Abstract:
|
The high dimensionality of the longitudinal neuroimaging data poses significant challenges for the traditional statistical methods. Recent proposals based on linear mixed effects models or generalized estimation equations either fail to account for the spatial correlation or involve a complex segmentation procedure for further statistical analysis. In our work, we propose a fast and accurate two-step approach to model brain structure longitudinal changes and their dependence on baseline covariates. In the first step, we estimate the voxel-wise brain volume change rate per subject using a linear regression model. In the second step, we model the estimated rate of change as a function of subjects' characteristics via general linear regression accounting for the spatial correlation. Specifically, we assume that each voxel is only correlated with voxels in its small neighborhood and the correlation magnitude depends on their distance. Using the longitudinal tensor-based morphometry data collected on chronically HIV-infected patients, we found that the ventricles expanded 0.5% to 1% faster than the other brain areas, which agrees with the scientific hypotheses about the brain atrophy.
|