Research Of Brain Structure Segmentation Algorithms Based On Sparse Representation And Deep Learning

Brain structure is importent to the body,brain structures include in the hippocampus,thalamus,pallidus and so on.Diseases such as alzheimer's,Parkinson's disease are associated with the abnormal brain structure,so the study of the brain structures' segmentation in MR image,not only can diagnose the illness as soon as possible,timely treatment,but also can help doctors diagnose,and decrease the workload.Hence it is of great significance.However,due to the inhomogeneity of gray scale,the blurring edge of brain structures and partial volume effect,the automatic segmentation of brain MR images is still a challenge.In this paper,the segmentation algorithm based on sparse representation and deep learning is deeply studied to perform 3d automatic segmentation of different brain structures.The main work and researching achievements of this paper are as follows:(1)This paper studies and implements the segmentation algorithm based on sparse representation combining with label priors.According to the distribution probability of the target structure in the label image,the threshold value is set adaptively to determine whether the pixel marked to belong to the target structure.In view of the complexity of the brain structures' surrounding tissues,a label-specific atlas patch partition method based on sparse representation algorithm is proposed,and the atlas patch is divided into a group of complementary image patches according to the label,which can enhance the boundary discrimination of the image patches.In addition,a post-processing method is proposed to correct the unstable region after initial segmentation with multi-scale information.It is proved that the algorithm segmentation accuracy can be improved by combining the label prior information by qualitative and quantitative analysis(2)Aiming at the problems of complex shape of brain structure,uneven gray scale and low contrast of brain images,a sparse representation segmentation algorithm combining texture structure features and deep learning features is proposed.In the sparse representation frame,multi-level features are added to reduce the misdirection of image patches' edges.The orientation-scale descriptor based on texture structure feature is integrated to solve the gray inconsistency between atlases,which improves the ability to distinguishing the edges of brain structures in images.By extracting deep-level features through convolutional neural network,the segmentation effect of complex brain structures is improved.Experimental results show that integrating multiple features can solve the problem of monotonous and inaccurate gray image patch information and improve the segmentation effect.(3)To solve the problem of more sample demand and poor effect of 3d network,information extracted in 2d network is insufficient information,this paper studies and proposes the brain structure segmentation algorithm based on multi-view collaboration networks.The three-dimensional brain structure images are sliced at multiple views and inputted into the twodimensional network,the output probability maps and the original image are concatenated as the input of 3d network.Then 2d dimensional networks' output and 3d networks' output are combined to get the final result.The proposed algorithm can use global information of the image and extract multiple views information,which makes the image detail is preserved and also can overcome the disadvantages of small sample.At the same time,the network cooperates training of two-dimensional and three-dimensional images,which can extract the information inside and between the slices of brain structure images.The accuracy of the algorithm is improved.