You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster E80, Friday, November 10, 10:00 – 11:15 am, Harborview and Loch Raven Ballrooms

Enhanced accuracy of lesion to symptom mapping with multivariate sparse canonical correlations

Dorian Pustina1, Brian Avants2, Olufunsho Faseyitan1, John Medaglia3, H. Branch Coslett1;1Department of Neurology, University of Pennsylvania, 2Department of Radiology, University of Pennsylvania, 3Department of Psychology, University of Pennsylvania

INTRODUCTION: Lesion to symptom mapping (LSM) is a critical tool for making inferences about the causality of brain-behavior relationships. LSM analyses are typically performed by running independent tests at each individual voxel (VLSM), also called “the mass-univariate approach”. Recent studies show that VLSM might produce displaced or inaccurate results (Mah 2014, Sperber 2017, Zhang 2015). Using a different approach, multivariate methods consider all the voxels as potential predictors of behavior. From a conceptual standpoint this approach may improve the mapping of functional areas. To evaluate this hypothesis, we developed and tested a multivariate LSM method based on Sparse Canonical Correlation Analysis for Neuroimaging (SCCAN; Avants 2010, 2014). Results obtained from simulated and real data using SCCAN and VLSM were compared in their ability to detect the putative functional brain units. METHODS: Lesion maps from 131 patients with left hemispheric chronic stroke were included in the study (age: 58.2+/-11.3, months post-onset: 44+/-63, lesion size: 100+/-82 ml). To achieve simulations from known functional brain areas, we first merged three atlases (Glasser 2016, Fan 2016, Hua 2008). The lesion load of each region in the combined atlas was used to simulate the behavioral deficit of each subject. A substantial amount of noise (50%) was injected in behavioral simulations to match the typical noise of real brain-behavior relationships. LSM analyses were performed either with univariate (non-parametric Brunner-Munzel tests) or multivariate (SCCAN) methods. Several scenarios were tested, including functional units composed of one, two, and three brain regions, different sample sizes (N=20-131), and different univariate thresholding mechanisms - Bonferroni, false discovery rate (FDR), permutation-based family wise error correction (FWER). The accuracy of each method was assessed by computing (1) dice overlap, (2) average displacement, (3) center of mass displacement, and (4) peak voxel displacement, between the simulated brain region and the LSM statistical map. The accuracy scores of VLSM and SCCAN results were compared with paired Wilcoxon tests. RESULTS: SCCAN produced more accurate results compared to VLSM, typically consisting in higher dice overlap and smaller average displacement (all p<0.001, 93 regions tested, Figure 1). This advantage persisted at different sample sizes (N=20-131) and with different multiple comparison corrections of VLSM (FDR, Bonferroni, FWER, Figure 2). For functional units composed of multiple brain regions SCCAN identified almost all the simulated areas, while VLSM missed more simulated areas. Under no circumstance could VLSM exceed the accuracy obtained with SCCAN. Functional mapping of real aphasia scores from the Western Aphasia Battery and the Philadelphia Naming Test revealed known language-critical areas with SCCAN, while VLSM either produced diffuse maps (FDR correction) or few scattered voxels (FWER correction). CONCLUSIONS: The newly developed SCCAN mapping produces systematically more accurate results than VLSM. This method will allow researchers to collectively draw firmer conclusions on the spatial topography of cognitive systems. A new software package is made available for LSM analyses ( This work represents a harmonization between the statistical theory behind construct psychology and multivariate neuroimaging analysis, which defines a frontier of modern cognitive neuroscience.

Topic Area: Methods

Back to Poster Schedule