Renal Cancer Prognosis Prediction Based On Pathology Images And Genomic Data

With the advent of whole slide scaner technology,large numbers of tissue slides are being scanned,which enables tissue slides to be represented and archived digitally.Digital pathology not only has substantial implications for telepathology and education,but also,as a new source of big data,provides huge research opportunities for medical image analysis.It is well-known that there is a lot of diagnostic and prognostic data embedded in pathology images.Computer image analysis and machine learning techniques mine useful information that doctors cannot observe in images or think of in mind,while avoiding some defects of doctors' diagnoses.For example,1)diagnoses are subjective and easily influenced by doctors' experience;2)it's easy to miss subtle lesions;3)it's difficult to obtain quantitative features.This dissertation deals with prognosis prediction for two most common renal cell carcinoma(clear cell and papillary renal cell carcinoma),using computerized analysis algorithms on digital pathology images.Novel topological features which describe tumor microenvironment are extracted from pathology images to predict papillary renal cell carcinoma prognosis,and some of them are found significantly related to survival.There are many kinds of cells in tumor microenvironment,including tumor cells,immune cells and stromal cells.As a highly heterogeneous disease,the progression of tumor is not only achieved by unlimited growth of the tumor cells,but also supported,stimulated,and nurtured by the microenvironment around it.However,traditional qualitative and/or semi-quantitative parameters obtained by pathologist's visual examination have very limited capability to capture this interaction between tumor and its microenvironment.We propose novel topological features that characterize spatial organization of different cell patterns in pathology images.Experimental results show that the proposed topological features provide better patient stratification than clinical prognostic factors.For another kidney cancer,clear cell renal cell carcinoma,we systematically analyze the relationship between tumor morphological features and gene expression features and find that these two types of data are complementary in predicting prognosis.In other words,combining these two types of data can significantly improve the capability of prognosis prediction.In our experiments,we also find that both pathology images and genomic data identify a poor prognosis subtype with high percentage of tumor stroma.The risk index of the prognostic model built on these two types of data is significantly related to overall survival and can also predict survival in early-stage(stage I and II)patients.Multivariate Cox regression analyses demonstrate that the prognostic value of our model is independent of other known clinical and molecular prognostic factors.Overall,this dissertation systematically shows the application of pathology image analysis and multimodal data analysis to renal cell carcinoma prognosis prediction,which is beneficial to clinical decision making and providing intelligent and personalized healthcare.