MIMoSA: A Method for Inter-Modal Segmentation Analysis

Abstract

Magnetic resonance imaging (MRI) is crucial for in vivo detection and characterization of white matter lesions (WML) in multiple sclerosis. While these lesions have been studied for over two decades using MRI technology, automated segmentation remains challenging. Although the majority of statistical techniques for the automated segmentation of WML are based on a single imaging modality, recent advances have used multimodal techniques for identifying WML. Complementary imaging modalities emphasize different tissue properties, which can help identify and characterize interrelated features of lesions. However, prior work has ignored relationships between imaging modalities, which may be informative in this clinical context. To harness the coherent changes in these measurements, we utilized inter-modal coupling regression (IMCo) to estimate the covariance structure across modalities. We then used a local logistic regression, MIMoSA, which leverages new covariance features from IMCo regression as well as the mean structure of each imaging modality in order to model the probability that any voxel is part of a lesion. Finally, we introduced a novel thresholding algorithm to fully automate the estimation of the probability maps to generate fully automated segmentations masks for 94 subjects. To evaluate the performance of the automated segmentations generated using MIMoSA we compared results with gold standard manual segmentations. We demonstrate the superiority of MIMoSA to other automated segmentation techniques by comparing it to the OASIS algorithm as well as LesionTOADS. MIMoSA resulted in statistically significant improvement in lesion segmentation.