A Protocol for Optimization-Independent Similarity Measure Evaluation

Abstract

Evaluation of a registration method is a complex and application-dependent task. The accuracy and robustness of registration depends on a number of factors, such as image acquisition protocols and parameters, image content, spatial transformation, similarity measure, and optimization. The complex interdependence of these factors makes the assessment of a particular factor on registration difficult even for very specific registration tasks. This paper deals with the evaluation of similarity measures. To reduce the degree of complexity or uncertainty in similarity measure evaluation, we propose an evaluation protocol that enables optimization-independent evaluation. Given the image data and parametric spatial transformation, similarity measure values are sampled equidistantly along random lines in I-dimensional parametric space. The obtained similarity measure profiles are then used to derive statistical estimation of capture range and smoothness of the similarity function and the accuracy, precision, and distinctiveness of its optimum. The proposed protocol is used to evaluate three similarity measures, i.e. mutual information, normalized mutual information, and histogram energy on simulated 2D T1, T2, and PD MR brain volumes. The protocol may be a useful tool, first, for studying the influence of many implementation issues on the similarity function and, second, for selecting the best combination of similarity measure and corresponding optimization method for specific applications.