Predicting Conversion from Clinically Isolated Syndrome to Multiple Sclerosis--A MRI Feature Based Machine Learning Approach

Purpose: MRI plays a central role in establishing the diagnosis of Multiple Sclerosis (MS). We hypothesized that studying MRI features of the lesions such as shape and brightness in the baseline scan of patients with a Clinically Isolated Syndrome (CIS) may predict conversion into MS. Methods: We performed a single centre analysis of a prospective cohort of 84 CIS patients, who were followed-up for at least 3 years. Conversion into definite MS was defined according to the 2010 McDonald criteria, i. e. encompassed clinical and radiological criteria. Lesions in 3D FLAIR and 3D T1 images were segmented semi-automatically. Shape and brightness features were automatically calculated from these masks and input into an oblique random forest machine learning model (RFM). Prediction accuracies were validated through a three-fold cross-validation. Results: 66 patients converted to MS and prediction of (non-)conversion was correctly in 71 patients in an RFM based on shape features. Brightness features did not improve the model's performance. This predictor was significantly more accurate than predicting with Barkhof's criteria (p < 0.001, McNemar's test) with a sensitivity of 94% and specificity of 50% (85% and 28% respectively for Barkhof's criteria). Conclusion: Our study shows that MRI shape parameters of lesions can contribute to predicting the future course of CIS patients more accurately.