Multivariate prediction of multiple sclerosis using robust quantitative MR-based image metrics

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Diagnostics

Background:

• Multiple Sclerosis (MS) diagnosis relies on clinical and imaging data.
• Quantitative Magnetic Resonance Imaging (qMRI) offers detailed tissue characterization.
• Motion artifacts are a common challenge in clinical MRI, potentially impacting diagnostic accuracy.

Purpose of the Study:

• To evaluate multivariate supervised machine learning models for predicting MS presence/absence using qMRI features.
• To assess model performance on MRI data with varying degrees of motion-induced artifacts.
• To compare multivariate analysis with univariate analysis for MS prediction.

Main Methods:

• 52 MS patients and 45 healthy controls underwent 3T MRI scans.
• Quantitative parameters (T1, T2*, total water content, myelin water content) were extracted from grey and white matter.
• Multivariate models were trained and cross-validated for MS prediction.
• Univariate analysis used optimized cut-offs for individual parameters.

Main Results:

• Cross-validated multivariate models achieved 83.7% correct classification on artifact-free data.
• Performance decreased to 74.5% with significant motion artifacts.
• T1 in grey matter and myelin water content in white matter were key discriminating features in multivariate analysis.
• Univariate analysis yielded up to 77% correct classification for specific parameters.

Conclusions:

• Quantitative MRI measures enable automated MS prediction with good specificity.
• Multivariate models effectively classify MS even with motion-degraded datasets, especially when combining grey and white matter features.
• Multivariate analysis demonstrates superior performance over univariate analysis for quantitative MR data in MS prediction.