Communication-Efficient Federated Learning for Multi-Institutional Medical Image Classification.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Distributed machine learning

Background:

• Federated learning (FL) is popular for privacy-preserving medical image analysis by keeping data localized.
• Non-identically and independently distributed (non-i.i.d.) data in FL can cause communication bottlenecks due to frequent model aggregation.
• Existing FL frameworks struggle with communication efficiency when dealing with heterogeneous data.

Purpose of the Study:

• To propose a communication-efficient federated learning framework for medical image analysis.
• To address the challenges posed by non-i.i.d. data in federated networks.
• To reduce communication overhead without compromising model accuracy.

Main Methods:

• Developed an adaptive server-client model transmission strategy.
• Implemented conditional model uploads based on probability and informative update thresholds.
• Incorporated a proximal term to handle data heterogeneity in federated networks.

Main Results:

• The proposed framework significantly reduces communication costs compared to other FL algorithms.
• Maintained high model accuracy on non-i.i.d. datasets, demonstrated on a diabetic retinopathy rating task.
• Effectively tackled data heterogeneity challenges in federated learning.

Conclusions:

• The novel FL framework enhances communication efficiency in medical image analysis.
• The adaptive transmission and data heterogeneity handling mechanisms are effective.
• This approach offers a practical solution for large-scale, privacy-preserving medical AI.