SPARSE-VIEW MEDICAL TOMOSYNTHESIS VIA MIXED SCALE DENSE CONVOLUTIONAL FRAMELET NETWORKS

Abstract

X-ray tomosynthesis is a low-dose and relatively inexpensive 3D imaging technique that relies on limited-angle and sparse-view tomography. Unfortunately, tomosynthesis often leads to reconstructed images that are corrupted by ripple artifacts. The current state-of-the-art for artifact suppression in tomographic data involves the use of Convolutional Neural Networks for mapping corrupted reconstructions into artifact-free images. Recently, Deep Convolutional Framelet Networks (DCFNs) were proposed in which max-pooling layers in the U-net were replaced by fixed Wavelet decompositions. In this work, we show that replacing the regular convolutional blocks in the DCFNs by Mixed Scaled Dense (MSD) blocks for exploiting multi-scale features allows us to better represent and hence suppress tomosynthesis artifacts at different scales. Experiments using simulated data show that our Mixed Scale Dense Convolutional Framelet Network (MSDCFN) outperforms the state-of-the-art methods in the vast majority of the tomosynthesis scans evaluated.