Research On Left Ventricular Myocardial Segmentation And Post-processing Technology Of Cardiac MR Image Based On Generative Adversarial Network

With the development of medical imaging technology,the estimation of cardiac function parameters based on left ventricular segmentation of cardiac MR(Magnetic Resonance,MR)images is of great significance for the diagnosis and treatment of related diseases.Traditional left ventricular segmentation methods have disadvantages such as low segmentation accuracy or inability to achieve fully automatic segmentation.With people’s in-depth research on deep learning,especially the application of convolutional neural networks in medical image segmentation,the segmentation accuracy rate of the left ventricle in cardiac MR images has also been greatly improved,and fully automatic segmentation has been achieved.In this paper,the left ventricle segmentation methods of cardiac MR images based on the generative adversarial network are researched,and a fast post-processing method for the defects of the segmentation results of the generative adversarial network is proposed.The main work and innovations of the thesis are as follows:(1)The performance of the generative adversarial network on the segmentation of the left ventricle of cardiac MR images is researched and analyzed.The structure and characteristics of the Generative Adversarial Network is introduced.An attention module is attempted to add into the Generative Adversarial Network to suppress the influence of the non-left ventricular region on the segmentation results,A multi-scale discriminant network is applied to train the Generative Adversarial Network by using hierarchical features to capture the long-distance and short-distance spatial relationships between pixels.Then introduced the public left ventricular segmentation data set is introduced,which was randomly divided into training set,validation set and test set as the training and test data of this article.In the experimental results,the generative adversarial network achieved Dice coefficients of 0.9312,0.9623,Jaccard coefficients of 0.8866,0.9276,and Sensitivity coefficients of 0.9260,0.9557 in the inner and outer membranes of the test set,respectively.(2)Aiming at the problem of insufficient left ventricular segmentation data set of cardiac MR images,the idea of migration learning is presented and the strategy of layer-by-layer fine-tuning is proposed in the course of training.By using the VGG16 network to pre-train on natural images,the pre-training network is migrated to the target network toperform secondary training,Meanwhile,the pre-trained network is fine-tuned layer by layer according to the size and similarity between the source domain and target domain to make the network adaptively adjust network parameters,avoiding direct fine-tuning of each network to cause excessive calculations.The effects of different improved methods on the segmentation effect are analyzed through experiments.From the segmentation effect of the test set,the generative adversarial network has achieved Dice coefficients of 0.9399,0.9697,Jaccard coefficients of 0.8968,0.9415,and Sensitivity coefficients of0.9363,0.9686 in the inner and outer membranes of the test set,respectively.(3)Aiming at the problem that the edge of the ventricular membrane is fuzzy and the output result of the edge is poorly segmented by the generating adversarial network,an improved fast post-processing algorithm based on two-level set is proposed.When the traditional level set method is used to segment the left ventricular and outer membrane,there are problems such as unstable evolution of the level set function,low accuracy of fuzzy edge segmentation and low segmentation efficiency.A new distance regularization function and anisotropic gradient vector flow are proposed to improve the energy function of the level set.The new distance regularization function can better constrain the edge regularity,and can make up for defects in the output of the generated adversarial network.The anisotropic gradient vector flow makes the model have better segmentation accuracy at the edge depression.Finally,the 0-level set function and the k-level set function are used to evolve to inner and outer membranes of the left ventricle respectively.Through ROI extraction,most of the regions that do not need to participate in level set optimization in the output results of the generated confrontation network are removed,which greatly improves the time efficiency.Experiments show that the Dice coefficient,Jaccard coefficient and Sensitivity coefficient are all improved after the output results of the confrontation network are generated and processed.