Flex-DD: Deep denoising model with flexible priors for sparse-view CT reconstruction

Abstract

Sparse-view Computed Tomography (SVCT) can effectively reduce radiation risk and improve scan-imaging speed. However, the severe streak artifacts will degrade the reconstruction results. Traditional iterative reconstruction methods rely on appropriate prior knowledge for achieving satisfactory results, while supervised deep learning techniques require large-scale paired training data that is challenging in practical CT application. In this paper, we propose a novel deep denoising model for SVCT reconstruction, which can jointly exploit deep prior and flexible hand-crafted prior. Specifically, we develop an ADMM algorithm for the optimization of Flex-DD. Moreover, we introduce a novel mechanism for flexible incorporation of Flex-DD model into the SVCT reconstruction task via HQS optimization framework, which significantly improves the reconstruction performance with good convergence. Extensive experiments on both simulated ellipses images and human CT images have demonstrated that our proposed method can achieves promising results in both qualitative and visual evaluations compared to popular state-of-the-art methods.