New Publication from NearData Partners: Federated EndoViT

We’re excited to highlight a new research article from our partners in the NearData project, titled:
“Federated EndoViT: Pretraining Vision Transformers via Federated Learning on Endoscopic Image Collections.”

About the Study

This publication addresses a key challenge in medical AI:
How can we develop robust and accurate AI models without compromising patient privacy?

The research team proposes a novel solution that combines federated learning and confidential computing to collaboratively train Vision Transformers on surgical data analysis—without sharing sensitive medical information across institutions.

The result is FL-EndoViT, a federated approach to model pretraining that allows for effective and secure collaboration between hospitals. Advanced techniques such as FedSAM and Stochastic Weight Averaging (SWA) were used to improve performance and stability across distributed datasets.

📄 Read the full article on arXiv: https://arxiv.org/pdf/2504.16612

Authors:
Max Kirchner, Alexander C. Jenke, Sebastian Bodenstedt, Fiona R. Kolbinger, Oliver L. Saldanha, Jakob N. Kather, Martin Wagner, and Stefanie Speidel

Why It Matters

This work is an important step toward building trustworthy, privacy-preserving AI systems in healthcare — proving that innovation and data protection can go hand in hand.