| With the development of deep learning technology,medical image segmentation technology has made great progress,especially in the background of the rapid development of artificial intelligence.The research on medical image segmentation using multi-layer perception also has made remarkable achievements in recent years.Supervised learning in the existing intelligent medical image segmentation can obtain stable features from multi-layer perceptual learning and has been applied in clinical practice.However,the supervised strategies need too much annotated data,which requires experienced experts to do the annotation.In order to solve this problem,this paper explores a semi-supervised learning strategy in this field.Specifically,the left ventricular segmentation method based on semi-supervised learning is studied and explored.The specific work is as follows:Firstly,a semi-supervised regularized residual learning network(ss RRLNet)is designed.Under the supervision of a small amount of annotated data,the learned features are transferred to unlabeled data learning process through transfer learning after the preliminary features are obtained.Furthermore,some labeled data are treated as unlabeled data,and the learning results from the regularized residual network are obtained under the constraint of minimum residual loss.Through validation and testing in public dataset ACDC,it is proved that this method can obtain satisfactory left ventricular segmentation results under the supervision of a small amount of labeled data.Secondly,a semi-supervised left ventricular segmentation method with dual path consistency constraints is proposed.On the basis of ss RRLNet network,DPCCNet(Dual Path Consistent Constraints Network)is designed along with a Siamese network.The input data is disturbed by the adjustment of gamma exposure.The original image and enhanced image are used as input data,and the Structural Similarity Index(SSIM)is introduced as loss control to strengthen the learning capability.Further,the consistency constraints are adopted to penalize and correct the network and improve the generalization ability.The experiments show that,under the same conditions,Dice Coefficients(DSC),Jackard Coefficients(JC)and other index values are satisfactory.It’s proved that this model has satisfactory segmentation performance compared with other relevant segmentation methods.In conclusion,this paper proposes two semi-supervised left ventricular segmentation methods,and explores an effective solution for the clinical problem in practice.The method studied can obtain satisfactory segmentation results under the supervision of a small amount of labeled data.It has been proved that this method has certain practical application values,and clinical application prospects. |
