Attention-Guided Decoder in Dilated Residual Network for Accurate Aortic Valve Segmentation in 3D CT Scans

摘要

Automatic aortic valve segmentation in cardiac CT scans is of high significance for surgeons' diagnosis on aortic valve disease and planning of aortic valve-sparing surgery. However, the very fast flapping speed, ambiguous shapes and extremely thin structures of the aortic valve lead to great difficulties in developing automatic segmentation algorithms. In this paper, we proposed an end-to-end deep learning method to address the problem of segmentation of the aortic valve from cardiac CT scans. Our method uses 3D voxel-wise dilated residual network (DRN) as backbone network and we equip it with novel attention-guided decoder modules to suppress non-valve artifacts and noise and pay attention on the fine leaflets in order to acquire accurate valve segmentation results. We conducted qualitative and quantitative analysis to compare with state-of-the-art (SOTA) 3D medical image segmentation models. Experiment results corroborate that the proposed method has very high competence.