Study On Retinal Vessel Segmentation Based On GAN Using Few Samples

There are more than 400 million patients with chronic eye diseases in China,but the number of clinicians is far from enough,and the community lacks automatic screening equipment,resulting in more and more early patients who cannot be diagnosed through retinal screening.Nowadays,deep learning has made achievements in various fields with large samples and high accuracy.However,although there are many patient samples in the medical field,it is difficult and expensive to obtain manually label,which makes it difficult to obtain labeled samples.At the same time,because the sample data of healthy people is easy to obtain,and various severe samples are difficult to obtain and have different shapes,most of the public data sets of retinal blood vessel segmentation have the problem of class imbalance of healthy people and patients.Network generalization will have an inhibitory effect if imbalanced sample sets are used for training.Research on the enhancement of retinal image data has found that reducing the total number of training sets based on the cardinality of the scarce samples in the training set and reducing the number of samples in different categories to a ratio of one can achieve a simple and effective plan to balance sample and keep information richness.At the same time,the problem of class imbalance is transformed into the problem of few-sample learning,and then the problem of sample scarcity is solved.After studying U-Net,it is found that the layer-jumping connection mechanism proposed by it has natural advantages for medical image processing,but it is still not enough to compensate for the performance degradation caused by fewer samples.Subsequent research on generative adversarial networks found that its zero-sum game idea can improve network performance through confrontation without increasing the complexity of the network.After combining the generative adversarial network with U-Net,it proves that the idea is feasible and solve some problem of the defects of the generative adversarial network.Based on the above research foundation,this paper proposes a few samples retinal blood vessel segmentation method based on a generative adversarial network.The generator in the generative adversarial system uses an improved U-Net for retinal blood vessel segmentation,and the discriminator in the generative adversarial system uses an improved VGG16.This neural network is responsible for building a confrontation system with the generator,forcing the generator to optimize under the constraints of the discriminator.In the experiment,it was carried out on the DRIVE data set and the mixed HRF data set.The training set has only 6 samples after class balance,and the test set remains unchanged.The segmentation performance of this method is better than that of DU-Net and Iter Net on these two data sets.Compared with DU-Net and Iter Net of trained with the complete data set,the accuracy rate of segmentation performance is only less than 1%.The accuracy rate of segmentation performance is slightly improved compared to the U-Net trained with the complete data set.At the same time,the network model is small,the training and testing speed is fast,and it is easy to implement in the terminal equipment in clinical applications,which proves this method is effective for the few samples retinal blood vessel segmentation and proposed a new solution to the problem of sample scarcity and imbalance in samples.In addition,after using the unbalanced training set in the control experiment,it is found that whether it is too many samples of healthy people or too many samples of patients,it has a strong inhibitory effect on the network segmentation performance,which proves that the training set class balance improves the accuracy of retinal blood vessel segmentation.These results further affirmed the importance of solving the problem of class imbalance.