| Cardiac magnetic resonance imaging(MRI)diagnosis is an important diagnostic method for heart diseases.A series of cardiac performance indicators and disease types can be obtained by studying the cardiac magnetic resonance imaging.One of the important detection methods is to accurately segment the ventricles in the images.In clinical practice,experienced doctors are required to perform segmentation manually.But this method of manual segmentation is time-consuming.Nowadays,image segmentation algorithms based on deep learning have played a huge role in practice and achieved good segmentation results.However,compared with natural image segmentation,ventricular segmentation has some obvious difficulties.On the one hand,the number of medical image segmentation methods used for deep learning network training is small and difficult to obtain.The features of low resolution,unclear image,blurred edge of ventricular image,and ventricle with small contraction area bring some difficulties to the training of deep learning network.To solve the problem that there are few ventricular images available for training and the ventricular region is small,this paper proposes a semi-supervised segmentation algorithm for left and right ventricles based on adversarial learning.The semi-supervised adversarial learning algorithm is based on a generative adversarial network(GAN),which replaces the generative network with a segmented network suitable for segmentation of the left and right ventricular regions,and replaces the discrimination network with a fully convolutional network that can generate a confidence map.The generation network and the discrimination network can confront each other and learn from each other during the training process,and jointly improve the segmentation ability and the discrimination ability.Only a small number of images with labeled were used in the initial training in the network.When the network has acceptable segmentation and discrimination capabilities,images with labeled and unlabeled images are input alternately for semi-supervised learning.This can make full use of a large number of unlabeled images.In addition,for the complex characteristics of the ventricular images,a MSPN module and a multi-scale dilated convolution module are used in the segmentation network to perform a multi stage and multi scale of feature extraction.In orderto reduce the amount of model parameters and avoid overfitting,a reasonable number of filters are set in the segmentation network.In the experiment,this paper used multiple data sets to verify the effectiveness of the algorithm,and compared with other excellent left and right ventricular segmentation results.The experimental results show that the algorithm improves several evaluation indicators.At the same time,in order to further improve the accuracy of the left and right ventricle segmentation of the model,and reduce the required training data,this paper also proposes a semi-supervised segmentation algorithm of ventricles based on the cycle generation adversarial network,and add the attention gate module.One group of adversarial networks consists of adversarial networks and generative networks capable of producing segmentation results,while the other set of generative networks is used to generate labels.Compared with the aforementioned algorithm,we add the process of generating labels.The experiment shows that this method can effectively improve the accuracy and robustness of segmentation with fewer labeled images. |
