AFF-NET: An Adaptive Feature Fusion Network For Liver Vessel Segmentation From CT Images

Accurate liver vessel segmentation from CT images is essential in computer aided diagnosis and surgery. However, due to the complex structures of liver vessels, it is difficult to extract small vessels and edge vessels from the images. Therefore, we propose an adaptive feature fusion network (AFF-Net) to accurately segment vessels from liver CT images. The AFF-Net contains three novel components: 1) An adaptive feature connection (AFC) module is designed to suppress image background noise to accurately extract small vessels; 2) An enhanced auxiliary (EA) module is proposed to fully utilize the topological information of vessels to improve the segmentation integrity; 3) A global information supervision (GIS) module is introduced to extract liver edge features to improve edge vessel segmentation accuracy. Experiments on public datasets show that our method achieves the Dice score of 0.72 and the sensitivity score of 0.73, showing much higher accuracy than related methods.