TADynFed: Dynamic modality-adaptive federated learning with tissue-aware disentanglement for cross-disease analysis

Area of Science:

• Medical image analysis
• Federated learning
• Artificial intelligence in healthcare

Background:

• Federated learning (FL) enables collaborative medical image analysis across institutions while preserving data privacy.
• Real-world FL faces challenges like modality heterogeneity (incomplete/varying data) and cross-disease generalization.
• Existing FL methods struggle with uniform modality assumptions and static client participation, hindering clinical deployment.

Purpose of the Study:

• To propose TADynFed, a novel framework for Heterogeneous Federated Learning (HFL).
• To address both data and modality heterogeneity in federated medical image analysis.
• To improve model performance under realistic clinical constraints, including missing modalities and diverse pathologies.

Main Methods:

• TADynFed employs a tissue-aware disentanglement strategy to decouple modality-specific and shared features.
• A dynamic prototype memory bank compensates for missing modalities.
• An adaptive aggregation mechanism considers client reliability and tenure for robust learning.

Main Results:

• TADynFed achieved an average mDice score of 66.03%, significantly outperforming baseline methods (PointTransformerFL: 58.60%, FedAvg: 53.06%).
• Demonstrated superior boundary alignment (ASD: 1.85 mm, HD95: 8.70 mm) and calibration stability (ECE: 0.09).
• Exhibited high communication efficiency (76 MB/round) and strong cross-disease transferability without retraining.

Conclusions:

• TADynFed establishes a new benchmark for realistic, heterogeneous, and mix-modal federated medical imaging systems.
• The framework effectively maintains high segmentation accuracy, boundary precision, and calibration stability while minimizing bandwidth.
• TADynFed shows robust performance across diverse datasets and pathologies, highlighting its clinical applicability.