Interpretable deep learning as a means for decrypting disease signature in multiple sclerosis.

Area of Science:

• Neuroimaging
• Neurology
• Artificial Intelligence in Medicine

Background:

• Mechanisms of multiple sclerosis (MS) remain poorly understood, necessitating advanced methods for early detection of tissue degeneration.
• Distinguishing between primary progressive MS (PPMS) and relapsing-remitting MS (RRMS) is crucial for effective patient management and treatment strategies.

Purpose of the Study:

• To identify and characterize microstructural differences between PPMS and RRMS using diffusion and structural magnetic resonance imaging (MRI).
• To leverage advanced computational models, including convolutional neural networks (CNNs), for analyzing complex MRI data and classifying MS subtypes.

Main Methods:

• Utilized novel microstructural descriptors derived from 3D-Simple Harmonics Oscillator Based Reconstruction and Estimation (3D-SHORE) and Rotation Invariant spherical harmonics Features (RI-SHF).
• Employed diffusion MRI (dMRI) metrics (fractional anisotropy, mean diffusivity) and T1-weighted images as benchmarks and for comparison.
• Applied CNN models and Layerwise Relevance Propagation (LRP) for feature analysis and visualization of relevant brain regions.

Main Results:

• Diffusion MRI features extracted from grey matter effectively differentiated between PPMS and RRMS patients.
• LRP heatmaps successfully highlighted key areas of relevance in the brain, consistent with established knowledge of MS pathology.
• The approach demonstrated potential for uncovering subtle data properties indicative of specific disease states.

Conclusions:

• Advanced dMRI analysis combined with CNNs and LRP offers a powerful tool for classifying MS subtypes.
• This methodology can aid in early disease detection and patient stratification by identifying critical microstructural signatures.
• LRP analysis has the potential to reveal previously unrecognized factors contributing to different MS disease states.