Super-resolution PET Brain Imaging using Deep Learning

Positron emission tomography (PET) is a noninvasive and reproducible way of medical diagnostic technology. However, due to its intrinsic imaging characteristics, the PET image resolution is inferior compared to other medical imaging modalities. PET image quality and quantitative accuracy are degraded due to its poor resolution. In this study, we present a super-resolution (SR) approach for PET images based on a deep learning network. The network was trained on images from the high-resolution research tomograph (HRRT) scanner and its blurred low-resolution (LR) counterparts matching the clinical mCT scanner resolution. Data augmentation methods were also applied to improve the training data’s generalizability. After the training, the model was validated using the blurred LR HRRT images. Also, the proposed network was applied to enhance the resolution of routine clinical brain images acquired on an mCT scanner. The results reveal that the proposed approach leads to a substantial improvement in PET image resolution and overall quality.