3458 - Validated Supervised Deep Learning on Pathology Whole-Slide Images for HER2 Mutation Prediction in Lung Adenocarcinoma

Y. Xu¹, B. Li¹, B. Zou¹, B. Fan¹, J. Yu², and L. Wang¹; ¹Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, ²Department of Radiation Oncology and Shandong Provincial Key Laboratory of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China

Purpose/Objective(s): Accurate evaluation of Human Epidermal Growth Factor Receptor 2 (HER2) is pivotal for the successful implementation of antibody-drug conjugate (ADC) therapy in lung adenocarcinoma patients. However, the prohibitive cost of genetic testing constrains the detection scope of HER2 mutations. To identify potential carriers of HER2 mutations, we propose a deep learning model utilizing a Visual Transformer (ViT) network on Hematoxylin and Eosin (H&E) stained specimens from lung adenocarcinoma patients.

Materials/

Methods: Lung adenocarcinoma patients were recruited from both Shandong Cancer Hospital and the TCGA database, undergoing separate processes for model training and external validation. Whole-slide images (WSIs) of H&E stained histopathological specimens were collected and segmented into patches of 1024 × 1024 pixels. The ViT-trained patch-level model identified predictive patches and generated corresponding probability distributions.

Results: Utilizing the ViT-Recursive Neural Network framework, we developed a prediction model for HER2 mutation status, subsequently subjecting it to external validation in the TCGA cohort. For comprehensive training and validation, we integrated 560 H&E stained histopathological specimens from 294 lung adenocarcinoma patients at Shandong Cancer Hospital and 60 WSIs from an additional 60 patients in the TCGA cohort. The model exhibited an accuracy of 87.3% in the internal validation cohort and 78.5% in the external validation cohort.

Conclusion: In conclusion, the ViT-Recursive Neural Network model, leveraging pathological WSIs, proves effective in predicting HER2 mutations in lung adenocarcinoma patients.