DualRes-UNet: Limited Angle Artifact Reduction for Computed Tomography

Abstract

In some special cases of computed tomography (CT), limited angle problem occurs due to the limitation of imaging configuration and system design, which could cause severe artifacts in reconstructed CT images. In this work, we explore a Convolutional Neural Network (CNN) architecture for limited angle artifact reduction in CT imaging, named as DualRes-UNet. In our network, we adopt continuous down-sampling layers similar to U-Net to obtain a large receptive field view so that it can capture high level structure of object. Like concatenation operations, the proposed DualRes modules introduce high resolution features into the up-sampling process, which are beneficial to preserve more details and textures in final images from CNN. Preliminary experiments of 120~150 degrees of limited-angle CT were conducted, which demonstrate that our DualRes-UNet can effectively suppress the limited angle artifact and improve the quantitative accuracy of CT images.