Image registrations that are based on similarity measures simply adjust the parameters of an appropriate spatial transformation model until the similarity measure reaches an optimum. The numerous similarity measures that have been proposed in the past are differently sensitive to imaging modality, image content and differences in the image content, selection of the floating and target image, partial image overlap, etc. In this paper, we evaluate and compare 12 similarity measures for the rigid registration. To study the impact of different imaging modalities on the behavior of similarity measures, we have used 16 CT/MR and 6 PET/MR image pairs with known 'gold standard' registrations. The results for the PET/MR registration and for the registration of CT to both rectified and unrectified MR images indicate that mutual information, normalized mutual information and the entropy correlation coefficient are the most accurate similarity measures and have the smallest risk of being trapped in a local optimum. The results of an experiment on the impact of exchanging the floating and target image indicate that, especially in MR/PET registrations, the behavior of some similarity measures, such as mutual information, significantly depends on which image is the floating and which is the target.

In this paper, we present a protocol for the evaluation of similarity measures for non-rigid registration. The evaluation is based on five intuitive properties that characterize the behavior of a similarity measure, i.e. the accuracy, capture range, distinctiveness of the optimum, number of local minima, and risk of non-convergence. These five properties are estimated locally from similarity measure values that correspond to a range of systematic local free-form deformations, obtained by displacing control points in random directions from the gold standard position. Global similarity measure properties are obtained by combining the local properties over image regions or over the entire image. The feasibility of the proposed evaluation protocol is demonstrated for three similarity measures: mutual information, normalized mutual information and correlation ratio. The evaluation is carried out on a number of MR and CT images: a pair of simulated MR T1 and MR T2 images of the head, three pairs of real MR T1 and T2 images of the head, six pairs of real MR T1 and CT images of the head, and pairs of MR and CT images of three vertebrae. The protocol may help researchers to select the most appropriate similarity measure for a non-rigid registration task.

The accuracy and robustness of a registration method depend on a number of factors, such as imaging modality, image content and image degrading effects, the class of spatial transformation used for registration, similarity measure, optimization, and numerous implementation details. The complex interdependence of these factors makes the assessment of the influence of a particular factor on registration difficult, although it is often desirable to have some estimate of such influences prior to registration. The similarity measure used to create the cost function is one of the factors that most influences the quality of registration. Traditionally, limited information on the behavior of a similarity measure is obtained either by studying the quality of the final registration or by drawing plots of similarity measure values obtained by translating or rotating one image relative to the "gold standard." In this paper, we present a protocol for a more thorough, optimization-independent, and systematic statistical evaluation of similarity measures. This protocol estimates a similarity measure's capture range, the number, location and extent of local optima, and the accuracy and distinctiveness of the global optimum. To show that the proposed evaluation protocol is viable, we have conducted several experiments with nine similarity measures and real computed tomography and magnetic resonance (MR) images of a spine phantom, MR brain images, and MR and positron emission tomography brain images, for which "gold standard" registrations were available. We have also studied the impact of histogram bin size on the behavior of nine similarity measures. The proposed evaluation protocol is useful for selecting the best similarity measure and corresponding optimization method for a particular application, as well as for studying the influence of sampling, interpolation, histogram bin size, partial image overlap, and image degradation, such as noise, intensity inhomogeneity, and geometrical distortions on the behavior of a similarity measure.

Purpose: A promising patient positioning technique is based on registering computed tomographic (CT) or magnetic resonance (MR) images to cone-beam CT images (CBCT). The extra radiation dose delivered to the patient can be substantially reduced by using fewer projections. This approach results in lower quality CBCT images. The purpose of this study is to evaluate a number of similarity measures (SMs) suitable for registration of CT or MR images to low-quality CBCTs. Methods and Materials: Using the recently proposed evaluation protocol, we evaluated nine SMs with respect to pretreatment imaging modalities, number of two-dimensional (2D) images used for reconstruction, and number of reconstruction iterations. The image database consisted of 100 X-ray and corresponding CT and MR images of two vertebral columns. Results: Using a higher number of 2D projections or reconstruction iterations results in higher accuracy and slightly lower robustness. The similarity measures that behaved the best also yielded the best registration results. The most appropriate similarity measure was the asymmetric multi-feature mutual information (AMMI). Conclusions: The evaluation protocol proved to be a valuable tool for selecting the best similarity measure for the reconstruction-based registration. The results indicate that accurate and robust CT/CBCT or even MR/CBCT registrations are possible if the AMMI similarity measure is used.