3D Image Segmentation For Hepatobiliary Based On Multi-scale Attention Mechanism

Hepatolithiasis is a common disease in the Department of hepatobiliary surgery,which is one of the main causes of death from non-neoplastic biliary tract diseases in China.Stone removal therapy for hepatolithiasis can employ minimally invasive surgical procedures,such as percutaneous transhepatic cholangioscopic technique.The surgery needs to identify the location of the stone in the hepatic bile duct,which commonly relies on preoperative imaging analysis of a patient’s abdominal contrast-enhanced CT scans.Because of the close proximity of the hepatic bile ducts and intrahepatic vessels,it is easy to incorrectly puncture into the intrahepatic vessels,which causes massive hemobilia.Currently,the locations of bile ducts and stones in abdominal CT images are subjectively determined by the radiologists,which is potentially influenced by their experience.Therefore,it is necessary to explore an automated image segmentation algorithm for hepatic bile ducts in CT images.The main contents of this thesis are described as follows.We construct a image dataset in collaboration with the First Affiliated Hospital of Guangzhou Medical University,which involves of the imags of portal venous phase of enhanced CT for hepatolithiasis.The image data are annotated by the tool of ITK-SNAP.Thus,3744 slice images from 21 cases are selected for training and 891 slice images from 5cases for testing.In this thesis,we propose a multi-scale attention based segmentation algorithm with residual convolution.It utilizes the encoding decoding network structure to automatically mine the intrinsic relationship of the features of the bile ducts.To integrate the shapes and edges of bile ducts,we design a multi-scale attention module based on the transformer for hop connection to integrate the features from different feature layers,which can learn their relationships from each each and optimize the segmentation results of hepaticbile ducts.The feature extraction network is further modified in this thesis,in which grouping convolution and deep convolution are employed to produce larger receptive fields to make the network structure more flexible.Upsampling and downsampling layers are decoupled from the feature learning module.The multiscale attention modules are modified to efficiently deal with the features of hepatic bile ducts from three different feature layers,which can deal with the split nature of feature fusion across other layer.In order to focus on the edge features of hepatic bile ducts,an cross entropy loss function with overlapping degree is designed to focus the edges,which is beneficial to let the network pay attention to the edge features of bile ducts to effectively improve the accuracy of segmentation.Experimental results show that both the proposed segmentation methods achieve good results for hepatic bile ducts.The self-attention segmentation network with the multiscale fusion strategy achieves 76.35% Dice coefficient（DSC）and 6.67 Hausdorf distance（HD）,which is superior to the existing networks.