Assessing the Accuracy of Non-Rigid Registration With and Without Ground Truth
We present two methods for assessing the performance
of non-rigid registration algorithms. We also show that assessment
can be carried out with or without the need for some form of ground
truth. One method utilizes a measure of overlap among data labels.
The other method exploits the fact that, given a set of non-rigidly
registered images, a generative statistical appearance model can be
constructed. The quality of the model depends on the quality of the
registration, and can be evaluated by comparing images sampled from
it with the original image set. We derive indices of model specificity
and generalisation, as well as introduce a formulation for overlap
among anatomical labels. We show that all of them demonstrate the loss of
registration as a set of correctly registered images is progressively
perturbed. Finally, we compare the sensitivities of these methods.