Abstract 2451: H&E whole-slide image (WSI) based artificial intelligence (AI) model to detect EGFR and ALK mutation in non-small cell lung cancer (NSCLC)

Abstract Detection of mutated epidermal growth factor receptor (EGFR) and anaplastic lymphoma kinase (ALK) are essential to guide therapeutic decisions for patients with NSCLC. Although traditional molecular testing methods-immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH)-and comprehensive genomic profiling (CGP) can provide extensive molecular insights, they are limited by prolonged turnaround time, higher tissue requirements, and elevated costs. To address this challenge, we tested a deep learning model that leverages digital pathology to analyze scanned H&E-stained WSIs for identifying EGFR and ALK mutations in NSCLC patients. The study utilized a commercial assay to predict EGFR and ALK genomic alterations directly from H&E-stained WSIs through a cloud-based platform. The assay was developed by Imagene Inc. using multiple instance learning (MIL) methods, trained with embeddings that were extracted by the commercial CanvOITM Foundation Model (FM). The FM ensured robust generalizability and computational power by leveraging 632, 608 digitized H&E-stained tissue samples. Thresholds were optimized on a held-out dataset to achieve positive predictive values (PPVs) of 50% and 40% for EGFR and ALK, respectively, while maintaining a 99% negative predictive value (NPV) for both genes. We tested the predictions in a blind retrospective evaluation including 164 digitized H&E images with EGFR molecular information and 158 images with ALK fusion data. In this study, 29 patients had EGFR-positive and 3 had ALK-positive mutations detected by commercial CGP providers. For EGFR mutation prediction, our model achieved a sensitivity of 48.3%, a specificity of 88.3%, a PPV of 66.7%, and an NPV of 100%. For ALK mutation prediction, the model achieved a sensitivity and NPV of 100%, specificity of 99.1%, and a PPV of 75%. Additionally, 54.88% (90/164) of EGFR and 28.48% (45/158) of ALK cases were classified as “intermediate, ” indicating uncertain results. Our study highlighted the potential of our prediction model for identifying EGFR and ALK mutations to guide targeted therapies in NSCLC patients. The model’s NPV of 100% demonstrates its strong ability to accurately rule out both EGFR and ALK mutations, ensuring that patients without these mutations are not subjected to unnecessary treatments. Furthermore, the ALK mutation prediction was promising with a sensitivity of 100% and specificity of 99.1%, demonstrating the model’s ability to detect all true positive cases while effectively minimizing false positive results. The presence of intermediate results suggests the need for additional diagnostic steps to confirm or exclude mutations and ensure accurate clinical decisions. Excluding EGFR and ALK alterations directly from the diagnostic H&E image addresses the demand for rapid genomic analysis driven by the rising use of neoadjuvant immunotherapy. Citation Format: Chan Mi Jung, Brandon Park, Hannah Son, Liam IL-Young Chung, Young Kwang Chae. H&E whole-slide image (WSI) based artificial intelligence (AI) model to detect EGFR and ALK mutation in non-small cell lung cancer (NSCLC) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2451.