AI-powered analysis of tissue slides to reveal the cellular composition and the spatial organization of the tumor microenvironment

Introduction - Digital pathology and artificial intelligence (AI) applied to histopathological images are gaining interest in immune-oncology, since they allow to streamline and ameliorate diagnostic and prognostic processes. The aim of this work was the development of a pipeline for the computational analysis of cancer tissues stained by H&E to find cell- or tissue-level features that could have a clinical relevance. Methods - The pipeline includes the adoption of machine and deep learning algorithms for what concerns cell segmentation, cell classification, tissue segmentation and spatial analysis. Specifically, open source platforms such as QuPath and RStudio were adopted to segment cells, classify them in a supervised manner, and perform spatial analyses. Results – On H&E images, we trained a machine learning classifier to detect tumor cells within the tumor region and then analyse their spatial clustering by exploiting the Ripley’s K function. Accordingly, patients were classified as “highly clustered”, “poorly clustered” or “uniformly distributed”. Then, another classifier was trained to distinguish lymphocytes from other cells, and their density was computed within and outside the tumor bed. According to this score, samples were classified as “immune desert”, “immune excluded” and “inflamed”. The combination of these two AI-based classifications significantly correlated with the prognosis. Conclusion – AI-powered H&E allowed us to classify samples based on quantitative data and the combination of the tumor and immune predictors generated clinical relevant results. These tools, once validated, may contribute to discover novel tumor and immune classifiers and human interpretable features.