PDF version of this entire document
In this chapter, NRR of a brain dataset is studied in order to show that Specificity and Generalisation are valid measures of the degree of misregistration of a group of images. It is expected that, as registration is degraded, Specificity and Generalisation should change accordingly. Experiments were performed to validate the approach for evaluating NRR, making use of ground-truth annotation to provide an independent measure of misregistration.
Taking a set of registered images, for which ground-truth labels were available, a series of controlled deformations were applied. Those deformations were chosen in such a way so that they change whole images smoothly and still move boundaries significantly. It is reasonable to have smooth warp fields in registration, so I applied similar warps to move them away from the correct answer. This produced progressively-increasing misregistration and made it possible to investigate how measures of Specificity and Generalisation varied as a function of the known misregistration. Generalised overlap (see Subsection 2.4.2) was also measured for each of the deformed image sets, using the ground-truth labels, to provide a comparison.
