Hippocampal shape is now at the forefront of brain imaging research. For example, analysis of shape deformations is being used for early diagnosis of dementia of the Alzheimer type (DAT) [2]. A means of quantifying 3D shape deformations and thus performing shape analysis is embedded in the segmentation technique described above.
If one looks at the grey-level and warp models of eqs.1-4 as orthonormal bases spanning n-dimensional spaces (where n is the number of voxels present), the distribution of the projection of training set subjects data in those spaces can be used as a probability function characterizing the type of population in question. A matching statistic - or measure of distance - can be assigned for grey-level images (and/or warp fields) of a new subject projected in the same space. This similarity measure can be associated with a particular pathological difference between the new subject and the training set. The training set could be homogeneous with respect to the absence or presence of the pathology under study. Training set/validation set pairs (e.g. normal/normal, normal/epileptic, epileptic/normal, etc.) can be easily constructed in this fashion. The aim is to derive diagnosis information correlated to a particular Principal Component or distribution in PCA space. This is close in spirit to the work of Csernansky [2].