Abstract 2461: Federated learning enables multi-institution collaboration without sharing Hand E slides
Kodi Taraszka, Oded Rosolio, Intae Moon, Yevgeniy R. Semenov, Alexander Gusev · Cancer Research · 2025
Abstract Hematoxylin and eosin slides provide vital visual information on a collected tissue sample and are a key component of cancer diagnostics. Previous work in computational pathology has utilized deep learning architectures to accurately and efficiently tackle complex tasks such as tumor classification. However, the generalizability of these algorithms is currently limited due to the inherent difficulty of cross institution collaboration. Here, we propose leveraging federated learning (FL) which aggregates model level information between institutions to enable collaboration without exposing sensitive data. This allows the deep learning model to incorporate diverse patient populations and heterogeneous data sources which are needed for generalizability while preserving patient privacy. As a proof-of-concept, we used attention based multiple instance learning within the FL framework to perform cancer subtype classification between lung adenocarcinoma and lung squamous cell carcinoma with three distinct cohorts: Dana-Farber Cancer Institute (DFCI), Massachusetts General Hospital (MGH), and The Cancer Genome Atlas (TCGA). We present comparisons using three approaches: FL, institution specific analyses, and a centralized model where data is combined at one site (gold standard). We find all three approaches have comparable overall AUC. Additionally, we see that there is significant heterogeneity in AUC between cohorts which may indicate FL is more robust to overfitting and more generalizable. Comparing features generated using the foundation pathology model (UNI) to those generated from a ResNet50 model pre-trained on ImageNet (ResNet), we observed more institutional level heterogeneity in the UNI results than in ResNet which may indicate UNI is learning institution specific features that limit its generalizability. Overall our work shows FL enables generalizable deep learning across institutions without risking patient privacy. Citation Format: Kodi Taraszka, Oded Rosolio, Intae Moon, Yevgeniy Semenov, Alexander Gusev. Federated learning enables multi-institution collaboration without sharing Hand E slides [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 2461.