Research On Segmentation Methods For Brain Megnetic Resonance Image Based On Multi-Atlas

Multi-atlas segmentation is one of the most widely used and successful automatic segmentation methods for brain MR images in biomedical field.The multi-atlas based segmentation method mainly includes atlas generation technology,registration technology,atlas selection technology,label propagation technology,label fusion technology and postprocessing technology,etc.This paper mainly studies the image registration technology,label fusion technology and atlas selection technology in the multi-atlas segmentation methods.The traditional image registration methods only use the grey scale value of the target itself in the target image to be segmented,without taking full advantage of the grey scale features of the target and its surrounding tissues,which is inconsistent with the fact that the same brain tissue may have different grey scale values and the same grey scale value may represent different brain tissues in brain MR images.At the same time,traditional multiatlas segmentation methods treat all voxels in brain MR images equally,without considering the characteristics of the target image to be segmented,and do not make full use of the spatial location information of brain tissue in brain MR images.To address these problems,a new target-oriented brain tissue segmentation method is proposed.The method adopts the segmentation region where the target image to be segmented is centred,and through the grey-scale transformation of the target image to be segmented and its corresponding atlas image,not only the grey-scale features of the image are fully utilised,but also the spatial location information of the image is effectively utilised,solving the problem that the traditional multi-atlas segmentation-based method does not fully consider the alignment accuracy of the region to be segmented in the image alignment stage.The experimental results show that the method can effectively improve the segmentation accuracy of brain MR images based on multi-atlas.The traditional atlas labels fusion methods do not consider those labels that may be anomalous as well as only the similarity factor between the atlas image and the image to be segmented in the calculation of label weights,without using the information of the boundaries and regions of the atlas label image.To address this problem,a new multi-atlas label fusion method based on similarity weighting is proposed.The method extracts the combined information of voxels to be labelled and the combined information of atlas voxels separately,and calculates the similarity of these two types of information as the weights of each atlas label involved in label fusion,and then reduces the influence of abnormal atlas labels on the weighted fusion effect of multi-atlas labels by means of abnormal atlas label detection rules.The method makes full use of the information of the atlas image,the boundary and region of the target image to be segmented and the atlas label image when calculating the weight of each atlas label participating in the label weighted fusion,and also solves the problem of the influence of abnormal atlas labels on the effect of multi-atlas label fusion well.The experimental results show that the method can achieve accurate automatic segmentation of brain tissue in brain MR images.The traditional atlas selection methods select a number of atlases from the atlas with the highest similarity to construct a subset of atlases,thus overcoming the interference caused by the diversity of the atlas.These methods only consider the correlation between the atlases and do not take into account the redundancy within the atlas.At the same time,these methods calculate the similarity between the atlas images and the target images in the atlas set as a whole,without considering the local nature of the target segmentation.To address these problems,a brain MR image segmentation method based on a two-step atlas selection strategy is proposed.The method solves the problems that the traditional atlas selection method only considers the similarity between images without considering the redundancy within the images and that the traditional rectangular region based method does not consider the original shape characteristics of the target tissue when calculating the similarity of the segmented region.The experimental results show that the similarity results obtained by this method are not affected by the target tissue size and target tissue location,and have higher robustness,which can effectively improve the segmentation accuracy of the multi-atlas-based method.