Recent studies based on deep learning on pathology images have shown the potential prognostic role of Tumor-infiltrating lymphocytes (TILs) in many cancer types. In general, most of these studies focused on detecting, quantifying and identifying TILs and spatial organization in entire whole slide images (WSIs) including both non-cancer and cancer regions. In this work, we hypothesize that TILs present in cancer regions not entire WSIs could have prognostic values. Here we present an automatic technique based on convolutional neural networks (CNN) that detects both cancer and TILs regions, and explores fraction and spatial organization of TILs in the detected cancer region from WSIs. A novel set of quantitative histological image features reflecting TILs fraction and spatial organization is extracted and used to stratify patients with distinct survival outcomes. In experiments using 277 different patient pathology slides selected from The Cancer Genome Atlas (TCGA) bladder cancer cohort, we show that bladder cancer patient survival is significantly correlated with our designed TIL-related image features, with a log-rank test p value below 0.05. Our method is extensible to histopathology images of other organs for predicting patient survival via TIL analysis.
- deep learning
- pathology image analysis
ASJC Scopus subject areas
- Biochemistry, Genetics and Molecular Biology(all)
- Agricultural and Biological Sciences(all)
- Immunology and Microbiology(all)
- Pharmacology, Toxicology and Pharmaceutics(all)