Research On The Segmentation Of Brains MRI Based On Deep Learning

With the development of medical imaging technology in recent years,computer-aided medical diagnostic methods have become an important research field in medical imaging,diagnostic radiation,and computer science in the medical field.With the increasing incidence of brain diseases,in order to more effectively use neuroimaging to evaluate the manifestation of diseases and the effectiveness of treatment,it is inevitable that highprecision repetitive measurement and evaluation of the lesion area are required.And the precise segmentation of MR images is a necessary step for measurement and evaluation.Brain tumors are very common diseases in brain diseases.According to the severity of the disease,it can be found that the tumor area will form a tumor core,or even an enhanced tumor core in the MR images of brain tumors,so the segmentation of these three related areas of tumor is necessary for the disease diagnosis in the medical diagnosis of brain tumors.In response to the above problems,the theory of fully convolutional neural network is adopted to focusing on the segmentation of the whole tumor,tumor core and enhanced tumor core.According to the distribution characteristics of the lesion area,the structure of full convolutional network is optimized to improve the accuracy of segmentation,so as to provide good computer assistance for the diagnosis of brain nerve diseases.First,by analyzing the segmentation method based on U-Net,a three-level cascade neural network was used to segment the whole tumor,tumor core and enhanced tumor core layer by layer,thereby reducing the number of parameters in network training for saving memory,speed up the training.And the bounding box is extracted from the previous level as the boundary constraint of the input of the latter level network,so that the latter level network uses only the tumor area of the brain MR image for training,focusing on the feature extraction of the tumor part.And an anisotropic network is used to convert the segmentation problem from 3D to 2D.The dilated convolution is added to reduce the loss of information and residual connections is used to to avoid gradient explosion problems caused by deepening the network to obtain better segmentation results.Secondly,the fusion method based on cascading U-Net with fully connected conditional random field and K-means clustering based on bounding box is deeply studied.By analyzing the rough edge problem that may be caused by the large receptive field,the fully connected conditional random field is discussed to use for detailed optimization of the segmentation results,and then the K-means clustering method is added to improve the segmentation results and further improve segmentation accuracy.The algorithm of this thesis was experimented on the brain tumor data set published by BraTS17,and a good segmentation effect of the whole tumor,tumor core,and enhanced tumor core with Dice scores of 0.902,0.815 and 0.779 was obtained.The experimental results show that the proposed method can solve the problem that MR image boundary blur is difficult to segment brain tumor accurately.This study not only has important significance for the diagnosis and treatment of neurological diseases,but also provides a new idea for the research of image segmentation.