PDF version of this entire document
A single NRR algorithm yields a particular solution to a problem, and different NRR algorithms tend to produce different results. It is hence necessary to have some way of quantifying the results of registration, to give some standard by which to compare the performance of various algorithms, or to guide improvement of a single algorithm.
Two main approaches to assessing the accuracy of NRR algorithms have been described in the literature - one based on the recovery of known deformation fields, the other based on measuring the overlap of ground-truth annotations after registration. Both approaches are valid, but neither is easy to apply routinely, and both are better suited to off-line evaluation of algorithms, rather than in-line evaluation of the results of NRR in practical applications [,,].