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

OBJECTIVES: The current work investigates the performance of different multivariate supervised machine learning models to predict the presence or absence of multiple sclerosis (MS) based on features derived from quantitative MRI acquisitions. The performance of these models was evaluated for images which are significantly degraded due to subject motion, a problem which is often observed in clinical routine diagnostics. Finally, the difference between a true multivariate analysis and the corresponding univariate analysis based on single parameters alone was addressed.

MATERIALS AND METHODS: 52 MS patients and 45 healthy controls where scanned on a 3T system. The datasets showed variable degrees of motion-associated artefacts. For each dataset, the average of T1 , T2 * , total and myelin bound water content was determined in white and grey matter. Based on these parameters, different multivariate models were trained and their cross-validated performance to predict the presence of MS was evaluated. Furthermore, the univariate distributions of each quantitative parameter were employed to define optimised cut-offs that differentiate MS patients from healthy controls.

RESULTS: For data not affected by motion, 83.7% of all subjects were correctly classified using a crossvalidated multivariate model. Inclusion of data with significant artefacts reduces the rate of correct classification to 74.5%. T1 in grey and myelin water content in white matter where the most discriminating variables in the multivariate analysis. In contrast, the total water content in white matter and the ratio of white and grey matter total water content each resulted in 77% correct classifications in a univariate regression analysis.

CONCLUSION: The results demonstrate that even simple quantitative MRI-based measures allow for an automated prediction of the presence/absence of multiple sclerosis with good specificity. Importantly, even highly degraded datasets due to motion-artefacts could be correctly classified, especially when pooling features derived from grey and white matter. Finally, the advantage of a multivariate over a univariate analysis of quantitative MR data was shown.