A Bayesian Estimation Formulation to Voxel-Based Lesion-Symptom Mapping

Studying brain-injured patients is important for investigating structure-function relationships using neuroimaging techniques. Voxel-based lesion-symptom mapping (VLSM) has increasingly been advocated as a relevant approach to detect structure-function associations in neuroimaging studies. The VLSM method involves mapping the relationship between brain injuries and behavioral performance on a voxel-by-voxel basis. This means that the statistical relationship be-tween damage and behavior (across patients) is computed separately for each voxel. However, one could expect voxels characterizing group differences to be localized into spa-tially consistent regions rather than randomly distributed over the brain. Thus, in this paper, we propose to depart from conventional models to characterize and exploit this spatial consistency. More precisely, we derive a Bayesian model that explicitly accounts for spatial correlations be-tween neighboring voxels using a Markov random field. Our results highlight that the proposed approach outperforms the conventional ones. Besides, it has the great advantage of possibly reducing the number of patients and identifying new language areas, which are two crucial insights in the targeted medical context.