A Lightweight Dual-Domain Attention Framework for Sparse-View CT Reconstruction
Chang Sun, Ken Deng, Yitong Liu, Hongwen Yang
Abstract
Computed tomography (CT) plays an essential role in clinical diagnosis. Sparse sampling is an effective way to reduce the radiation on patients, but it will lead to severe artifacts on the reconstructed CT images, and artifacts may interfere with the diagnosis. Therefore, it is significant to ensure the quality of the region of interest, with extra parameters as few as possible. In this paper, we design a novel lightweight network called CAGAN and propose a dual-domain reconstruction pipeline for parallel beam sparse-view CT. CAGAN is an adversarial auto-encoder with the coordinate attention (CA), which preserves the spatial information of features and learns to focus on the textures. Also, the application of lightweight convolutional blocks reduces the parameters by a quarter without performance loss. Taking 180 views as full-view, our model reaches a PSNR of 41.975 given 45 views and a PSNR of 34.546 given 10 views, which is extremely sparse.