SCU-Net: A Shape-Supervised Contextual-Fusion U-Net for the Dilated Biliary Tree Segmentation

Segmentation of the dilated biliary tree in abdominal CT is very important for the diagnosis and later treatment of biliary diseases. However, heterogeneity, abrupt deformation, and blurred boundary of the dilated biliary tree pose a challenge for current deep learning based methods, which have not been sufficiently studied. This work proposes a shape-supervised contextual-fusion U-Net (SCU-Net) to address this. Specifically, the network adds to the classic U-Net a novel feature fusion module to fuse the coarse-to-fine information while supervising segmentation integrity by the shape-aware distance map. Experiments showed that our model outperforms existing segmentation algorithms, obtaining a dice score of 76.4%, which was 3% higher compared with the advanced segmentation method nnU-Net.