Research On Brain MR Image Segmentation Based On Multi-atlas Label Fusion

Brain MR image segmentation is to identify the area or boundary of the brain tissue.There are a lot of semi-automatic and automatic MR image segmentation methods,in which the semi-automatic segmentation method needs manual intervention,which is time-consuming and laborious,and the repeatability is poor.The existing automatic segmentation methods mainly include graph cut based method,fuzzy clustering based method,multi-atlas based method,etc.The multi-atlas based method is currently considered one of the best methods that has best segmentation performance.The problem of multi-atlas label fusion is a very important step and has an important impact on the segmentation results.Aiming at the problem of the calculation of label weight in the process of label fusion does not make full use of label information,a new label fusion method based on label patch sparse representaton is proposed.The feature of this method is that it makes full use of label information in the process of label weight calculation.The label information of the atlas often contains abundant information of boundary and region,in the calculation of label weight,the approximate segmentation result of the image to be segmented can be sparse represented by a dictionary composed of atlas label patches,thereby the sparse coefficients can be used as the label weights.In this process,we not only make full use of the rich information contained in the label of the atlas,but also achieve the purpose of correcting the approximate segmentation result of the image to be segmented,especially in the boundary of the target area.The experimental results show that it can improve the segmentation accuracy of brain MR image by introducing the label information in the process of label weight calculation.In the process of label weight calculation of traditional label fusion method only the image information is used but the label information is ignored,and in the label patch sparse representation label fusion method only the label information is used but the image information is not used.In order to use the image information and label information at the same time we propose a label fusion method via combined information based on image information and label information sparse representation.In this method,the combined information based on image information and label information is combined through the improved local binary pattern.The specific processing steps are,at first the corresponding voxel of each label that participate the label fusion process of unlabeled target voxel is represented as a combination information,and the combination information of the unlabeled target voxel can be sparse represented by a dictionary composed of combination information of each voxel,finally the sparse coefficients are used as label weights for label fusion.The experimental results show that the proposed method can solve the problem that the calculation of label weight is dependent on the information of the image or the information of the label of atlas and the combination of these two kinds of information can be used to improve the reliability of label weight and can improve the segmentation accuracy of brain MR image.Aiming at the problem of interpolation in the process of label mapping,a new method for the representation of the weight of label based on probability map is proposed,and a new solution for two-fold weighting of the label is proposed in the label fusion process,which makes the weight of the label more scientific and reliable.The experimantal results show that the segmentation effect of brain MR image is improved by introducing the method of two-fold label weighting.In order to solve the problem in binary processing of the result of label fusion in label fusion method based on label weighting does not take into account the distribution characteristics of the target,a new label fusion method based on label weighting is proposed to solve the problem,which is based on the distribution of the target population.The experimental results show that the proposed method can effectively improve the segmentation effect of brain MR images with the same label weight.Several label fusion methods mentioned above,effectively solve the problem that the calculation of label weight only depends on image information but ignores label information of the atlas,and the binary processing of label fusion result has no reference to the overall distribution characteristics of the target.The segmentation performance of brain MR image segmentation method based on multi-atlas label fusion is improved.