Federated learning using model projection for multi-center disease diagnosis with non-IID data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Multi-center disease diagnosis requires a global model but faces privacy barriers with centralized learning.
• Federated Learning (FL) enables collaborative model training while preserving local patient data privacy.
• Non-Independent and Identically Distributed (Non-IID) data in FL causes catastrophic forgetting and slow convergence.

Purpose of the Study:

• To address the challenges of catastrophic forgetting and invalid aggregation in federated learning for multi-center disease diagnosis.
• To propose an innovative federated learning approach, Federated learning using Model Projection (FedMoP), to improve model performance and convergence.

Main Methods:

• Federated learning using Model Projection (FedMoP) is introduced to ensure local model performance is not degraded after local training.
• FedMoP guarantees that the global model's performance on local data improves after aggregation, enhancing convergence.
• The method operates without direct access to global or local data during critical training phases.

Main Results:

• FedMoP significantly outperforms state-of-the-art FL methods in accuracy, convergence rate, and communication cost.
• Experimental results demonstrate that FedMoP achieves accuracy comparable to or exceeding centralized learning.
• The proposed method effectively mitigates catastrophic forgetting and invalid aggregation issues inherent in Non-IID federated learning.

Conclusions:

• FedMoP offers a privacy-preserving solution for multi-center disease diagnosis using federated learning.
• The approach enhances model accuracy, convergence speed, and reduces communication overhead.
• FedMoP presents a viable alternative to centralized learning, delivering superior or equivalent performance with enhanced privacy.