Similarity metrics for intensity-based registration using breast density maps

Abstract

Intensity-based registration algorithms have been widely used in medical image applications. This type of registration algorithms uses an object function to compute a transformation and optimizes a measure of similarity between the images being registered. The most common similarity metrics used in registration are sum of squared differences, mutual information and normalized cross-correlation. This paper aims to compare these similarity metrics, using common registration algorithms applied to breast density maps registration. To evaluate the results, we use the protocols for evaluation of similarity measures proposed by Škerl et al. They consist in defining a set of random directions in the parameter space of the registration algorithm and compute statistical measures, such as the accuracy, capture range, number of maxima and risk of non-convergence, along these directions. The obtained results show a better performance corresponding to normalized cross-correlation for the rigid registration algorithm, while the sum of squared difference obtains the best result for the B-Spline method.