PDF version of this entire document
We conducted a series of experiments to test the hypothesis that reduced registration accuracy can be detected using model entropy. An equivalent 2D mid-brain T1-weighted slice was obtained from each of 37 subjects using a 3D acquisition. Fixed binary labels were positioned manually on the right- and left-hand-side grey matter, white matter, lateral ventricles, and caudate nucleus. These labels were used to establish ground truth for our overlap-based assessment approach.
The images and their accompanying labels were then non-rigidly
registered using a groupwise, minimum description length-based algorithm.
A statistical appearance model was constructed using the methods described
in Section 2. It used the set of landmark coordinates, which had been
extracted from the registration, to form the shape vector 
for each image. We then applied a series of warps, based on biharmonic
clamped-plate splines, to the training images and labels, resulting
in successively decreasing registration. Each warp resulted in increased
displacement, which corresponds to degraded NRR performance. Entropy
results were obtained for a range, using . Corresponding
results were obtained for the overlap-based method, which used the
resampled anatomical labels to assess label overlap.
The intent of these validation experiments was to show that the model-based approach lies in tight agreement with the overlap-based approach, which uses ground truth.