Research On Survival Analysis Of Pathological Images Based On Deep Learning Method

Disease-free survival of patients is an important part of prognosis research in clinical medicine.It provides guidance for doctors to formulate therapeutic interventions to improve patient survival.Survival analysis based on clinical data can effectively predict patient prognosis,but there are many challenges in using complex and diverse clinical data for prognosis prediction.pathological images record the microscopic morphology of cells and the tumor microenvironment,and are the gold standard for cancer diagnosis.They reflect the severity of the disease and contain important prognostic information.However,manual reading by pathologists is extremely time-consuming and subject to subjective bias.Computer-based assisted analysis methods can effectively alleviate this problem.Methods for prognosis prediction using pathological images are roughly divided into two categories: methods based on region of interest(ROI)and methods based on whole slide image(WSI).Obtaining image patches from ROIs annotated by pathologists,and extracting artificial features such as tissue texture and morphology for prognostic analysis of patients,the performance is poor.This type of method has the following problems: 1)the cost of manual annotation by pathologists is huge and subjective bias? 2)the color difference leads to inaccurate description of the extracted artificial features? 3)the extracted tissue features are poorly interpretable.Automatically extract global information of whole slide pathological images from WSI for prognosis,reduce labor costs,eliminate subjective errors,and improve performance.However,the traditional WSI method does not filter out invalid patches and cannot effectively extract and aggregate prognostic features.In view of the above problems,this thesis studies the prognostic feature extraction method from the perspectives of cell microscopic morphology and full-image macroscopic expression of pathological images,and constructs two types of risk prediction models.The main work and contributions are summarized as follows:1)Deep Mitosis Surv method for mitotic cell features.Mitotic count is an important parameter for assessing tumor aggressiveness in most cancers.Based on the object detection network,the mitotic cell features were extracted as covariates,and the Cox proportional hazards model was constructed.This method is superior to prognostic method based on follow-up data and is interpretable.2)Deep GCNMIL method for whole slide pathological images.A three-layer graph convolutional network is used in a multi-instance learning framework to learn each pathological image patch phenotype representation,introducing attention-based multi-instance learning pooling to aggregate phenotype cluster features to output prognostic risk.The concordance index of this method on the NLST lung cancer dataset and the TCGA?BRCA breast cancer dataset leads the second place by 0.026 and 0.014,surpassing the current similar mainstream methods.For large-scale pathological images with GB pixels,this method can effectively assess patient prognostic risk and help provide personalized medicine.3)Deep GATMIL method for whole slide pathological images.Based on Deep GCNMIL,a three-layer attention graph network is introduced to learn the representation of each phenotype,using multi-head attention to aggregate phenotype cluster features.Its concordance index on the NLST and TCGA?BRCA datasets is 0.042,0.024 ahead of the second place.