Research On Segmentation Of Brain Mri Based On Deep Subspace Clustering

Brain is an important organ of human body,which plays an important role in human health.In medical imaging technology,Magnetic Resonance Imaging(MRI)has the advantages of high resolution and low radiation,which is widely used in medical diagnosis and treatment.Brain image segmentation plays an important role in the diagnosis,location and treatment of brain diseases.Therefore,it has an important guiding significance for medical assistant diagnosis to achieve the automatic segmentation of brain MRI.The research goal of this thesis is to segment the cerebrospinal fluid,gray matter and white matter in brain MRI automatically and accurately,so as to provide service for medical diagnosis and treatment.Supervised learning in deep learning has achieved very good results in the field of image segmentation,but it limits its application scenarios to relying on a large number of labeled data.Therefore,this thesis proposes an unsupervised learning method,which combines the local homogeneity of supervoxel and the nonlinear transformation of deep subspace clustering to segment the cerebrospinal fluid,gray matter and white matter in three-dimensional brain MRI.Firstly,the supervoxel of 3D brain MRI are obtained by using the supervoxel generation algorithm,and multi-scale features are extracted to describe the supervoxel.Compared with the traditional voxel based method,the subsequent processing time can be greatly reduced by transforming a large number of voxels into a small number of supervoxels.Then,we use the deep subspace clustering model to cluster the supervoxel into different tissue categories.Finally,the clustering results of supervoxel are mapped to the original image to obtain the tissue segmentation results.Different from the previous subspace clustering,this method can learn the non-linear transformation and self-expression coefficient of features at the same time,which improves the accuracy of clustering results,and then obtains good brain tissue segmentation results.In order to further improve the segmentation results,this thesis further improves the deep subspace clustering model with the help of clustering results information of supervoxel,and proposes a supervoxel clustering model based on self-supervised deep subspace to improve the accuracy of brain tissue segmentation.In each iteration,the clustering results are used to supervise the learning of self-expression layer and feature extraction layer,so as to improve the accuracy of coefficient in self-expression layer.During supervising self-expression layer,the clustering results make the coefficients of the supervoxels which do not belong to the same class in the self-expression matrix tend to zero and eliminate the influence of the uncorrelated supervoxels.On the other hand,our thesis adds a classification layer in the behind of the encoder,and the clustering result is used as a "weak label" to supervise the network to extract nonlinear feature.This self-monitoring mechanism adds two loss functions to the network and is integrated into the total objective optimization function.Compared with the supervised learning method in deep learning,the method proposed in this thesis completes the segmentation of brain MRI without relying on the labeled image,and increases its application scene.In order to verify the effectiveness of this method,experiments are carried out on IBSR18 and Brainweb20 datasets.The results show that this method can effectively segment the three kinds of tissue in brain MRI.Compared with other methods,the segmentation result is significantly improved.