Effects of Angular Sampling in Deep-Learning based PET Reconstruction for Total-Body PET

In recent years, deep-learning (DL) has shown significant promise in positron emission tomography (PET) image reconstruction. Initially, the use of single view histo-images as input for end-to-end DL reconstruction offered easier handling by convolutional neural networks. Thereafter, the use of multi-view histo-images in DL-based PET reconstructions demonstrated better performance. Especially for Total-Body PET or dynamic imaging, DL-based reconstruction is useful in handling large data sizes since it decreases reconstruction speed when compared to classical tomographic approaches. For example, the PennPET Explorer data is sampled to a total of 180 views ie., 20 azimuthal-angular and 9 polar-angular samples. In this work, we investigate the effects of angular sampling strategies of PET data in DL based reconstruction approaches, which are quite different from the classical tomographic approaches following angular sampling requirements.