Research On Cell Recognition And Tissue Classification For Pathological Images Of Breast Cancer

Pathological image,as an important member of medical images,is the gold standard for disease diagnosis.Quantification description of pathological images,such as the morphology or spatial distribution of cells,and the type or size of tissues,etc.,helps to understand the potential patterns of the occurrence and development of diseases,thereby indicating that some cancers can be prevented,diagnosed,and intervened early to reduce the incidence and improve the survival rate.However,the high spatial resolution of pathological images brings great challenges to manual quantitative evaluation.The computer-aided analysis methods can effectively alleviate the situation that pathologists are difficult to accurately quantify pathological images in the clinical decision-making process,and assist pathologists to perform rapid and precise clinical diagnosis,course analysis and prognostic prediction in a more objective and robust automated way.In this thesis,we aim at the critical and difficult issues in the quantitative analysis of hematoxylin and eosin(H&E)-stained breast cancer pathological images,and carry out the following deep-learning-based research work at the cell level and tissue level:(1)Breast cancer lymphocyte recognition based on the dense dual-task network.To address the difficulty of manual quantification caused by the small size and complex spatial distribution of lymphocytes in breast cancer,this work proposes a novel Dense Dual-Task Network(DDTNet)for simultaneous and accurate automatic detection and segmentation of lymphocytes in H&E-stained pathological images of breast cancer.The feature encoding backbone network is used to extract multi-scale deep features related to lymphocyte position and morphology;the detection and segmentation decoders transform the deep features into the position and morphological probability maps of lymphocyte,respectively;a feature fusion strategy is utilized to efficiently introduce multi-scale features with lymphocyte location information for the segmentation decoder to improve cell recognition performance.Experiments on three independent breast cancer lymphocyte datasets demonstrate that DDTNet achieves excellent performance in both cell detection and segmentation metrics.In addition,this work also proposes a semi-automatic lymphocyte annotation method(TILAnno)based on level set,which provides the source for the development of the data-driven cell recognition framework.(2)Breast cancer tissue classification based on the efficient region-aware network.To address the difficulty of manual quantification caused by the variety and complex distribution of breast cancer tissues,and the real-time requirements for quantitative analysis of pathological images,this work proposes an Efficient Region-Aware Network(ERANet)for fast and accurate automatic classification of tissue structures in H&E-stained pathological images of breast cancer.The lightweight multi-scale feature extraction module efficiently aggregates multiscale context information under different receptive fields for local regions;the lightweight dynamic region attention module enables dynamic region-by-region enhancement of global features by learning region correlations;the sub-region classification strategy efficiently encodes neighborhood context information and quickly outputs a region classification map;the loss function based on regional regularization helps the model to establish the unique distribution relationship among various tissues of breast cancer.Experiments on three independent breast cancer tissue classification datasets demonstrate that ERANet achieves a good trade-off between model accuracy and inference speed.In summary,this thesis shows the feasibility of accurate quantification analysis of digital pathological images based on deep learning methods from two different perspectives,providing powerful tools and theoretical basis for the development of computational pathology.