There is an urgent need for neuroimaging biomarkers of Alzheimer's disease (AD) that correlate with cognitive decline, and with accepted measures of pathology detectable in cerebrospinal fluid (CSF). Ideal biomarkers should also be able to predict future decline, and should be computable automatically from hundreds to thousands of images without user intervention. Here we used our multiatlas fluid image alignment method (MAFIA ), to automatically segment parametric 3D surface models of the lateral ventricles in brain MRI scans from 184 AD, 391 MCI, and 229 healthy elderly controls. Radial expansion of the ventricles, computed pointwise, was correlated with measures of (1) clinical decline, (2) pathology from CSF, and (3) future deterioration. Surface-based correlation maps were assessed using a cumulative distribution function method to rank influential covariates according to their effect sizes. The resulting approach is highly automated, and boosts the power of fluid image registration by integrating multiple independent registrations to reduce segmentation errors.