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 inhibitor therapy. EGFR detection requires tissue samples, which are difficult to obtain in some patients, costing them the opportunity for further treatment. To realize EGFR mutation prediction without molecular detection, we aimed to build a high-accuracy deep learning model with only haematoxylin and eosin (H&E)-stained slides. Methods: We collected 326 H&E-stained non-small cell lung cancer slides from Beijing Chest Hospital, China, and used 226 slides (88 with EGFR mutations) for model training. The remaining 100 images (50 with EGFR mutations) were used for testing. We trained a convolutional neural network based on ResNet-50 to classify EGFR mutation status on the slide level. 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, 33% of the patients could be predicted by the deep learning model as EGFR positive or negative with a sensitivity and specificity of 100.0% and 87.5%. The remaining 67% of the patients got an uncertain result and will be recommenced to perform further examination. By incorporating adenocarcinoma subtype information, we achieved 100% sensitivity in predicting EGFR mutations in 37.3% of adenocarcinoma patients. Conclusions: Our study demonstrates the potential of a deep learning-based EGFR mutation prediction model for rapid and cost-effective pre-screening. It could serve as a high-accuracy complement to current molecular detection methods and provide treatment opportunities for non-small cell lung cancer patients from whom limited samples are available.