Survival Prediction Of Intrahepatic Cholangiocarcinoma Based On Histologic Image Analysis

Intrahepatic cholangiocarcinoma(ICC)is one kind of epithelial cancer that located in the liver and originated from the secondary hepatic duct or its branches and it is a highly malignant cancer.Due to the lack of evident early clinical symptoms,most of the patients are diagnosed at the advanced stage,and even if eligible patients were operated on,the 5-year postoperative survival rate is only around 30%.We intend to extract image information that is closely related to patients' survival though performing quantitative analysis of multiple tissue and immune cell on whole slide images of ICC patients to achieve the goal of survival prediction of ICC patients.In this paper,we first introduce a multi-tissue semantic segmentation and a cell detection models based on the deep learning.In the multi-tissue segmentation pipeline,we compared our model with other commonly used segmentation models and get a better performance.Towards cell detection,we propose an automatic two-step t-lymphocyte detection pipeline based on the box regression and the cell salient.The comparison experiments based on different datasets show that our model has strong robustness.Secondly,we do further image computing based on the results of the segmentation and detection pipeline.We build a series of histological features using the tissue,cell component,cell population and distribution.We could thus quantitatively describe the whole slide images though these features.Finally,combining the m RMR and t-test for feature selection with LDA,SVM and Bagged C4.5 for classifiers,we construct the effective survival prediction model.In the individual testing set,our model gets AUC = 0.72 for predictively stratifying ICC patients into low survival risks and high risks.Additionally,to further verify the effectiveness of our model on survival prediction of patients,we do Log-rank test for univariant analysis and get ? 0.05.Simultaneously,Cox multi-variant regression analysis is also implemented to validate our model's performance compared with traditional clinical outcome related signatures.The result shows that ? 0.05 and HR=13.7594(95%CI: 5.8146-32.559)which means the great ability of patient risk identification and the high value of survival prediction ability.Therefore,the survival prediction model proposed in this paper can empower clinicians to better understand tumors on the basis of accurately predicting the survival risk of patients,thereby assisting doctors to develop more appropriate treatment strategies for ICC patients and help achieve the goal of precision medicine.