We conducted a series of experiments to test the hypothesis
that reduced registration accuracy can be detected using model specificity
and generalisation. An equivalent 2D mid-brain T1-weighted slice was
obtained from each of 36 subjects using a 3D acquisition. A fixed
number (167) of landmark points were positioned manually on the cortical
surface, ventricles, caudate nucleus and lentiform nucleus, and used
to establish a ground-truth dense correspondence over the set of images,
using locally affine interpolation. A statistical appearance model
was constructed using the methods described in 4.3, with the set of
landmark coordinates forming the shape vector for each
image. Keeping the shape vectors fixed, we then applied a series of
smooth pseudo-random spatial warps to the training images, resulting
in successively increasing mis-registration. Each warp resulted in
an average point displacement of between one and two pixels. Specificity
and Generalisation results were obtained for 0, 1, 5, and 10 warps
per image, using
.