Research On Pathological Image Classification And Survival Analysis Method Based On Deep Learning

The latest statistics from the American Cancer Society show that there were a total of19.3 million new cancer cases worldwide in 2020,with about 10 million cancer-related deaths.In China,there were about 4.064 million new cancer cases and about 2.413 million cancer-related deaths in 2016,posing a serious threat to the health and lives of the people.Pathological examination is the "gold standard" for cancer diagnosis,playing a crucial role in diagnosis,prognosis prediction,and other aspects.H&E stained pathological images are the most common pathological images,widely used in clinical practice due to their simple staining operation,good stability,and low cost.In traditional analysis,pathologists observe pathological sections under a microscope and evaluate patients’ conditions in conjunction with other clinical diagnostic information,which is labor-intensive,time-consuming,and has poor repeatability.The emergence of digital pathology scanners,as well as the development of computer hardware and deep learning technology,has brought new opportunities for computer-aided pathological image analysis.This thesis conducts two researches on H&E stained pathological image analysis based on deep learning technology:(1)The sizes of regions of interest in pathological images vary greatly,posing a significant challenge for AI-assisted diagnosis.To address this challenge,this article proposes a dual-branch fusion model(Bi Fusion Net)that combines graph neural networks and convolutional neural networks.Pathological images are processed into fixed-size images and graphs that maintain the original spatial structure of the images.To make the two branches complementary in performance,Focal loss is added to the graph neural network branch on top of the cross-entropy loss,and higher weights are set for difficult-to-classify samples.The final model is trained under the joint supervision of cross-entropy loss and Focal loss.The model achieved the best classification performance of 67.03% ± 2.04% on the breast cancer dataset BRACS and 97.33% ± 1.25% on the colorectal cancer CRA dataset.(2)Pathological images contain a wealth of prognostic information.To improve patient survival prediction,this thesis constructs a survival analysis framework based on H&E stained pathological images.The framework includes three parts: first,to obtain accurate cell nucleus category information,the existing Ho Ver-Net model is improved by using tissue information to correct misclassified cell nuclei;second,cell graphs are constructed using cell nucleus information,and features related to tumor-lymphocyte interaction are extracted.This feature is further quantified to construct a global representation of tumor-lymphocyte interaction(Tumour-Lymphocyte Spatial Interaction Score,TLSI-score);finally,prognostic validation and interpretability analysis are performed on lung adenocarcinoma pathological image datasets from three hospitals.Experimental results show that the TLSI-score is an independent prognostic factor related to disease-free survival.Combined with some clinical information,better survival prediction performance can be achieved compared to existing clinical models,with C-index values of 0.716,0.666,and 0.708 in the training set and two external validation sets,respectively.