-Millimeter Resolution Dual-Panel Clinical PET System

We are developing a 1-mm resolution clinical positron emission tomography system dedicated to head-and-neck cancer imaging. Due to the dual-panel structure of the system, the angular sampling is limited. In this work, we propose to employ deep learning to facilitate image reconstruction using limited angle tomography for the system using the image-to-image translation neural network model known as "Pix2Pix". Inputs were created by running Monte-Carlo simulations with voxelized phantoms in two orthogonal projection angles of the opposing dual-panels, and target images were acquired by simulating the phantoms with an ideal cylinder scanner. The model was able to generate predicted images that highly match the target images. Using this model, the mean values of normalized root mean square errors (NRMSE) and structural similarity index measures (SSIM) of predicted full-angle images for ≥10 OSEM iterations improved by 47.0% and 18.2%, respectively, compared to the input images. The mean values of NRMSEs, SSIMs of region-of-interest, normalized contrast recovery ratio, and contrast-to-noise ratio for ≥10 iterations were improved by 43.4%, 19.8%, 46.7%, and 16.2%, respectively, compared to the input images.