Abstract 5048: Imaging based EGFR mutation subtype classification using EfficientNet

Abstract

Abstract This study aimed to determine whether EfficientNet-B0 was able to classify EGFR mutation subtypes with H&E stained whole slide images of lung and lymph node tissue. Background: Non-small cell lung cancer (NSCLC) accounts for the majority of all lung adenocarcinomas, with estimates that up to a third of such cases have a mutation in their epidermal growth factor receptor (EGFR). EGFR mutations can occur in various subtypes, such as Exon19 deletion, and L858R substitution, which are important for early therapy decisions. Here, we propose a deep learning approach for detecting and classifying EGFR mutation subtypes, which will greatly reduce the cost of determining mutation status, allowing for testing in a low resource setting. Methods: An EfficientNet-B0 model was trained with whole slide images of lung tissue or metastatic lymph nodes with known EGFR mutation subtype (wild type, exon19 deletion or L858R substitution). Regions of interest were tiled into 512x512 pixel images. The RGB .jpeg tiles are augmented by rotating 90°, 180°, 270°, and mirroring. The model was initialized with random parameters and trained with a batch size of 32, a learning rate of 0.0001 for 1 epoch before the validation loss increased for the next 5 epochs. Results: The model achieved a slide AUC of 0.8333, and a tile AUC of 0.8010. Slide AUC is the result of averaging all tiles within a slide and measuring performance based on correctly predicted slides (n=18). Tile AUC is the result of measuring performance based on correctly predicted tiles (n=102,000). Conclusion: Using EfficientNet-B0 architecture as the basis for our EGFR mutation classification system, we were able to create a top performing model and achieve a slide AUC of 0.833 and tile AUC of 0.801. Healthcare providers and researchers may utilize this AI model in clinical settings to allow for detection of EGFR mutation from routinely captured images and bypass expensive and time consuming sequencing methods. Table 1. Number of image tiles used and the number of slides they were extracted from. Train Validation Test Exon19 tiles 187,384 47,904 33,096 L858R tiles 166,288 19,512 26,136 Wild type tiles 225,944 27,696 42,768 Exon19 slides 47 6 6 L858R slides 46 6 6 WIld type slides 43 6 6 Citation Format: Daniel L. Franklin, Tara Pattilachan, Anthony Magliocco. Imaging based EGFR mutation subtype classification using EfficientNet [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5048.