NeuroMoE: A Transformer-Based Mixture-of-Experts Framework for Multi-Modal Neurological Disorder Classification

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Neurological disorders (NDs) pose diagnostic challenges, particularly in distinguishing overlapping conditions.
• Current deep learning (DL) methods struggle to integrate multi-modal Magnetic Resonance Imaging (MRI) and clinical data effectively.
• A gap exists in mapping the continuum of ND progression from prodromal to established stages.

Purpose of the Study:

• To develop a novel DL framework for enhanced ND classification using multi-modal data.
• To improve the accuracy of diagnosing neurological disorders in clinical settings.
• To address the limitations of existing DL approaches in leveraging diverse medical data.

Main Methods:

• A proprietary multi-modal clinical dataset for ND research was utilized.
• A transformer-based Mixture-of-Experts (MoE) framework was proposed, integrating anatomical MRI (aMRI), Diffusion Tensor Imaging (DTI), and functional MRI (fMRI) with clinical assessments.
• Transformer encoders captured spatial relationships, while modality-specific experts and adaptive fusion enhanced feature extraction and integration.

Main Results:

• The proposed multi-modal framework achieved a validation accuracy of 82.47%.
• The approach significantly outperformed baseline methods by over 10% in diagnostic accuracy.
• The framework demonstrated particular effectiveness in distinguishing between overlapping neurological disease states.

Conclusions:

• The novel MoE framework effectively integrates multi-modal MRI and clinical data for improved ND diagnosis.
• This approach shows significant potential for enhancing diagnostic accuracy in real-world clinical applications.
• Advancements in multi-modal learning are crucial for understanding ND progression and differentiating complex disease states.