Classification of Multiple Sclerosis Clinical Profiles via Graph Convolutional Neural Networks

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Neurological diseases like Multiple Sclerosis (MS) require precise characterization for effective treatment.
• Advances in neuroimaging and deep learning offer new avenues for disease classification.
• Current methods may not fully capture the nuances of different MS clinical profiles.

Purpose of the Study:

• To develop and evaluate a neural network-based approach for classifying MS patients into four distinct clinical profiles.
• To assess the utility of structural connectivity information derived from diffusion tensor imaging (DTI).
• To compare classification performance using unweighted and weighted connectivity matrices and investigate graph-based features.

Main Methods:

• Utilized diffusion tensor imaging (DTI) to obtain structural connectivity information, represented as graphs.
• Employed a neural network architecture to classify patients based on graph representations.
• Evaluated classification performance using both unweighted and weighted connectivity matrices.
• Investigated the contribution of graph-based features to classification accuracy.

Main Results:

• Neural network methods demonstrated high performance in classifying MS clinical profiles.
• Weighted connectivity matrices significantly improved classification compared to unweighted ones.
• Local graph metrics did not enhance classification, indicating the importance of neural network-derived latent features.
• Graph weights effectively captured information crucial for discriminating between different MS clinical forms.

Conclusions:

• Deep learning, particularly neural networks, shows great potential for classifying MS clinical profiles using brain connectivity data.
• Structural connectivity, especially when represented with weighted matrices, contains vital information for differentiating MS subtypes.
• The intrinsic feature extraction capabilities of neural networks outperform traditional graph metrics for this classification task.