A very effective strategy for placing a boundary on true registration accuracy is the use of internal inconsistencies generated by the registration method itself. In the simplest case, image A can be registered to image B and the inverse of the resulting transformation can be compared to the result of registering image B to image A. Some registration methods are deliberately designed to produce transformations that are exact inverses of one another in this context, and a slightly more complex strategy involving three images is required. In this case, image A is registered to image B, and image B is registered to image C. The results of these two registrations are combined to compute the transformation needed to register image A to image C. The algorithm is then used to register image A to image C directly, and the two estimates of this transformation are compared. To the extent that these two estimates differ, the algorithm must have been inaccurate. By assuming that the inaccuracies are optimally distributed among the three registrations performed, a best-case estimate of registration inaccuracy can be obtained. Unlike many of the other validation methods discussed here, these inaccuracies cannot overestimate the true errors since they are based on real data and involve no external standards. Use of this validation strategy is illustrated in Woods et al. .
Was this article helpful?