Medical Image Segmentation Based On Multi-atlas Registration

Medical image segmentation has been widely used in clinical diagnosis, pathological analysis, surgical planning, image information processing, computer-assisted surgery and other areas of medical research and practice. Compared to the time-consuming manual segmentation and semi-automatic segmentation, automatic segmentation has important significance and value in use. Among the many automatic segmentation algorithms, atlas-based segmentation algorithm has outstanding performance, its use on the brain image segmentation is the recent research focus. Atlas-based segmentation divides into two types:a single atlas-based segmentation and segmentation based on multi-atlas, the former is mainly used to study the relationship between the structures or areas, the latter is mainly used for measuring the volume and morphological studies and has higher accuracy than the former. This article is write to segment brain structure using multi-atlas segmentation algorithm (MAS).An atlas is defined as a gray image and its corresponding manual segmentation image. First step of MAS is image registration. The algorithm sets the image to be segmented as the reference image, atlases as floating images, then the registration parameters obtained is used to deform the manual segmentation, the labels on the deformed segmentation images are mapped to the target image coordinate, fusion these labels will get the final image. An MAS algorithm can be subdivided into eight components:generation of atlases, offline learning, registration, atlas selection, label propagation, online learning, label fusion and post-processing. While realizing the whole segmentation process, the paper gives a detail description about registration, label fusion and post-processing, because they make important decision on the finial accuracy.Registration is the process of making an image matching with another image by the same anatomical points, the substance of the registration is a process finding deformation parameters. Registration algorithm has four elements:feature space, the search space, similarity measure and search strategy. Using image gray level information, this paper chooses the normalized mutual information as similarity measure, uses Powell algorithm to find the similarity transformation parameters globally between the gray image and the image to be segmented.Among label fusion algorithms, those that derive weight from local similarity between the atlas and the image to be segmented have been most successful in practice. The paper makes an analysis on the traditional weighted fusion algorithms, inherits their advantages and takes the relationship between atlases into consideration, and uses the local search strategy to further improve the accuracy of the algorithm.Segmentation error occurred primarily due to systematic errors, post processing of this paper is performed using classifier constructed on Adaboost. The classifier learns systematic errors from training data, and then is used to detect and correct segmentation errors. Experiments show this strategy can significantly improve the accuracy of segmentation. This algorithm can be applied to a wide range of image segmentation problems to correct systematic errors and improve accuracy.