The Deep Steerable Convolutional Framelet Network for Suppressing Directional Artifacts in X-ray Tomosynthesis

Tomographic artifacts are unwanted distortions and/or structures not present in the scanned body that may appear in the reconstructed images. Recent deep learning-based methods for suppressing artifacts in tomographic images are currently not informed by the nature of these artifacts in the design of their network architectures. In this work, we present the Deep Steerable Convolutional Framelet Network (DSCFN), inspired by the theory of deep convolutional framelets, which exploits the regular pattern of ripple artifacts presented in sparse-view X-ray tomosynthesis images. Experiments with simulated data show that the DSCFN outperforms the regular U-net and its deep convolutional framelet, the tight U-net, in terms of PSNR.