ClusMFL: A Cluster-Enhanced Framework for Modality-Incomplete Multimodal Federated Learning in Brain Imaging...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Multimodal Federated Learning (MFL) shows promise for collaborative model training in healthcare.
• Modality incompleteness, where specific brain imaging data (PET, MRI, CT) is missing, is a significant challenge in real-world applications.
• Existing MFL methods often assume complete data or oversimplify missing-modality scenarios.

Purpose of the Study:

• To propose ClusMFL, a novel MFL framework designed for cross-institutional brain imaging analysis under realistic modality incompleteness.
• To address both client-level and instance-level modality incompleteness.
• To enable effective knowledge transfer and model training even with missing data modalities.

Main Methods:

• ClusMFL leverages feature clustering using the FINCH algorithm to create modality-label specific cluster centers.
• Supervised contrastive learning is employed for feature alignment within modalities.
• Cluster centers act as proxies for missing modalities, facilitating cross-modal knowledge transfer.
• A modality-aware aggregation strategy is utilized to enhance performance in severely incomplete data scenarios.

Main Results:

• ClusMFL was evaluated on the ADNI dataset using structural MRI and PET scans.
• The framework demonstrated state-of-the-art performance compared to baseline methods across various levels of modality incompleteness.
• Results show ClusMFL's effectiveness in handling missing data and improving cross-institutional brain imaging analysis.

Conclusions:

• ClusMFL provides a scalable and effective solution for multimodal federated learning in brain imaging analysis, particularly under modality incompleteness.
• The proposed feature clustering and modality-aware aggregation strategies significantly improve model performance in challenging real-world scenarios.
• This framework advances collaborative medical image analysis by robustly handling missing data.