Research On Image Segmentation For Bile Ducts And Hepatolith Based On U-Net

Hepatobiliary is one of the most common surgical diseases in the world,especially in Asia.Every year,the second operation caused by unclean hepatolith not only increases the trauma and pain of the patient,but also causes a certain economic loss to the society.In order to improve the safety of surgery,complete and correct preoperative diagnosis is essential.Hepatobiliary surgeons rely on clinical experience to locate lesions and hepatolith in CT images of the abdomen,which is greatly affected by subjective factors.Therefore,automatic segmentation of bile ducts and hepatolith in CT images of the abdomen helps doctors quickly and accurately locate focal areas.This task is very challenging due to the high deformability of hepatobiliary ducts and the small size of hepatolith.To the best of our knowledge,there is no studies on image segmentation for bile ducts and hepatolith.In order to help hepatobiliary surgeons to perform surgery more efficiently,this thesis proposes a medical image segmentation network based on U-Net,which can effectively segment bile ducts and hepatolith simultaneously.The main work of this thesis is:1.There is no public dataset for segmentation of bile ducts and hepatolith.With the cooperation with the Department of Hepatobiliary Surgery,the First Affiliated Hospital of Guangzhou Medical University,a new data set termed GZMU-HS is constructed based on the abdominal CT images of clinical cases;2.A new end-to-end fully convolutional network is designed to automatically segment bile ducts and hepatolith,which is based on U-Net.For simply description,it is named as M-Net according to the shape of the network.In the M-Net,two encoder-decoders are composed of four streams.In different streams,multi-scale sparse convolutions are designed to extract rich semantic features and multi-scale context information at different scales.In order to make full use of the advantages of multi-scale feature maps,a multi-stream feature fusion strategy is proposed to merge richest semantic features generated in the first stream into other streams.3.To further improve the segmentation performance,a new loss function has been defined,which is the bootstrap cross-entropy function.Its function is to focus the feature extraction of M-Net on pixels of indistinguishable bile ducts and hepatolith.By discarding the calculated loss values of easily distinguishable pixels with higher classification probabilities,the overall loss in the network is concentrated on these indistinguishable pixels,thereby making the training more effective and targeted.Experimental results show that the M-Net is superior to many latest deep learning methods,and can simultaneously segment bile ducts and hepatolith in CT images of the abdomen.The M-Net can achieves the image segmentation performance with 98.678%recall rate,84.427% accuracy,89.831% DICE,and 90.998% F1 score for bile ducts,;and99.894% recall rate,55.132% accuracy,71.248% DICE,and 71.051% F1 score for hepatolith.