Specificity-Aware Federated Learning With Dynamic Feature Fusion Network for Imbalanced Medical Image Classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) enables multi-client medical image classification while preserving privacy.
• Existing FL methods often overlook local model specificities and struggle with imbalanced medical datasets.

Purpose of the Study:

• To propose a novel specificity-aware federated learning framework to address limitations in current medical image classification FL approaches.
• To enhance global model performance by considering individual client model characteristics and mitigating class imbalance issues.

Main Methods:

• Introduced an Adaptive Aggregation Mechanism (AdapAM) utilizing reinforcement learning for adaptive weighting and aggregation of local model parameters based on data distribution and performance feedback.
• Developed a Dynamic Feature Fusion Strategy (DFFS) to address class imbalance by dynamically fusing majority class features and collaborating other features based on min-batch imbalance ratios.

Main Results:

• The proposed framework achieved state-of-the-art performance on both dermoscopic and fundus image datasets.
• Experimental results demonstrated the effectiveness of AdapAM in handling local model specificities and DFFS in managing class imbalance.

Conclusions:

• The specificity-aware federated learning framework with AdapAM and DFFS offers a robust solution for privacy-preserving medical image classification.
• This approach significantly improves classification accuracy in real-world medical applications with imbalanced data.