Abstract
Optimally-Discriminative Voxel-Based Analysis (ODVBA) (Zhang and Davatzikos, 2011) is a recently-developed and validated framework of voxel-based group analysis, which transcends limitations of traditional Gaussian smoothing in the forms of analysis such as the General Linear Model (GLM). ODVBA estimates the optimal non-stationary and anisotropic filtering of the data prior to statistical analyses to maximize the ability to detect group differences. In this paper, we extensively evaluate ODVBA to three sets of previously published data from studies in schizophrenia, mild cognitive impairment, and Alzheimer's disease, and evaluate the regions of structural difference identified by ODVBA versus standard Gaussian smoothing and other related methods. The experimental results suggest that ODVBA is considerably more sensitive in detecting group differences, presumably because of its ability to adapt the regional filtering to the underlying extent and shape of a group difference, thereby maximizing the ability to detect such difference. Although there is no gold standard in these clinical studies, ODVBA demonstrated highest significance in group differences within the identified voxels. In terms of spatial extent of detected area, agreement of anatomical boundary, and classification, it performed better than other tested voxel-based methods and competitively with the cluster enhancing methods.