To conduct our experiments we used two different sets of MR images of the brain. The first, which we will refer to as the 'MGH Dataset' (see Acknowledgements), was a set of 2D transaxial mid-brain slices, extracted at an equivalent level from each of a set of affinely aligned T1-weighted 3D MR scans of 76#76 normal subjects. As well as the images themselves, we had access to ground-truth data, in the form of dense (pixel by pixel) anatomical label maps for the gray and white matter, the caudate nucleus, and the lateral ventricles. These labels were further divided into left and right hemispheres. The anatomical labels were obtained by manual annotation under conditions of rigorous quality control. An example image and the corresponding label maps are shown in Figure .
The set of images was non-rigidly registered using a Minimum Description Length (MDL) NRR algorithm , and this registration was used as the starting point for a systematic evaluation of the effects of misregistration.
The second set of images, which we will refer to as the 'Dementia Dataset', consisted of a set of 2D transaxial mid-brain slices, extracted at an equivalent level from each of a set of affinely aligned T1-weighted 3D MR scans of 78#78 subjects entered into a clinical study of dementia.