Generative Adversarial Network Based Medical Image Segmentation

With the continuous advancement of modern medical technology,the amount of medical image data continues to grow rapidly,which has caused computer-aided medical image analysis algorithms to attract widespread attention from both industry and academic field in recently years.Medical image segmentation have huge application requirements in location of lesion tissue,anatomical structure learning and other situations.An important research task of medical image segmentation is how to train high-performance segmentation models with limited fine-labeled data.However,in actual application,the cost of doctors' calibration data is very expensive and time-consuming,and a large amount of unlabeled data cannot be effectively introduced into the model training process.In order to cope with this situation,in recent years,the academic community has carried out tasks on medical image segmentation in fully supervised and semi-supervised modes.The algorithm of full supervised mode focuses on using limited refined labeling data to improve the segmentation performance as much as possible,and the semi-supervised algorithm introduces unlabeled data into the model training based on the fullysupervised method to further improve network performance,which can alleviate the current mainstream segmentation model's need to rely on a large amount of fine-labeled data training.Inspired by the idea of adversarial,this paper combines the algorithms of generative adversarial networks to achieve two tasks respectively.Most existing segmentation algorithms based on generative adversarial networks use discriminators in the form of classifiers to optimize the segmentation network.Such discriminators can only provide global adversarial information,which is insufficient for segmentation tasks that require a large amount of local detailed information.Therefore,this paper proposed a fullysupervised adversarial learning algorithm based on a fully convolution discriminator.On this basis,we introduced an attention mechanism to design and improve the segmentation network.Specifically,for small 3D MRI,CT datasets and 2D medical image datasets,we designed and improved the 2D segmentation network U-Net.For large 3D MRI and CT datasets,an attention mechanism was introduced to improve V-Net.Models that increase attention gate training can automatically highlight salient features that are helpful for a particular task,and thus eliminate the need to use a positioning module.In the experiment,the method has obtained good results.After the fully convolutional discriminators is introduced,the model performance is further improved.Further,on the basis of completing the fully supervised algorithm,this paper proposed a semi-supervised adversarial learning algorithm guided by importance weights named Importance Guided Semi-supervised Adversarial Learning for Medical Image Segmentation(ISDNet).The ISDNet method continues to use adversarial learning to optimize the segmentation network on the basis of the full convolution discriminator.It is particularly capable of implementing semi-supervised algorithms in combination with traditional self-training algorithms.Based on this,the importance of combining unlabeled data Sex weights further improve the performance of segmented networks.The BSN method can effectively utilize unlabeled data to further optimize the segmentation network.