Research On Segmentation Algorithm Of Infant Brain MR Images Based On Neural Network

Modern medical imaging technology has become a clinical diagnosis tool of brain disease.In the existing medical imaging technology,magnetic resonance(MR)imaging has become the first choice to find the brain structural lesions of infant,according to its characteristics,such as no ionizing radiation damage to the human body,high spatial resolution and getting high-resolution imaging of the brain and nervous system.The segmentation of infant brain tissue into WM,GM,and CSF plays an important role in studying early brain development in health and disease.It's also important for the doctor's analysis of the disease and following treatment.Therefore,in this thesis,we studied 2D and 3D infant brain MR image segmentation algorithms.In summary,the major work and innovation of this thesis include:(1)Learned SOM and then found the defects of this algorithm,such as over-segmentation problem.Improving the direction of search neighborhood and the evaluation index of search direction can solve the over-segmentation problem.Combined SOM with GA is used to reduce the feature dimensions effectively which implements the adaptive input of SOM neural network.PSO algorithm is used to optimize the entropy/standard deviation threshold,which solves the shortcoming of empirically determined threshold.The algorithm proposed in this thesis significantly improves the accuracy and stability of the imfant brain MR image segmentation,which is improved search direction based on mixed neighborhood and minimum standard deviation to improve the gradient entropy to evaluate the image pixel gray value discrete degree.(2)Learned segmentation algorithm of infant brain MR image based on convolutional neural network.The thesis improves the defect of single grayscale information in the input of original CNN,and improved network input based on GLCM which takes local features into account,such as symbiosis and mean,energy and so on.According to the anatomical characteristics of infant brain MR images,we proposed three three different network structures for infant MR image segmentation,and the design reason and network parameters were analyzed in detail.The experimental results show that the three CNN structures proposed in this thesis can complete the segmentation task of 2D infant brain MR images,and improve the segmentation accuracy,especially for the CSF segmentation.(3)To study and finish the task about the segmentation of 3D infant brain MR images.We proposed an algotithm named "three-channel-two-stage" network structure,which considered the anatomic structure characteristics of the 3D infant brain MR images and 3D convolution neural network.The proposed algorithm in this thesis which is based on "three-channel-two-stages" network structure and improved classification estimation of SOM can complete the take of brain MR segmentation such as brain tissue segmentation and brain structure segmentation.The experimental results show that the proposed algorithm can improve the accuracy,enhance the robustness,and enhance the flexibility of the algorithm.