Bile Ducts And Hepatoliths Segmentation From Sequence Images Based On Correlation Mechanism

Hepatolithiasis is a common and frequently-occurring disease in surgery,which is the main reason of death from benign biliary tract diseases.Minimally invasive surgery for removing stones is one of its main treatment methods.The success of the operation largely depends on the manual identification and localization of the bile ducts and hepatoliths in CT images by experienced doctors before operation.However,this work is time-consuming and labor-intensive.Also,it demands that the doctor have rich clincal experience.Specifically,the evaluation of CT images are quite different due to different doctors with subjective experience.Therefore,automatic and accurate segmentation of bile ducts and hepatoliths from CT sequence images is significant for clinical application and research.Although there are some studies on segmentation of bile ducts and hepatoliths,they do not perform well due to three limitations.1.Bile ducts are highly deformable,and the hepatoliths are small in size and unevenly distributed;2.Bile ducts,hepatoliths and adjacent tissues are quite similar in grayscale,increasing the difficulty of segmentation;3.Hepatoliths often merges with a variety of complex lesions,which indicates that isolated cross-sectional CT two-dimensional medical images cannot show sufficient biliary stricture and stone location.To deal with these problems,a segmentation method for bile ducts and hepatoliths is proposed in this thesis based on correlation mechanism,which can provide the distribution of bile ducts and hepatoliths for the patients,and is helpful for stone removal operations.The main work and contributions of this thesis are described as following.First,to deal with the problem of orphaned cases in the existing hepatolithiasis image data et GZMU-HS,we have expanded the image dataset with more cases and labelled the adjacent slices of isolated two-dimensional images.This job is aided by highly qualified doctors in the Department of Hepatobiliary Surgery at the First Affiliated Hospital of Guangzhou Medical University.The expanded image dataset facilitates our subsequent study.Secondly,to overcome the above three limitations,a contextual information correlation mechanism for sequential images for hepatolithiasis,including the extraction of low-level multi-scale temporal context information,the extraction and fusion of high-level bidirectional temporal context information,and weighted distribution loss function indicating the correlations among CT slices.The meachanism makes full use of multi-scale contextual feature maps and high-level bidirectional contextual feature maps in an end-to-end full convolutional network,which contributes to good segmentation.The proposed segmentation method based correlation mechanism can better focus on contextual information between CT slices,learn contextual features between CT slices and fuse multi-scale features in the CT slices.Thus,rich multi-scale contextual features can be achieved for segmentation.In the training process,the contextual information between CT slices is more accurately involved due to the optimization of the weight distribution loss function for CT slice correlation.Experimental results show that the proposed method is superior to a variety of deep learning methods with a Dice coefficient of 91.793%,a recall rate of 99.424%,and an accuracy of 89.214% for bile ducts,and a Dice coefficient of 74.632%,a recall of 99.898%,and an accuracy of 60.564% for hepatoliths.