| Intrahepatic cholangiocarcinoma(ICC)is a malignant tumor,the second most common primary liver cancer,with a high mortality rate and poor prognosis.With its increasing incidence in recent years and the huge population base in China,the extent of its harm to people’s health cannot be ignored.ICC has no obvious symptoms in the early stage,so it is not easy to be detected and is usually found in the middle to late stage.The main treatment for ICC is resection of the lesion,but the postoperative effect is not satisfactory.Therefore,it is especially important to develop individualized treatment plans for different patients’ conditions.The degree of pathological differentiation is one of the key indicators of tumor pathology,which can reflect the development of the disease to a certain extent.However,preoperative detection is usually performed by tissue biopsy,which is painful to patients and cannot provide a comprehensive understanding of the degree of differentiation of the entire lesion area.Therefore,it is of great importance to construct a model that can accurately predict the degree of ICC pathological differentiation.In order to efficiently predict the degree of pathological differentiation in patients with ICC,a prediction model based on a self-supervised approach and a prediction model for the long-tailed distribution problem are proposed in this thesis,respectively.At the same time,a high-quality dataset containing 408 cases and 12,075 CT images was finally constructed by screening 455 cases.Most of the traditional studies on ICC use radiomics,machine learning and other methods.These methods have some subjective nature of human intervention.In view of the powerful capability of deep learning methods for feature extraction,this thesis proposes two deep learning-based models for predicting the degree of pathological differentiation of ICC.First,this thesis briefly introduces and analyzes the collected dataset,and explores the preprocessing methods of CT images accordingly,and summarizes the method of fixed-size cropping based on the center of the lesion applicable to ICC.In this thesis,a prediction model of pathological differentiation degree of ICC based on a selfsupervised method is proposed.Through the introduction of a novel self-supervised task,the feature extraction capability of deep convolutional networks is further enhanced to improve the accuracy of prediction.Second,medical data sets always present the problem of long-tail distribution due to the difficulty of data collection in the medical field and the specificity of the source population.The long-tail distribution problem can largely affect the effect of deep learning models.In this thesis,we propose a method to solve this problem by expanding the data under multiple windows with loss weighting,and construct a pathological differentiation degree prediction model for ICC with long-tail distribution.Extensive experiments show that the pathological differentiation degree prediction model constructed by the two methods proposed in this thesis has good performance and generalizability. |
