Federated learning improves site performance in multicenter deep learning without data sharing.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Multi-institutional collaboration is crucial for developing robust AI models in healthcare.
• Sharing sensitive patient data across institutions poses significant privacy and logistical challenges.
• Federated learning (FL) offers a decentralized approach to model training.

Purpose of the Study:

• To demonstrate the feasibility and effectiveness of FL for multi-institutional training of deep learning models.
• To evaluate the performance and generalizability of an FL model compared to individually trained models.
• To confirm that FL can be implemented without centralizing or sharing underlying physical patient data.

Main Methods:

• Deep learning models were trained locally at each participating institution using their respective clinical datasets.
• A global model was trained collaboratively across all institutions using the federated learning approach.
• Model performance was assessed using held-out test sets from each institution and an external dataset.

Main Results:

• The federated learning model demonstrated superior performance and generalizability compared to models trained at single institutions.
• The FL model's overall performance significantly surpassed that of any individual institutional model.
• The study successfully implemented FL across three academic institutions, preserving data privacy.

Conclusions:

• Federated learning is a powerful methodology for enabling collaborative model development across multiple institutions.
• FL effectively addresses privacy concerns by eliminating the need for data transfer and pooling.
• Further research into FL is warranted to accelerate the development of generalizable clinical AI models.