Digital Pathological Image Analysis Based On Deep Learning

Pathological examination is the golden standard of numerous diseases in clinics.Powered by image processing and artificial intelligence techniques,modern computer-aided diagnostic(CAD)models are able to detect lesions in histopathological images.In the context of whole-slide images,more information such as tumor microenvironment,immuno-microenvironment can be obtained.This information plays a crucial role in risk stratification and may support disease management.The challenges in digital pathological image analysis mainly include the huge data volume,the histological and apparent heterogeneity,the cost of data annotation,and the requirement of model interpretability.This thesis focuses on deep learning-based digital pathological image analytic systems,carries on the investigations from three aspects.The main contents are as follows:1.We made a systematic summary of recent advances in computational pathology,with emphasis on its applications in lung cancer,colorectal cancer,breast cancer,immunohistochemistry-stained slides,and non-tumor disease.At the same time,we showed a full development procedure of a computational pathology model in the context of a foreground segmentation task.2.To solve the challenge of histologic and apparent heterogeneity,we propose an analytic tool to visualize the staining pattern of large cohorts of digital slides.The method is composed of procedures including sampling,clustering,defining and solving distances between color spectrums,and dimensional reduction.Using this tool,we were able to analyze the development and deployment of a CAD model,leading to several conclusions that have practical guiding significance.3.To deal with the expensiveness of data annotation,we propose to combine deep pre-encoding and logistic regression models.Because of the high efficiency,it can be integrated into an active learning framework,serving as a potential solution for interactive data annotation.We verified the method on three large-scale datasets,and conducted intensive comparative studies regarding the type of encoder and active learning strategy.4.We developed a number of computational pathology models for specific diseases,and made investigations towards model interpretability and its clinical value.(a)We proposed a clinical heart failure detection model in endocardial myocardial biopsy images;along with the training of VGG models for detection task,we propose to use techniques including Grad-CAM,UMAP,and overlapped cross-validation to obtain interpretable predictions,adding convincement of the models.(b)We propose an automatic framework to quantify tumor microenvironment(TME)in colon adenocarcinoma slides,and verify the prognostic value of tumor stroma,necrosis,and lymphocytes distribution.(c)We also propose a U-Net-based model for cancer detection in lung specimens.Combined with the spatial distribution map of lymphocytes,we quantified the interaction between tumor cells and lymphocytes,which turned out to has prognostic value.