Self supervised learning improves dMMR/MSI detection from histology\n slides across multiple cancers

Abstract

Microsatellite instability (MSI) is a tumor phenotype whose diagnosis largely\nimpacts patient care in colorectal cancers (CRC), and is associated with\nresponse to immunotherapy in all solid tumors. Deep learning models detecting\nMSI tumors directly from H&E stained slides have shown promise in improving\ndiagnosis of MSI patients. Prior deep learning models for MSI detection have\nrelied on neural networks pretrained on ImageNet dataset, which does not\ncontain any medical image. In this study, we leverage recent advances in\nself-supervised learning by training neural networks on histology images from\nthe TCGA dataset using MoCo V2. We show that these networks consistently\noutperform their counterparts pretrained using ImageNet and obtain\nstate-of-the-art results for MSI detection with AUCs of 0.92 and 0.83 for CRC\nand gastric tumors, respectively. These models generalize well on an external\nCRC cohort (0.97 AUC on PAIP) and improve transfer from one organ to another.\nFinally we show that predictive image regions exhibit meaningful histological\npatterns, and that the use of MoCo features highlighted more relevant patterns\naccording to an expert pathologist.\n