Generalizable biomarker prediction from cancer pathology slides with self-supervised deep learning: A retrospective multi-centric study

• DOI: 10.1016/j.xcrm.2023.100980
• Publication Date: 2023-03-22T15:12:54Z
• Authors: Jan Moritz Niehues; Philip Quirke; Nicholas P. West; Heike I. Grabsch; Marko van Treeck; Yoni Schirris; Gregory P. Veldhu...; Gordon G.A. Hutchins; Susan D. Richman; Sebastian Foersch; Titus J. Brinker; Junya Fukuoka; Andrey Bychkov; Wataru Uegami; Daniel Truhn; Hermann Brenner; Alexander Brobeil; Michael Hoffmeister; Jakob Nikolas Kather
• Keywords: Proto-Oncogene Proteins B-raf; MICROSATELLITE INSTABILITY; Class I Phosphatidylinositol 3-Kinases; ARTIFICIAL-INTELLIGENCE; IMMUNE; ASSOCIATION; GENE; Article; 3. Good health; COLORECTAL-CANCER; MODEL; Proto-Oncogene Proteins p21(ras); Deep Learning; WNT PATHWAY; SURVIVAL; Humans; Microsatellite Instability; Colorectal Neoplasms; Biomarkers; Retrospective Studies

Deep learning (DL) can predict microsatellite instability (MSI) from routine histopathology slides of colorectal cancer (CRC). However, it is unclear whether DL also other biomarkers with high performance and predictions generalize to external patient populations. Here, we acquire CRC tissue samples two large multi-centric studies. We systematically compare six different state-of-the-art architectures pathology slides, including MSI mutations in BRAF, KRAS, NRAS, PIK3CA. Using a validation cohort provide realistic evaluation setting, show that models using self-supervised, attention-based multiple-instance consistently outperform previous approaches while offering explainable visualizations the indicative regions morphologies. While prediction BRAF reaches clinical-grade performance, mutation PIK3CA, NRAS was clinically insufficient.