Octahedron-shaped Convolution for Refining Aorta Semantic Segmentation

EasyChair Preprint 5965

Abstract

Refining 3D aorta segmentation is essential for clinical aorta analysis. The small tubular diameter of the aorta branches and the discontinuity of neighbouring information make it difficult to get a continuous semantic segmentation map. In this paper, we proposed a novel adaptive octahedronshaped convolution (AOSC) based on VNet and signed distance map(SDM). AOSC aimed to aggregate more contextual information for each sample point in the aortic branches with smaller tubular diameters. The weights of feature fusion introduced SDM as auxiliary information to measure the similarity of neighbouring points. Furthermore, we embedded the learned 3D offset field into AOSC to avoid inaccurate segmentation on the region around the narrow tubular structures. The AOSC module prolonged the predicted length of small aorta branches and then improved the tubular continuity of the aorta segmentation map. We evaluated the AOSC module on our-collected dataset and MICCAI ASOCA2020 coronary artery dataset. Our method achieved the state-of-the-art results in terms of Dice and Jaccard metrics.

Keyphrases: Aorta Segmentation, aorta branches, contextual information, tubular continuity, tubular diameter

@booklet{EasyChair:5965,
  author    = {Xi Xiang and Gongning Luo and Pengfei Zhao and Wei Wang and Kuanquan Wang},
  title     = {Octahedron-shaped Convolution for Refining Aorta Semantic Segmentation},
  howpublished = {EasyChair Preprint 5965},
  year      = {EasyChair, 2021}}