Robust Multi-modal Image Registration Based on Prior Joint Intensity Distributions and Minimization of Kullback-Leibler Distance
Albert C. S. Chung, Rui Gan, William M. Wells · 2007
Robust registration is essential for image-guided therapy as well as structural and functional analysis. In this work, we present a new method that can give superior robustness in multi-modal image registration. This method is based on the a priori knowledge of the joint intensity distributions between image pairs at different image resolutions and the Kullback-Leibler distance (KLD) similarity measure. Expected joint distributions are estimated from pre-aligned training images. Two image volumes are registered when the value of KLD is minimized. Two thousand randomized registration experiments on clinical brain CT – T1 image pairs from the Retrospective Image Registration Evaluation (RIRE) project have been performed and evaluated independently by the project. The results demonstrate that, as compared with the conventional Mutual Information (MI)-based method and the Normalized Mutual Information (NMI)-based method, the proposed KLD-based method can significantly increase the registration success rates. To increase the registration accuracy, we further propose a refinement as the last step of the KLD-based method. The refinement step can be based on either MI or NMI, namely the KLD-MI-based method and KLD-NMI-based method respectively. Experimental results on CT – T1 image pairs show that the KLD-MI-based and KLD-NMI-based methods consistently give higher registration accuracy than the KLD-based method. The success rates of the KLD-MI-based and KLD-NMI-based methods are high. In addition, the effects on the performance of KLD-based methods under different histogram bin sizes, intensity inhomogeneity and different noise levels are analyzed using simulated BrainWeb T1 – T2 image pairs.