Multimodality Medical Image Registration By Combining Mutual Information And Gradient Information

For many advantages，Image registration technology based on mutual informationis widely employed in multimodality medical image registration.Although differentmodal medical images to be registrated are different in gray level,the imaging objectis the same organ or issue.when the two images are aligned exactly, the mutualInformation is maximized. So the best registration position is the place where themutual Information is maximization.Methods based on the maximization of mutualinformation image registration chieve higher robustness and registration accuracythan those based on geometric features.However, the mutual information have twomain drawbacks in medical image registration, Firstly, mutual information is affectedby the overlapped portion between the images.We can deal this problem bynormalizing mutual information.Secondly, mutual information has an inherent defectin the expression of spatial information because of ignoring the spatial information ofthe image, which may fail to properly reflect the relationship between the images.Asthe defect mentioned may result in mismatches, we include gradient information to fixthis deficiency. This paper presents a new approach that combines NMI and gradientinformation as registration measure and uses genetic algorithm and Powell algorithmas a hybrid algorithm to optimize the registration parameters. Analyzing of thecharacteristics of the new registration function, we can see that new registrationfunction has a certain improvement in smoothness than traditional mutualinformation,and still shows a local extreme characteristic that leads local optimizationalgorithm,such as Powell,to local extreme with a poor accuracy or causes thealgorithm fail to work out the registration parameters. To solve this problem, we use aglobal optimization strategy for finding the optimal value for new similarity measure.The hybrid optimization algorithm mentioned above consist of genetic algorithm andPowell algorithm.It use genetic algorithm to finding a global registration pose underconditions of low precision, then give rough global registration as initial point ofPowell algorithm. Hybrid optimization algorithm integrated the strong globaloptimization ability of genetic algorithm and the strong local extremum searchcapabilities of Powell algorithm,and could converges to the global optimal value withhigh accuracy.Experiments are presented that demonstrate using a new hybridoptimization strategy the approach in this paper yields better accuracy and precisionof registration with the addition of the spatial information and the normalized mutualinformation.