Deep-learning-based Quantification And Integrated Analysis Of Histopathology Images And Multiomics Data For Breast Cancer

The cancer initiation and development can be characterized as the dynamic interaction of tumor and its microenvironment(TME).TME consists of epithelial tissue,stromal tissue and other tissues,among which there are a wide spectrum of cells including tumor cell,fibroblast and different immune cells.Whole-slide image(WSI)can provide rich information of TME,and there have been many clinical observations indicating that the spatial features of different tissues and cells on WSI have significant diagnostic and prognostic values in multiple cancers.Imaging-genomics,as one of the current research focuses,has combined the image-based features with bioinformatics as well as statistical technologies to explore the progression of cancer,and thus to bring new strategies for clinical treatment of cancer.For now,the evaluation and analysis of histopathology image is mainly carried out by pathologists,which can be time-consuming and may induce human bias.Due to the lack of automatic image processing system,the research on further analysis of the TME heterogeneity in specific cancer subtypes and the molecular processes governing these differences still has major limitations and remains further exploration.In this paper,we combined the image analysis models in machine learning with bioinformatic methods to quantify the TME in different breast cancer subtypes,and we further explored the associations between these image features and genomic data as well as patient prognosis.Below are the three main contents of this article(1)Segmentation and quantification of histopathological tissues for breast cancer based on DCNN model.The epithelial and stromal tissues are two fundamental and the most common tissues in histopathological image of breast cancer.We proposed an image processing pipeline based on Deep Convolutional Neural Network(DCNN)model to classify epithelial and stromal regions in Whole-slide image(WSI).Inspired by the clinical analysis of histopathology images,this system consists of three steps:(1)identification of Region of Interest(ROI);(2)distinguishing of different histopathological tissue regions;(3)quantification and evaluation of different histopathological tissues.We first trained the proposed DCNN model on a dataset which consists of H&E patches with tissue annotations,and then applied the well-trained model to 1,000 WSIs to perform tissue segmentation.Finally,based on the segmentation results,we calculated the ratio between the area of epithelial and stromal tissue to the overall tissue region for each WSI.(2)Fully automatic immune characterization of breast cancer based on cascaded-trained U-net model.Immunotherapy is a new and promising target for tumor treatment in the past 10 years.Tumor-infiltrating Lymphocyte(TIL)is one type of the immune cells which can be observed on H&E stained histopathological image,and is also one of the current research focuses of immunotherapy.We developed a cascaded-trained U-net model to automatically identify TILs in histopathological image.Based on the results of detection,we further extracted 43 quantified features of TILs,which mainly evaluated the spatial distribution of immune hotspots on WSI,including number of hotspots,within-group dispersion,between-group dispersion,and so on.(3)Integrated analysis of histopathological image features and multiomics data for different subtypes of breast cancer.We first divided all the breast cancer patients into three subtypes,i.e.ER-positive,ER-negative and Triple negative,and then performed integrated analysis for each subtype,respectively.(1)We used the statistical correlation analysis and gene enrichment analysis to explore the relationship between gene expression and epithelial and stromal tissues.The results indicate that the same histopathological tissue was associated with similar biological processes in different breast cancer subtypes,whereas each subtype had its own idiosyncratic biological processes governing the development of these tissues.(2)We applied bioinformatics technologies as well as Cox prediction model on integrated histopathological image data and multiomics data to demonstrate the molecular mechanism and prognostic prediction of the immune phenotypes in different breast cancer subtypes.We found that the immune phenotypes in ER-positive and ER-negative breast cancer subtypes are regulated by similar biological processes,but the immune phenotype Triple negative subtype has very unique molecular process.We also found that the Clustering Dispersion pattern of TILs on image is related to both immune-related genes and patient prognosis,which suggests that pathologists should pay more attention to the clumping patterns of TILs during clinical diagnosis.These findings are expected to bring new insights into the clinical immunological assessment and treatment.