The Research Of Brain MR Image Registration Technique Oriented By Segmentation Target

Medical image registration is an image processing technique which matches two medical images to make the anatomical position aligned in same space. It is a key procedure in image segmentation of human tissues and organs. As the hippocampus is a very vital organ, the accurate segmentation is quite important mean to disease diagnosis about hippocampus. Since the shape of hippocampus is complicated, the contour between it and its adjacent organs is fuzzy and their intensities are close to each other in MRI, the registration of hippocampus image is a difficulty in the field of image registration.Image registration involves several steps such as image preprocessing, space transformation, similarity measurement, and iterative refinement. Among them, similarity measurement is the critical factor in registration accuracy. However, the classic registration method ignores the similarity information of target position and size. If those target information can be taken into accounted in registration during segmenting of hippocampus and other similar organs, the registration result must be helpful to the segmentation. Based on above idea, a registration method oriented by segmentation target using weighted MMI is put forward. It carries out hippocampus segmentation of brain MR Images based on atlases. Firstly, the ROI centered on the target area is set up by constructing an image mask. Secondly, pixels of the target and its neighbor in the ROI are assigned with different weights according to the weighted image. Last, the registration is achieved by taking the MMI of segmentation target as standard instead of the alignment of whole images. This new method actually improves local registration accuracy of the segmentation target while sacrificing global registration accuracy. In addition, a new and modified registration flow based on the features of medical images is proposed, utilizing multi-resolution sampling, compound transformation and so forth.Base on the modified registration method, a 20-round circulated comparative experiment is designed to verify the effectiveness of the novel algorithm above. The experiment results show that the Dice overlap metric of segmentation results has a 3.4% around improvement compared to the classic registration method by using weighted MMI, and the computation time consumes about 100 seconds. So this new registration method is a practical technique.