FedFound: Federated Foundation Model for GI endoscopy imaging

Gastrointestinal (GI) endoscopy is essential for identifying abnormalities and enabling early diagnosis, thereby improving diagnostic accuracy and facilitating timely clinical intervention. While deep learning has shown promise in supporting GI diagnostics, most models require carefully labeled datasets that are costly to acquire and limited in scope. Foundation models offer a paradigm shift by learning general-purpose representations that can be adapted for diverse tasks, addressing data scarcity and variability. However, training such models on medical imaging data remains challenging due to privacy regulations, security concerns, and fragmented digital infrastructures that prevent centralized data aggregation, even as hospitals collectively generate vast amounts of endoscopic data. To overcome these barriers, we propose a federated learning (FL) framework for GI foundation models, allowing data to remain local while collaboratively training a shared model. We extensively evaluate multiple FL algorithms under both homogeneous and heterogeneous settings across three clinically relevant downstream tasks-classification, detection, and segmentation. We further evaluate our framework with a larger number of clients and across diverse pretraining strategies to assess its scalability and robustness. Our results demonstrate consistent performance gains, underscoring the potential of Federated Foundation Models to enable clinically validated AI applications and support secure, regulation-compliant deployment in real-world gastroendoscopy practice.