Population based ranking of frameless CT-MRI registration methods

Background: Clinical practice often requires simultaneous information obtained by two different imaging modalities. Registration algorithms are commonly used for this purpose. Automated procedures are very helpful in cases when the same kind of registration has to be performed on images of a high number of subjects. Radiotherapists would prefer to use the best automated method to assist therapy planning, however there are not accepted procedures for ranking the different registration algorithms. Purpose: We were interested in developing a method to measure the population level performance of CT-MRI registration algorithms by a parameter of values in the [0,1] interval. Materials and Methods: Pairs of CT and MRI images were collected from 1051 subjects. Results of an automated registration were corrected manually until a radiologist and a neurosurgeon expert both accepted the result as good. This way 1051 registered MRI images were produced by the same pair of experts to be used as gold standards for the evaluation of the performance of other registration algorithms. Pearson correlation coefficient, mutual information, normalized mutual information, Kullback-Leibler divergence, L-1 norm and square L-2 norm (dis)similarity measures were tested for sensitivity to indicate the extent of (dis)similarity of a pair of individual mismatched images. Results: The square Hellinger distance proved suitable to grade the performance of registration algorithms at population level providing the developers with a valuable tool to rank algorithms. Conclusions: The developed procedure provides an objective method to find the registration algorithm performing the best on the population level out of newly constructed or available preselected ones.