The tumor microenvironment has attracted increasing attention because of its close association with tumor development and progression. The interactions between various types of cells (e.g., tumor-stroma and tumor-lymphocyte interactions) and the overall ecological diversity during cancer development are essential characteristics of prognosis, classification, and formulation of therapeutic strategies. Systematic investigation of the tumor microenvironment is crucial for patient stratification and precision medicine by incorporating tumor heterogeneity.

Histopathological images derived from tissue slides contain rich information about cell morphologies and tissue structures, which make it a cost-efficient and easily accessible tool to dissect tumor microenvironment for clinical decision-making. Spatial organization features (SOFs) in the tumor microenvironment based on histopathological images provide new insights into cancer biology and demonstrated strong clinical implications and have been exploited as independent diagnostic, prognostic, and predictive biomarkers. However, the scoring of SOFs relies heavily on visual judgment, which is inherently restricted by its subjectivity, lack of quantitation, and potential discrepancies between pathologists. In recent years, artificial intelligence (AI) has been put to the forefront of clinical research, significantly extending the capability of traditional pathology. AI-empowered digital pathology can not only distinguish between different cell and tissue types accurately but also connect image-based features with molecular characteristics such as gene mutations, microsatellite instability, and molecular subtypes. Without additional immunohistochemical or molecular tests, image-based deep learning provides an extremely inexpensive assay based on histology slides that are ubiquitously existing and can be performed easily in the setting of the routine histopathological evaluation.

The work in this thesis focuses on employing deep learning-derived spatial organization features to dissect the cancer microenvironment, improve the prediction of patient prognosis, and achieve cancer subtyping prediction on histopathology images. In specific, We propose a deep learning framework for automated tissue classification of H&E-stained whole-slide images followed by quantification of spatial organization features (SOFs). The framework we proposed is based on tissue classification, which is more efficient than the quantification of cell-based features. The annotations for different types of tissue are also more convenient, which is critical for data-driven deep learning model. The systematic profiling of spatial characteristics in the tumor microenvironment is achieved by identifying novel spatial organization features (SOFs) from multi-scale level and collecting reported indices. When compared with the deep (interlayer) features, the SOFs are based on direct quantification of clearly defined features, and hence the superiority in pathological and biological interpretation. Finally, we transfer the deep learned framework to different types of cancer with the help of pathologists and investigate the prediction performance for prognosis and cancer subtyping of the mentioned SOFs. The prognostic power and clinical associations are further validate in independent public and/or inhouse cohorts. In total, the deep learning framework we proposed in this thesis is reliable enough to extract sufficient spatial characteristics from routine H&E images and may providing a cost‐efficient tool for more quantitative analysis of tumor microenvironment and stratification of patients for more optimized clinical management.