High accuracy epidermal growth factor receptor mutation prediction via histopathological deep learning

Abstract Background: The detection of epidermal growth factor receptor (EGFR) mutations in patients with non-small cell lung cancer is critical for tyrosine kinase inhibitors therapy. EGFR detection requires tissue samples, which are difficult to obtain for some patients, losing the opportunity for further treatment. To realize EGFR mutation prediction without molecular detection, we aim to build a high accuracy deep learning model with only hematoxylin-eosin-stained slides. Methods: A total of 326 hematoxylin-eosin-stained non-small cell lung cancer slides were collected from Beijing Chest Hospital, China. After digitalized into whole slide images, 226 slides (88 with EGFR mutation) among them were fed to a convolutional neural network for model training. The remaining 100 images (50 with EGFR mutation) were used as the test set. Results: The sensitivity and specificity of the model were 76% and 74%, respectively, with an area under the curve of 0.82. When applying the double threshold approach, the deep learning model could screen out 40% of the patients with a sensitivity and specificity of 96% and 94%, respectively. By further involving adenocarcinoma subtype information, 37.3% of the adenocarcinoma patients could be screened out with 100% sensitivity and specificity. Conclusions: In this research, we show rapid and inexpensive pre-screening potentials of the deep learning-based EGFR mutation prediction model. It could not only act as a high-accuracy complement to current molecular detection techniques, but also provide opportunities for non-small cell lung cancer patients with limited sample to receive further treatment.