A Multi-scale Multiple Sclerosis Lesion Change Detection in a Multi-sequence MRI

Multiple sclerosis (MS) is a disease characterized by demyelinating lesions in the brain and spinal cord. Quantification of the amount of change in MS lesions in magnetic resonance imaging (MRI) over time is important for evaluation of drug effectiveness in clinical trials. Manual analysis of such longitudinal datasets is time- and cost prohibitive, and also prone to intra- and inter-rater variability. Accurate automated change detection methods would be highly desirable. We propose a new MS lesion change detection method that integrates a voxel’s multi-sequence MR intensity with its immediate neighborhood context and the texture of the extended neighborhood in a machine learning framework. On our dataset of 15 patients, the proposed method had higher performance (median AUC-ROC = 0.97, AUC-PR = 0.43, Wilcoxon’s signed rank test, p < 0.001) than implemented baseline methods. As such, the proposed method has potential clinical applications as an efficient, low-cost algorithm to capture and quantify local lesion change and growth.