Evaluation of nine similarity measures used in rigid registration

Since image registration is carried out by optimising a criterion function or similarity measure, the behaviour of a similarity measure greatly influences the final result of registration. In this paper, we describe a protocol to quantitatively evaluate the behaviour of similarity measures. Registered MR/CT images of the spine and MR/MR brain images were used to evaluate nine similarity measures. Besides, the effects of binning and noise on the similarity measures were studied. The normalised and non-normalised mutual information were the two measures that performed the best.