Abstract
This paper describes a novel approach to the spatial filtering of deformation tensor derived maps of shape difference and shape change estimated in multi-subject studies of spatially normalised brain anatomy. We propose a spatial shape filter that combines tensor values locally which fall within regions of similar underlying intensity and therefore tissue. The filter additionally incorporates information derived from the spatial normalisation process to focus filtering more strongly within specific MRI intensities in regions where tissue intensities have been most consistently aligned in the spatial normalisation process. This is achieved using a statistical framework to introduce a measure of uncertainty of regional intensity correspondence. Results comparing the approach to conventional Gaussian filtering in the analysis of tensor derived measures of brain shape change in Alzheimer’s disease and normal aging indicate significantly improved delineation of local atrophy patterns, particularly in cortical gray matter, without the need for explicit tissue segmentation.