Thursday, May 26, 2022 • 9:45–11:00 am (CT) • Light Hall, Room 512
Organizer: Elizabeth Sweeney, University of Pennsylvania
Chair: Rajarshi Guhaniyogi, Texas A&M University
Data-driven subtyping of multiple sclerosis lesions on quantitative magnetic resonance imaging
Elizabeth Sweeney, University of Pennsylvania
Jordan Dworkin, Columbia University
Melanie Marcille, Weill Cornell Medical College
Thanh Nguyen, Weill Cornell Medical College
Susan Gauthier, Weill Cornell Medical College
Subtyping of multiple sclerosis (MS) lesions based on their presentation on magnetic resonance imaging (MRI) has historically been done through visual inspection and manual categorization along with validation on histopathology. These approaches have been successful in uncovering meaningful lesion subtypes, the presence of which correlate with patient disability. However, it is likely that there is relevant heterogeneity contained within lesions based on multi-sequence MRI intensity distributions that cannot be characterized by visual inspection. We propose a data-driven method for subtyping MS lesions on quantitative MRI sequence, namely the quantitative susceptibility map (QSM), myelin water fraction (MWF), T1 mapping (T1map) and fractional anisotropy (FA) sequences. In a population of 176 MS patients with 4,186 manually identified lesions, we first estimated lesion-level empirical quantile functions of the lesion intensities on each of the QSM, MWF, T1map, and FA sequence. We performed functional principal component analysis (fPCA) on the empirical quantiles for each sequence to reduce the dimensionality of the data. To subtype lesions, we applied k-means clustering to the fPCA scores and detected four lesion subtypes based using the silhouette method. Next, in a patient-level model, we examined the relationship between disability, as measured by the expanded disability status score (EDSS), and the count of lesions in each subtype in a linear model adjusting for sex, age, disease duration, disease type, and total lesion volume. In the linear model patients' lesion counts in each subtype explained an additional 7% of the variation in EDSS. Two of the four subtypes showed positive associations with EDSS (p = 0.019 and 0.015), suggesting these subtypes may represent more aggressive lesions.
Image analysis approaches to MS diagnostics
Russell (Taki) Shinohara, University of Pennsylvania
Lesions in the white matter of the brain, including those that arise in multiple sclerosis, are abnormalities measurable on MRI. While much literature has focused on identifying these lesions, less work has focused on the nature of these lesions. As new imaging modalities arise that allow us to interrogate these lesions better, new statistical modeling problems that include spatial constraints and overlapping analysis domains are increasingly important. Leveraging multi-modal imaging approaches that focus on knowledge about etiology is critical for developing the next generation of robust and generalizable imaging biomarkers.
Quantifying and testing differential spatial patterns among MS lesion subtypes
Jordan Dworkin, Columbia University
Elizabeth Sweeney, University of Pennsylvania
Thanh Nguyen, Weill Cornell Medical College
Susan Gauthier, Weill Cornell Medical College
The characterization, detection, and assessment of lesion subtypes has become an increasingly important topic in multiple sclerosis imaging. Expert-defined subtypes—like T1 black holes and rim-positive lesions—have shown differential clinical impacts as compared to commonly measured T2-hyperintense lesions; more recent work has also found that data-driven lesion subtyping reveals relevant heterogeneity. Because both expert-driven and data-driven subtypes are derived from image intensity patterns, the extent to which the subtypes diverge along other domains is often unknown. Here, we propose measuring characteristic differences across lesion subtypes using a generalized additive mixture model (GAMM). In addition to measuring subtypes’ associations with traditional measures like lesion size and shape, the multinomial GAMM allows for straightforward assessment of spatial differences in subtypes’ relative prevalence across the brain. The incorporation of subject random effects additionally facilitates tests of between-subject variability in subtypes’ relative prevalence. To illustrate the value of this approach, we present simulation studies comparing its performance to (a) tests of subtype-level summary statistics and (b) voxel-level mass univariate hypothesis testing. We additionally apply this approach to a sample of lesion subtypes derived from quantitative imaging intensities to measure their spatial and structural variability.
