Super-Resolution PET Using A Very Deep Convolutional Neural Network

The quantitative accuracy of PET is degraded by partial volume effects caused by the limited spatial resolution capabilities of PET scanners. In this paper, we present a resolution recovery technique for PET based on a very deep super-resolution convolutional neural network (VDSRCNN). Our network incorporates high-resolution anatomical information to assist the deblurring process. It introduces the spatial location information of the input image patches to accommodate the spatially-variant nature of the blur kernels in PET. To validate this method, we have performed realistic simulations using the BrainWeb digital phantom. We compared our results with existing PET image deblurring approaches based on a total variation penalty function and an anatomical joint entropy prior function. Our results show that VDSRCNN leads to superior performance relative to the two reference approaches both qualitatively (e.g. edge recovery) and quantitatively (as indicated by two metrics: peak signal-to-noise-ratio and structural similarity index).