| With the development of modern medical imaging technology,magnetic resonance imaging(MRI)has become an important means of assisting physicians in the diagnosis of brain diseases.A good computer aided diagnosis system can greatly reduce the workload of doctors,make the diagnosis more timely,and the diagnosis result is more accurate.Compared with the adult brain MR image processing,infant brain image preprocessing,segmentation and other operations have more difficult.The main reasons are:the infant brain image leads to a smaller capacity,low imaging resolution,gray matter reversal problem during brain development and large noise,volume effect make the segmentation algorithm using on the adult brain can’t be used for infant brain images.At present,there is still no perfect algorithm for infants to make a high accuracy of the brain image segmentation or focus region segmentation.Therefore,this thesis aimed at the characteristics of infant brain image,construct two kinds of segmentation method for infant brain image:1,the method based on SVM;2,the method based on asymmetric and multi-scale image block with sparse representation.The main work and research results of this thesis are:(1)Study and use the brain image segmentation method for infants based on SVM.According to the inherent characteristics of 2D infants brain images,construct the features based on the image gray,space and probability characteristics.During the three stages of the model training,and constantly improve the training data for the targeted,improve the image segmentation accuracy,and get the final results of the two-dimensional segmentation results.(2)Study on the label fusion strategy under the general probability model,and propose an improved algorithm based on non-local image block,to deal with the problem of a few training images and reduce dependence on the quantity and quality of the registration.The algorithm still can obtain satisfactory segmentation results in the case of a small number of atlas.(3)Based on the non-symmetric and multi-scale image block,the construction method of the similarity set and the sparse representation strategy are proposed.Using the initial segmentation results,the low confidence region is extracted according to the confidence measure,which is used as the target area of the fine segmentation.The thesis construct similar sets using the idea of non-symmetric and multi-scale image blocks.Using the same idea,the thesis use sparse representation on the object image block,and the final segmentation result is obtained.(4)Experiment on the segmentation strategy which is proposed in this thesis.Experiment data on the 7 groups of infant brain MR images.In this thesis,a method of leave-one-out cross validation is used to test the segmentation performance and robustness.Finally,analyze and evaluate the experimental results. |
