Direct Prediction of BRAFV600E Mutation from Histopathological Images in Papillary Thyroid Carcinoma with a Deep Learning Workflow

Papillary Thyroid Carcinoma (PTC) is the most common type of thyroid cancer. BRAFV600E is a prominent oncogenic mutation and has been found to have strong associations with the mortality and recurrence of PTC. In this paper, we propose a workflow to show that BRAFV600E mutation status can be directly predicted from histopathological images with deep learning. Our method mainly consists of two steps, tumor detection and mutation classification; each of them contains a Convolutional Neural Network (CNN). The information derived from the two steps are combined to predict mutation. We propose three different strategies and build a PTC dataset of 6,541,586 512×512-pixel patches from 439 H&E stained Whole Slide Images (WSIs) to perform our experiments. In the PTC-V600E strategy, we use patches from 50 PTC WSIs and 202 WSIs of 200 cases to train the two networks, respectively, and get an AUC of 0.884 on the test of 187 WSIs of 186 cases. In PCam-V600E and PAIP-V600E strategies, we use public datasets, PCam and PAIP 2019, of other cancer types instead of 50 PTC WSIs and get AUCs of 0.884 and 0.860. All the three strategies separate mutation-positive and negative cases successfully in our experiments, demonstrating the availability and feasibility of our work and its potential in further research and applications.