PDF version of this document
We have devised and shown the practicality of a model-based approach to evaluation of non-rigid image registration. To carry out such an evaluation, no ground truth needs to be involved and yet accuracy of the method is comparable with that of methods that exploit known solutions. To show that the method is both practical and accurate, we performed a validation by perturbing the ground-truth and confirming that our evaluation method detects the degree of perturbation.
We then proceeded to evaluating various registration algorithms, most notably the pair-wise and group-wise approach and showed that better results are obtains when the latter gets used.
Thus approach which was developed in the second year is an important advancement, which makes NRR assessment an objective process that does not require ground-truth solutions. The method can therefore be applied readily to evaluate and compare a variety of registration methods, just as it was able to assess and distinguish between pair- and group-wise registration.
In addition to the method which is able to assess registration, we have also developed a model-base objective function, which takes our evaluation criterion and uses it to guide registration. While the approach seems to work in theory, it is rather computationally-expensive and is therefore impractical to consider in 2- and 3-D.