Effect Of Gray-scale Pre-segmentation On3D Medical Image Registration

Medical image registration is a fundamental task of medical image processing, and the goalof registration is to determine the geometric transformation between the two images to make suretheir corresponding points can coincide on the same coordinate. The registration based on mutualinformation is widely used and doesn’t need any preprocessing; it makes use of the gray-scaleinformation, but local extrema is one of the difficult problems that still trouble the imageregistration. In this paper, the research is focus on the registration based on mutual information,and the work is as follows:(1) This paper firstly introduces the theory of image registration. As we know, imageregistration consists of three parts: registration measure, interpolation method and optimizationalgorithm. These three parts interrelated with each other, and influence each other. In this study,the local extrema caused by the partial volume interpolation method is discussed, and also thelocal extrema caused by noise is analyzed, so the causes of the local extrema are clear.(2) Considering the causes of the local extrema, gray-scale pre-segmentation is proposed.This paper discusses how the pre-segmentation can avoid local extrema, and also it verifies theapplicability of several segmentation methods, such as differential method, classic Otsu method,multi-threshold Otsu method and Gaussian mixture model segmentation algorithm.(3) At last, mono-modal registration and multimodal registration are conducted in order toverify the feasibility of the method, experiment results shows the gray-scale pre-segmentationcan avoid the local extrema, and it can also increase the mutual information of the two images.At the same time, the influence of binary pre-segmentation and multi-threshold per-segmentationis studied, experiment results show that they both can avoid the local extrema of mutualinformation, but multi-threshold pre-segmentation is easily influenced by noise, so its influenceis related to the images in registration.Finally, a conclusion is made and the future research directions in this field are proposed.