A distribution information sharing federated learning approach for medical image data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Federated learning (FL) is crucial for cross-silo scenarios like healthcare due to privacy benefits.
• Non-Independent and Identically Distributed (non-IID) data is a common challenge in FL, degrading model performance.
• Existing FL algorithms face performance degradation when applied to heterogeneous medical datasets.

Purpose of the Study:

• To propose a novel federated learning approach, FedDIS, to address the non-IID problem in medical image classification.
• To enhance FL performance by reducing data heterogeneity across clients through local data augmentation.
• To maintain patient privacy while enabling effective data distribution sharing.

Main Methods:

• Federated training of a Variational Autoencoder (VAE) to learn data distribution.
• Sharing encoded data distribution information among clients.
• Augmenting local datasets using the VAE decoder based on shared distribution information.
• Training the final classification model using combined local and augmented data in a federated manner.

Main Results:

• The proposed FedDIS method significantly improves federated learning performance on non-IID datasets.
• Experiments demonstrated effectiveness on Alzheimer's disease MRI dataset and MNIST classification.
• FedDIS successfully mitigates performance degradation caused by data heterogeneity.

Conclusions:

• FedDIS offers an effective solution for federated learning in medical imaging under non-IID conditions.
• The approach balances performance improvement with essential patient privacy protection.
• Sharing distribution information is a viable strategy to overcome data heterogeneity in FL.