Modality-Agnostic Federated Learning With Adaptive Updates for Heterogeneous Medical Image Tasks

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) allows collaborative training on decentralized medical data, preserving privacy.
• Data heterogeneity in imaging modality (CT, MRI) and tasks (segmentation, classification) limits FL adoption.
• Existing FL methods struggle to create unified models for diverse medical datasets.

Purpose of the Study:

• To propose FedCMT, a modality-agnostic FL framework to address data heterogeneity in medical imaging.
• To enable flexible adaptation to various input modalities and local tasks within a federated network.
• To enhance collaboration and generalization in FL for medical image analysis.

Main Methods:

• FedCMT incorporates group-wise adapters and personalized decoders for modality- and task-specific feature capture.
• A conflict-averse module extracts modality-invariant representations, mitigating inter-client feature conflicts.
• Global-to-local knowledge distillation balances global consistency and local specialization.

Main Results:

• FedCMT demonstrated stability and fostered shared knowledge across diverse medical imaging modalities.
• Evaluated on ten CT and MR datasets with up to eight clients and varied tasks.
• Achieved an average improvement of 4.76% over state-of-the-art FL baselines and 4.01% over standalone training.

Conclusions:

• FedCMT effectively handles data heterogeneity in medical FL, outperforming existing methods.
• The framework supports flexible input modalities and diverse local tasks.
• FedCMT shows promise for real-world medical image analysis applications.