Hepatic Vein Segmentation Via Sequence Context Association Mechanism

In recent years,the incidence of liver diseases is increasing due to ecological degradation and irregular lifestyle and diet.Surgical resection is commonly used for liver diseases with salient lesions,such as liver tumors,intrahepatic bile duct stones and liver cysts.To remove the liver lesion precisely and to avoid the hemorrhage due to incision into the vessels,it is necessary to implement the resection along the surrounding hepatic veins according to the Couinaud segmentation method.Thus,quantitative analysis and accurate modelling of the hepatic veins are required prior to surgery.Clinically,the hepatic veins in CT images are manually outlined by the surgeon,which is time-consuming and demands rich clinical experience.Therefore,it is clinical urgent to automatically and accurately segment the hepatic veins from medical images.Several issues should be addressed in the hepatic veins segmentation.（1）The vessel size is small and unevenly distributed.Thus,the existing deep learning networks easily ignore the subtle features of these fine vessels during multiple downsampling,resulting in poor segmentation for fine vessels.（2）The lesions around the vessels will make the distributions and morphologies of the vessels irregular.Isolated CT slices cannot fully show the complex alignment and location information of the vessels,resulting in poor continuity of the segmented vessels.To address the above issues,the main work of this thesis are summarized as follows.In this thesis,we propose a hepatic veins segmentation method based on the intra-slice attention mechanism,which uses encoding and decoding modules to obtain the vein features in different dimensions.These features involve a large amount of fine vein vessels details and semantic information.Meanwhile,we design an intra-slice attention module to fully integrate the high-dimensional semantic features and low-dimensional detail features of vein vessels.The network focuses on the hepatic veins features in regions and channels to improve the segmentation of fine hepatic veins.We also design the intra-slice and inter-slice attention mechanism and the intra-slice and inter-slice graph connection mechanism to enhance the correlation between adjacent CT slices,which can extract the sequence context information.The sequence context information can reveal the complex alignments,distributions and locations of hepatic veins,which is beneficial to improve the continuity of vein segmentation.In addition,this thesis proposes a loss function for the vein region combined with an edge constrained term to optimally adjust the vein edges.Experiments show that the two hepatic veins segmentation methods achieve good segmentation results.Specifically,the segmentation network with intra-slice and inter-slice information association is achieves the segmentation performance of 84.5% Dice,85.6% precision,86.6% sensitivity,and 86.1% F1,which outperforms several deep learning networks.