Research On Automatic Brain MR Image Segmentation Algorithm Based On Multi-Atlas

Brain image segmentation is essential for the diagnosis and treatment of brain diseases,and it is also the basis for three-dimensional reconstruction of brain tissue structure and quantitative analysis of lesions.The accuracy of segmentation directly affects focus tissue localization,lesion shape and size measurement,and clinical diagnosis and treatment planning.Manual segmentation is a commonly used method,but this method is time-consuming and subjective.At present,a large number of brain images are generated every day in the hospital.It will bring a greater burden to doctors if this images are all processed by manual segmentation.Therefore,an automatic segmentation technique is needed to deal with the routine analysis of clinical brain MR images.Many automatic image segmentation methods have been proposed,such as graph-based methods,fuzzy clustering-based methods,multi-atlas based methods,and machine learning segmentation methods.These methods can effectively alleviate the burden of manual segmentation,but their segmentation speed and accuracy still need to be improved.Multi-atlas based image segmentation method can use the priori knowledge of atlas to segment the target images,and it has shown better segmentation performance than other algorithms in brain image segmentation challenge.This method mainly includes three steps:atlas pre-selection,atlas registration and label fusion.First,Using a atlas preselection method to select the atlas that is similar to target image.Second,Using a atlas registration method to map the atlas label information to target image space.Finally,assigning corresponding label to each pixel in the target image through a label fusion step.This paper proposes a target tissue sub-image extraction algorithm,a atlas sub-image pre-selection algorithm based on neural network,a multi-atlas label fusion algorithm combined with pixel grayscale probability information,and a segmentation result optimization method for multi-atlas image segmentation methods.The specific research contents and innovations are as follows:1.Target tissue sub-image extraction algorithmIn the brain image segmentation task,the image areas outside the neighborhood of target tissue are invalid.They not only do not benefit segmentation accuracy,but also reduce segmentation time.For this problem,this paper proposes a target tissue sub-image extraction algorithm.This method extracts the target tissue sub-image based on target tissue position information and gray distribution information of target tissue and entire image.This method can not only reduce the image data,but also improve the local registration accuracy of the target tissue.2.Atlas sub-image preselection algorithm based on neural networkAiming at the problem that atlas sub-image with low similarity to target image participating in label fusion step will reduce segmentation accuracy of the image.This paper proposes a atlas sub-image preselection algorithm based on neural network.This algorithm uses a deep neural network model to pre-select the registered atlas sub-images in terms of intensity similarity and shape similarity.The segmentation experiments of subcutaneous tissues show that the algorithm can further improve the segmentation accuracy of target tissues.3.Multi-atlas label fusion algorithm combined with pixel grayscale probability informationIn view of the problem that currently proposed label fusion method does not use pixel grayscale probability information of the atlas,this paper proposes a multi-atlas label fusion algorithm that combines pixel gray probability information.This algorithm first uses a B-spline method to perform atlas registration,then uses a weighted voting label fusion method and a sparse representation label fusion method to fuse registered atlas labels image,and finally introduces pixel gray probability information to fuse the fusion result that obtained by the above two methods.The pixel gray probability information is obtained through segmentation training of atlas.This method combines the advantages of weighted voting label fusion method and sparse representation label fusion method,and makes full use of atlas prior information.This algorithm achieves better segmentation performance than other methods in brain MR image segmentation applications.4.Experiment and analysis of subcutaneous tissue segmentation based on multi-map automatic segmentation algorithmThe algorithm proposed above are combined into a multi-atlas brain MR image automatic segmentation algorithm.To solve the problem that the tissue outline in segmentation result is not smooth,which is not conducive to tissue subsequent 3D reconstruction,this paper proposes a segmentation result optimization method.This method takes advantage of the active contour model,and can obtain smooth target tissue contours.Firstly,take label fusion result as initial template,secondly,use template optimization algorithm to automatically extract initial active contour curve,finally,use active contour model to optimize the target tissue contour.The combined multi-atlas automatic segmentation algorithm is applied to segmentation task of subcutaneous tissue in 48 real human brains.The experimental results show that the segmentation algorithm proposed in this paper has better segmentation performance than stat-of-the-art automatic segmentation methods,and the segmentation results are closer to expert manual segmentation results.