Research On Cardiac Magnetic Resonance Image Segmentation Based On Deep Learning

Cardiovascular disease has been the leading cause of death among all kinds of diseases.It is of great significance to use computer-aided diagnosis system to determine whether the heart is diseased.Cardiac magnetic resonance imaging is regarded as the gold standard for quantitative cardiac analysis.Accurate segmentation of cardiac magnetic resonance images has become a major problem due to the complex structure of heart tissue,the large variation of ventricular shape and the existence of blurred image edges.Therefore,a multi-scale context residual attention network model,MCRAU-Net,is proposed in this paper,and the network model is compressed to achieve accurate segmentation of cardiac magnetic resonance images.To solve the problem of cardiac magnetic resonance image segmentation,an improved MCRAU-Net network model based on U-Net was proposed.The multi-scale context attention module and decoder-guided attention module were introduced into U-Net.The multi-scale context attention module captures and filters the context information,and the decoder-guided attention module solves the problem of too large semantic gap in the process of jump connection.At the same time,a residual structure is introduced in the decoding stage to alleviate the problems of gradient disappearance and gradient explosion.The experimental results show that the network model proposed in this paper achieves good segmentation effect.Aiming at the problem that the heart region occupies a relatively small part of the magnetic resonance image,a hybrid loss function is proposed in this paper.This loss function combines Dice and weighted cross-entropy(CE)loss functions.Different coefficients of CE and Dice loss functions are set to achieve faster convergence of the model,and different coefficients of weighted CE are set to alleviate the problem of class imbalance in the image.Experimental results show that the proposed hybrid loss function improves the scores of both Hausdorff and Dice metrics compared with Dice and CE loss functions.In view of the problem that adding multiple different modules in the improved network model will introduce a large number of hyperparameters and computation,the proposed network model was designed for lightweight.In this paper,the lightweight convolution operation and the multi-scale efficient channel attention module were designed and implemented to replace the traditional convolution operation and the channel attention module to achieve the compression of the network model.At the same time,in order to further reduce the computational load of the network model,the region of interest(ROI)was extracted from the dataset.The experimental results show that the network model is compressed and the segmentation performance is also improved.