Study On Cardiac MR Image Segmentation Model Based On GANs

Medical image segmentation can extract specific information of organ or tissue regions,which plays an important role in disease evaluation.Multi-modal images(such as multi-model cardiac MR images)can provide different kinds of useful information,and combining multi-modal medical image information for image segmentation can help doctors make more accurate diagnosis.However,the current medical image segmentation field has the problem of limited available sample data.Generative adversarial network(GAN)has the ability of data generation,so it is of great application value to use small sample data to establish a GAN-based cardiac MR images segmentation model.In order to solve the problem of less available samples in cardiac MR image segmentation,this paper uses generative adversarial network to focus on two aspects of research work:(1)Proposed a two-stage GANs method to solve the problem of segmentation of LGE modal images with a small number of annotations in multi-modal cardiac MR images.Firstly,an image translation between LGE images and bSSFP images is realized based on Cycle-GAN,and then based on pix2pix,a multi-cascade pix2pix network is proposed for indirect segmentation of LGE images.In order to constrain the segmentation results from the feature layer,this paper adds the perceptual loss to the loss of the multi-cascade pix2pix network.Simulation experiments show that the segmentation results can be improved by adding perceptual loss.Meanwhile,the proposed method has better segmentation performance for left ventricle and myocardium compared with FCNs and other methods.(2)An image segmentation network with a random data expansion module is designed to segment bSSFP images.The data expansion module is a generative adversarial network(MSG-GAN),which can simultaneously generate bSSFP images and their corresponding segmentation labels,and use them with the real data sets for image segmentation training.At the same time,the data expansion module is optimized.Experiments show that the method proposed in this paper can improve the image segmentation effect and is better than the image segmentation method using traditional image expansion under certain conditions.