Modeling the variability in brain morphology and lesion distribution in multiple sclerosis by deep learning

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Multiple sclerosis (MS) is characterized by brain morphology changes and white matter lesions.
• The variability of these pathological features beyond simple volume measurements is not well understood.
• Existing methods struggle to model the complex spatial distribution of MS lesions.

Purpose of the Study:

• To develop a statistical model for automatically discovering spatial patterns of variability in brain morphology and lesion distribution in MS.
• To enhance the understanding of complex MS pathology through advanced computational techniques.
• To create a model that can capture concurrent morphological and lesion patterns.

Main Methods:

• Utilized a deep belief network (DBN), a layered neural network, for statistical modeling of brain images.
• Developed separate DBNs for brain morphology and lesion distribution, and a joint DBN for combined patterns.
• Leveraged the DBN's ability to learn parameters directly from training images, avoiding the need for pre-defined distance metrics for lesions.

Main Results:

• The DBN model successfully identified established patterns of MS pathology.
• The model also uncovered more subtle, previously unrecognized spatial patterns.
• Computed model parameters demonstrated significant correlations with clinical scores in MS patients.

Conclusions:

• The proposed DBN-based statistical model offers a powerful tool for analyzing complex MS brain imaging data.
• This approach effectively captures spatial variability in morphology and lesions, offering deeper insights into MS pathology.
• The findings suggest potential for using this model in clinical assessment and understanding disease progression.