Evaluation of Four Similarity Measures for 2D/3D Registration in Image-Guided Intervention

2D/3D medical image registration in image-guided intervention is crucial to assist the clinician to establish the space corresponding relationship between image information and patients' anatomy, which can be quantified by a similarity measure. Among similarity measures, mutual information and its derivatives, were used widely for image registration, and showed significantly differences in the performances of registrations. However, the comparison of their registration performances has not been studied quantitatively yet. Therefore, in this paper, four similarity measures were evaluated for 2D/3D rigid registrations, which are mutual information (MI) and its three derivatives (distance coefficient mutual information (DCMI), distance weighted mutual information (DWMI), gradient weighted mutual information (GWMI)). They were applied to implement registrations based on porcine skull phantom datasets from the Medical University of Vienna, and were evaluated through the mean target registration errors (mTRE) for the registrations. The results demonstrated that the performance of DCMI was the most accurate and robust, and MI was the least effective of the four similarity measures. Moreover, due to the presence of a great amount of soft tissues, GWMI also had the low performance with its mean of mTRE even greater than that by MI, which suggested that intensity gradients were not always having a positive impact for 2D/3D rigid registration when involving a great amount of soft tissues. Between DCMI and DWMI, there were a significant difference in terms of accuracy and robustness, despite using the same image information for them, which means that the construction of an ideal measure should consider not only the image information to be involved but also the construction way of these information.