MIMoSA: An Automated Method for Intermodal Segmentation Analysis of Multiple Sclerosis Brain Lesions

Abstract

Background and purpose: Magnetic resonance imaging (MRI) is crucial for in vivo detection and characterization of white matter lesions (WMLs) in multiple sclerosis. While WMLs have been studied for over two decades using MRI, automated segmentation remains challenging. Although the majority of statistical techniques for the automated segmentation of WMLs are based on single imaging modalities, recent advances have used multimodal techniques for identifying WMLs. Complementary modalities emphasize different tissue properties, which help identify interrelated features of lesions.

Methods: Method for Inter-Modal Segmentation Analysis (MIMoSA), a fully automatic lesion segmentation algorithm that utilizes novel covariance features from intermodal coupling regression in addition to mean structure to model the probability lesion is contained in each voxel, is proposed. MIMoSA was validated by comparison with both expert manual and other automated segmentation methods in two datasets. The first included 98 subjects imaged at Johns Hopkins Hospital in which bootstrap cross-validation was used to compare the performance of MIMoSA against OASIS and LesionTOADS, two popular automatic segmentation approaches. For a secondary validation, a publicly available data from a segmentation challenge were used for performance benchmarking.

Results: In the Johns Hopkins study, MIMoSA yielded average Sørensen-Dice coefficient (DSC) of .57 and partial AUC of .68 calculated with false positive rates up to 1%. This was superior to performance using OASIS and LesionTOADS. The proposed method also performed competitively in the segmentation challenge dataset.

Conclusion: MIMoSA resulted in statistically significant improvements in lesion segmentation performance compared with LesionTOADS and OASIS, and performed competitively in an additional validation study.

Keywords: Automatic segmentation; lesion detection; logistic regression; multiple sclerosis.

Conflict of interest statement

Disclosure

All authors do not have any relevant financial conflicts of interest to disclose.

Copyright © 2018 by the American Society of Neuroimaging.

Figures

Figure 1

Features for MIMoSA including normalized images as well as an example of the T2-weighted Fluid-Attenuated Inversion Recovery (FLAIR) smoothed volumes (Gaussian smoother using σ=10mm,20mm) and a map for T1-weighted (T1) on FLAIR inter-modal coupling regression slopes.

Figure 2

The inter-modal coupling (IMCo) pipeline is shown below. Both X, Y, and the subject-level map are for illustrative purposes only and do not depict data collected. We create IMCo for each pair of imaging modalities allowing each modality in the pair to be Y on X and X on Y. The MIMoSA model utilized the intercepts and slopes generated as features therefore we calculate 2 subject-level maps. While the subject-level map is for illustrative purposes, the coloring represents spatial variability where some regions have higher versus lower slopes or intercepts within subject.

Figure 3

Probability maps for MIMoSA and OASIS as well as the difference (MIMoSA-OASIS) are shown in the second row. Using the thresholding algorithm, lesion segmentations for respective models are also shown in row 1 along with LesionTOADS hard segmentations. LesionTOADS and manual segmentations are also shown. The green box indicates a false negative result for OASIS and the blue box indicates a false positive result for LesionTOADS.

Figure 4

In the first row we compare Sørensen-Dice coefficient (DSC) scores for the Challenge Data when using rater 1 and rater 2 as the manual segmentation, respectively. In the second row we compare MIMoSA segmentation volume with manual segmentation volume reported by the Challenge in mL. Each column depicts the models trained using rater 1, rater 2, the label-fused STAPLE segmentations, and the predicted segmentations from training on rater 1 and rater 2 label-fused post prediction. Solid lines represent y=x while dashed lines indicate the fitted linear regressions also provided as equations in the bottom-right of each panel.