Application Of Generative Adversarial Networks Based On Domain Adaptation In Fundus Image Segmentation

Today,glaucoma is one of the main eye diseases that blind patients.The cup-to-disk ratio obtained from the fundus image is an important basis for diagnosing this eye disease.The cup-to-disk ratio is calculated based on the maximum vertical diameter of the optic disc and cup regions divided in the image,and the precise segmentation of these two regions usually requires an ophthalmologist with many years of experience to complete.Manual segmentation is inefficient in reality,so it is valuable to complete such segmentation tasks with the help of models.Previous works on fundus image segmentation models often pursue segmentation effects on a single dataset.However,for slightly different fundus images obtained by different fundus cameras,if the model is used without retraining,the segmentation effect usually declines significantly.In other words,there is a problem of domain shift in segmentation tasks.Since the fundus image dataset with segmentation labels is scarce,retraining the model on fundus images obtained from different devices does not work.Considering the above situation,that is,for the problem of domain shift,this paper studies the fundus image segmentation model based on domain adaptation.Specifically,the model in this paper is used to complete the joint segmentation of the optic disc and the optic cup in the target fundus image.A common segmentation model is to train the model on the training set,and then perform a model check on the test set from the same dataset.Different from this,when there is a difference between the source and target domains,the idea of the domain adaptation model is to use the labeled source domain dataset and the unlabeled target domain dataset to jointly train the model,and then test on the target domain dataset.The former has segmentation labels,and the training process belongs to supervised learning.The latter has no segmentation label and the training process is unsupervised learning.The previous domain adaptation model used the domain adaptation method at the input or output of the segmentation model.The innovation of the model in this paper lies in the use of a generative adversarial network structure with a domain adaptive function on the input and output ends of the segmentation model.From the perspective of function implementation,the entire model can be divided into two submodels.One is the image translation submodel,which is implemented by a generative adversarial network.It is used to extract the shared features of the fundus image data in the source and target domains.The other is the image segmentation submodel,which completes the segmentation task based on the shared features obtained by the translation submodel.Further,from a structural point of view,the image translation submodel includes three parts: an encoder,a decoder,and a discriminator.The shared features are generated by the encoder.The image segmentation submodel includes two parts: a segmentation network and a discriminator.The combination of the segmentation network and the discriminator also belongs to a generative adversarial network.The former produces segmentation maps,and the latter performs discrimination and plays the role of domain adaptation.The shared features obtained by the translation submodel and the discriminator in the segmentation submodel are two designs that are different from ordinary segmentation models,so that the entire model has the ability to reduce the performance degeneration.That is,it can better complete the segmentation task for the fundus images of the target domain,and alleviate the challenge of domain shift.This paper uses REFUGE,Drishti-GS1,and RIM-ONE＿r3 datasets for quantitative and qualitative analysis.Among them,the training set and test set of the REFUGE dataset are significantly different because they come from different devices.Therefore,we regard them as different datasets,which are respectively denoted as REFUGE＿1 and REFUGE＿2.Quantitative experiments include comparative experiments between our model and other different models and an ablation study that verifies the effectiveness of each part of the model.In the quantitative experiments,the REFUGE＿1 dataset is used as the source domain dataset,and the REFUGE＿2,Drishti-GS1,and RIM-ONE＿r3 datasets are used as the target domain dataset.Comparative experimental results show that the model in this paper can effectively alleviate the domain shift.Taking the F1 score index as an example,the model of this paper on the REFUGE＿2 test set has an F1 score of 0.9512 for the optic disc and an F1 score of 0.8465 for the optic cup.On the test set of the Drishti-GS1 and RIM-ONE＿r3 datasets,which are more different from the source domain dataset,the F1 score of the optic disc is 0.9403 and 0.9072,respectively,and the F1 score of the optic cup is 0.7749 and 0.7232,respectively.The above indicators have significantly improved the corresponding results compared with the segmentation model without using the domain adaptation method.The results of the ablation study verify the effectiveness of the domain adaptation method applied to the input and output of the segmentation network.The results of qualitative experiments intuitively corroborate the conclusions of the quantitative experiments.