Completed Feature Disentanglement Learning for Multimodal MRIs Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Multimodal Magnetic Resonance Imaging (MRI) is vital for clinical diagnosis and treatment.
• Feature disentanglement (FD) methods enhance multimodal learning (MML) by separating shared and specific features.
• Existing FD methods struggle with >2 modalities, losing critical shared information and lacking feature relationship interpretation.

Purpose of the Study:

• To address limitations in current FD-based MML methods for multimodal MRI.
• To propose a novel Complete Feature Disentanglement (CFD) strategy to recover lost shared information.
• To introduce a Dynamic Mixture-of-Experts Fusion (DMF) module for effective feature integration.

Main Methods:

• Developed a Complete Feature Disentanglement (CFD) strategy to identify modality-shared, modality-specific, and modality-partial-shared features.
• Introduced a Dynamic Mixture-of-Experts Fusion (DMF) module to dynamically integrate decoupled features by learning local-global relationships.
• Validated the approach on classification tasks using three multimodal MRI datasets.

Main Results:

• The proposed CFD strategy successfully recovers lost shared information among subsets of modalities.
• The DMF module effectively integrates decoupled features, capturing complex relationships.
• Experimental results show superior performance of the proposed approach over state-of-the-art MML methods on multimodal MRI classification.

Conclusions:

• The CFD strategy and DMF module significantly enhance MML for multimodal MRI analysis.
• This novel approach improves prediction accuracy by preserving and effectively utilizing shared feature information.
• The method demonstrates robust performance and outperforms existing techniques in multimodal MRI classification tasks.