No abstract

Tissue biomarkers are crucial for cancer diagnosis, prognosis assessment, and treatment planning. However, few of current biomarkers used in clinics are robust enough to show a true analytical and clinical value. Thus the search for additional tissue biomarkers, including the strategies to identify them, is imperative. Recently, the capabilities of deep learning (DL)-based computational pathology in cancer diagnosis and prognosis have been explored, but the limited interpretability and generalizability make the results difficult to be accepted in clinical practice. Here we present an interpretable human-centric DL-guided framework—PathFinder (Pathological-biomarker-finder)— that can inspire pathologists to discover new tissue biomarkers from well-performing DL models, which bridges the gap between DL and clinical prognosis. By combining sparse multi-class tissue spatial distribution information of whole slide images (WSIs) with attribution methods, PathFinder can achieve localization, characterization, and verification of potential biomarkers, while guaranteeing state-of-the-art prognostic performance. With the inspiration of PathFinder, we discovered that tumor necrosis in liver cancer, a long-neglected factor, has a strong relationship with patient prognosis. Thus we proposed two clinically independent indicators, including necrosis area fraction and tumor necrosis distribution, for practical prognosis, and verified their potentials in clinical prognosis according to Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK)-derived criteria. Our work demonstrates a successful example of introducing artificial intelligence (AI) into clinical practice in a knowledge discovery way, which can be adopted in identifying biomarkers in various cancer types and modalities.