TAGNet: A transformer-based axial guided network for bile duct segmentation

Automatic segmentation of intrahepatic bile ducts and common bile ducts plays an essential role in interventional surgery for cholangiocarcinoma, directly related to the success rate of the operation. However, the large shape and appearance variances make it challenging to segment bile ducts, especially for 3D CT images. In this study, we propose a transformer-based axial guided network, dubbed TAGNet, to automatically segment intrahepatic and common bile ducts by exploiting intraand inter-slice context modeling. The pivot is to take advantage of CNN-transformer hybrid architecture to simultaneously explore local and global contextual information from multiple adjacent slices. Especially a novel slice-axial-attention transformer module is imposed at multi-scales concurrently to capture the intraand inter-slice feature representations along each direction, boosting long-distance contextual modeling while limiting the computation cost. Moreover, a slice guided consistency loss function is advanced to enforce anatomical prior consistency among adjacent slices in a semi-supervised manner, thus enhancing the spatial topology of bile duct segmentation. Extensive experimental results on an in-house bile ducts CT dataset demonstrate that our method is capable of achieving promising performance, which achieves at least a 4.5% improvement in Dice and a reduction of 1.5 in HD95 than other state-of-the-art methods, indicating its potential for automated intrahepatic and common bile duct segmentation. We have made our code publicly available via https://github.com/zephyrize/TAGNet.