Evaluating Non-Rigid Registration without Ground Truth

We present a generic method for assessing the quality of non-rigid registration (NRR), that does not require ground truth, but rather depends solely on the registered images. We consider the case where NRR is applied to a set of images, providing a dense correspondence between images. Given this correspondence, it is possible to build a generative statistical model of appearance variation for the set. We observe that the quality of the resulting model will depend on the quality of the correspondence. We define measures of model specificity and generalisation that can be used to assess the quality of the model and, hence, the quality of the correspondence from which it is derived. The approach does not depend on the specifics of the registration algorithm or the form of the model. We validate the approach by measuring the change in model quality, as the correspondence of an initially registered set of MR images of the brain is progressively perturbed, and compare the results with those obtained using a method based on the overlap of groundtruth anatomical labels. We demonstrate that, not only is the proposed approach capable of assessing NRR reliably without ground truth, but that it also provides a more sensitive measure of misregistration than the overlap-based approach. Finally we apply the new method to compare the performance of repeated pairwise and fully groupwise registration of MR images of the brain.