A novel approach combined transfer learning and deep learning to predict TMB from histology image

Tumor Mutation Burden(TMB) is a quantifiable clinical indicator that can be used to predict the responses to immunotherapy of a range of tumors. However, the current DNA sequencing-based TMB measurement method represented by Whole Exome Sequencing (WES) is expensive and time-consuming, which limits its utilization in clinical practice. In this paper, we design a method through deep learning in order to predict TMB from available H&E stained whole slide images of gastrointestinal cancer. Experimental results demonstrate that our approach is capable of distinguishing high and low TMB with an AUC higher than 0.75. We further performed post-processing to improve the accuracy on both test sets to above 0.7 (0.71 accuracy for TMB-STAD and 0.77 accuracy for TMB-COAD-DX). Furthermore, the predicted low and high TMB patients with gastric and colon cancer have different survival rates, with p values of 0.348 and 0.8113, respectively, which indicates that our study is potentially helpful for practical treatment.