PDF version of this entire document

Data-Driven Evaluation of Non-Rigid Registration and Appearance Models

Roy S. Schestowitz*, Vladimir S. Petrovic, Carole J. Twining, Timothy F. Cootes, William R. Crum, and Christopher J. Taylor

Abstract:

The paper presents a generic approach, which can be used assess the quality of appearance models of the brain. Moreover, this approach is fully capable of assessing and comparing non-rigid registration (NRR) algorithms without exploiting any form of ground truth. We base this approach on the observation that a statistical appearance model can be constructed from a set of non-rigidly registered images. A model can be evaluated by comparing images generated by it with the image set from which it was constructed. The quality of the model depends on the quality of its seminal registration. A registration can also be evaluated by constructing and evaluating models that it produces. Indices are derived which reflect on model specificity and generalisation. We show that these indices are surrogates of Shannon's entropy, which can in itself be used to assess NRR. All of these measures are negatively affected as a set of correctly registered images is progressively perturbed. We compare our results against those which are obtained using overlap-based NRR assessment, which is based on ground-truth anatomical labels. We demonstrate that not only is our approach capable of assessing NRR without ground truth, but it is also more sensitive than the ground-truth-dependent approach. Finally, to demonstrate the practicality of our method, different NRR algorithms - both pairwise and groupwise- are compared in terms of their performance on MR brain data.