Research On Cardiac MR Image Segmentation Technology Based On Improved U-NET

In recent years,the frequency of cardiovascular disease has increased year on year and cardiovascular disease is now the leading cause of death from noncommunicable diseases each year.Cardiac MR（Magnetic Resonance,MR）images play a vital role in the calculation of cardiac-related functional parameters and the diagnosis of related diseases.In clinical practice,the segmentation of the various parts of the heart image is mainly performed manually by experts,and it is more difficult to segment the various parts of the heart,as the right ventricle varies greatly in shape and changes significantly from one stage to another,which is not only timeconsuming,but also produces subjective differences when segmented by experts.（1）Due to the small number of samples in the existing public dataset of cardiac MR images,which is not enough samples as a training set,a data enhancement operation is performed on the dataset,and the data is normalized to grey-scale values.In-depth study of U-Net and implementation of U-Net network for segmentation of right ventricular MR images was carried out.Combined with the Dice coefficient as an evaluation index,the deficiencies of the network are analyzed.（2）The ordinary convolution in the down sampling stage is replaced by extended convolution to solve the problem of information loss caused by cardiac MR images during pooling.The network loss function was improved and the attention module was introduced in the up-sampling stage to solve the situation that the right ventricle is easily confused with other tissues during segmentation,resulting in missegmentation,and to increase the network segmentation accuracy.To address the situation that the U-Net network has slow training speed due to gradient explosion caused by too deep network layers in the information transfer process,three residual modules are added to the encoding and decoding stages of the original network respectively to improve the network training speed,resulting in the improved U-Net network（i.e.,RAU-Net）.Experiments showed that the loss value of the improved network decreased from 0.4 to 0.1 and the Dice coefficient increased from 0.84 to0.91.（3）In order to extract more feature information to further improve the segmentation accuracy,two RAU-Net networks with the same structure were connected（i.e.,LRAU-Net）to form a multi-pair coding and decoding structure.In addition,as the layers deepen further after the networks are connected,the residual module is further improved in order to prevent the training speed from decreasing.Experiments show that the improved network has a loss value of 0.09 and a Dice coefficient of 0.95,confirming the effectiveness of the improvements.（4）Combined with the improved network,the automatic right ventricular image segmentation system was designed according to the actual clinical needs.It contains permission management module,data display module,image segmentation module,information remark module,manual annotation module and image storage module.After the system retrieves the patient’s cardiac MR image,it can achieve automatic segmentation and calculate the relevant cardiac index,which provides reference for doctors’ diagnosis.