| With the development of medical imaging processing technology,estimation of cardiac function parameters based on cardiac MR(Magnetic Resonance Imaging,MR)image of left ventricular segmentation is of great significance for the diagnosis and treatment of related diseases.The traditional left ventricular segmentation method has the disadvantages of low segmentation accuracy or can not achieve full automatic segmentation.With the development of deep learning,especially the application of fully convolutional neural networks in medical image segmentation,the accuracy of left ventricular segmentation has also been greatly improved,and fully automatic segmentation has been achieved.The method of left ventricle segmentation of the cardiac MR image based on the fully convolutional neural network,and a fast post-processing method for the defects of the segmentation result of the fully convolutional neural network are researched in this paper.The main work is as follows.(1)The performance of U-net network on left ventricular segmentation of cardiac MR images is studied and analyzed.The structure and characteristics of U-net network are introduced.At the same time,batch normalization algorithm is added to U-net network and the loss function based on Dice coefficient is used to train U-net network during network training.Subsequently,the Sunnybrook data set which is a public left ventricular segmentation data set is introduced,and the data set is randomly divided into a training set,a validation set,and a test set as the training and test data in this paper.In the experimental results,the U-net network achieved a Dice coefficient of 0.892,a recall rate of 0.893 and an accuracy rate of 0.921 in the test set.(2)In view of the shortcomings of the U-net network in the segmentation of MR images of the left ventricle of the cardiac,an improved U-net network is researched.These three aspects are mainly improved: a multi-input multi-output network structure is used to solve the problem of loss of image space details due to maximum pooling,and an attention module is added to the jump connection to suppress the segmentation of non-left ventricular regions.the loss function based on Tversky coefficients is used to train the network to solve the problem that the left ventricle occupies a relatively small portion in the cardiac MR image.The effects of different improved methods on the segmentation result are analyzed through experiments.From the perspective of the segmentation effect of the test set,the improved U-net network has a Dice coefficient increased by 3.3% and a Recall value increased by 4.8% compared to the U-net network trained based on the Dice loss function.(3)Since the the U-net network outputs the probability that each pixel belongs to the left ventricle,the probability value needs to be binarized.The common method of binarization is to directly use 0.5 as the threshold of left ventricle and non-left ventricle to binarizes the output probability map.Such a processing method will cause the edges of some segmentation results to be unsmooth or of noise points appearing.To solve this problem,the graph cut algorithm combined image morphology operation is proposed as the post-processing method of the U-net network.Image morphology operations are used to remove most of the U-net network segmentation results that do not require graph cut optimization,thereby greatly reducing the running time of the graph cut algorithm.The graph cut will combine the output probability map of the U-net network with the grayscale features in the neighborhood of pixels to construct an energy function and establish the graph model,the minimum value of the energy function is solved by the maximum flow / minimum cut algorithm,and the binary segmentation result is obtained.Compared with the method of binarizing the probability graph directly according to the threshold of 0.5,the edge of the segmentation result after graph cut optimization is closer to the real situation and has high computing efficiency. |
PDF Full Text Request