From publication:
Darko �kerl, Bo�tjan Likar, and Franjo Pernu�, "A Protocol for Evaluation of Similarity Measures for Rigid Registration," IEEE Transactions on Medical Imaging, vol. 25, No. 6, pp. 779791, June 2006.
The protocol enables a thorough, optimizationindependent, and systematic statistical evaluation of important similarity measure properties:
 Accuracy (ACC)
 Distinctiveness of the Optimum (DO)
 Capture Range (CR)
 Number of Local Minima (NOM)
 Risk of Nonconvergence (RON)

All properties are statistical estimations, obtained by randomly sampling the parametrical space and computing the corresponding similarity measure values.
The similarity measure evaluation consists of three steps:
 Step 1: Sampling of the parametrical space (Figure 1).
 Step 2: Computation of similarity measure values.
 Step 3: Computation of similarity measure properties.

Sampling of the parametrical space requires that the �gold standard� registration, normalization of the parametrical space by which the unit for each parameter is defined, and the values of R, M and N are known. Similarity measure values {SM(Xn,m)} are computed in the second step for the set of sampling points {Xn,m} and finally passed to the third step, which computes the properties of the similarity measure.
The first step, i.e. parametrical space sampling, and the third step, i.e. computation of similarity measure properties, do not depend on the specific implementation of a similarity measure. Consequently, the first and the third steps were made publicly available.
For the given registration task, the user selects the dimension of the parametrical space, normalization parameters, number of lines N, number of points on the line M, radius of the hyper sphere R, and gold standard registration parameters. For the selected parameters the set of sampling points {Xn,m} are computed in the first step of the evaluation protocol and sent to the user, which then computes the values of the similarity measure that correspond to the set of sampling points for an arbitrary number of registered images. The user then submits the similarity measure values and the corresponding set of sampling points to the third evaluation step in which the similarity measure properties are computed and returned to the user. An example of how the parameters should be selected is presented for rigid registration of simulated BrainWeb images.
