Fundus Image Generation Based On Generative Adversarial Networks And Deep Learning For Retinal Vessel Segmentation

Since the introduction of the Generative Adversarial Networks(GAN)by Lan Goodfellow in 2014,GAN has become a very hot research area in artificial intelligence,and a large number of GAN-related articles have been published.At the same time,GAN’s technology has also been applied to various fields,such as image generation,speech generation,natural language sequence generation,and style migration.Because of GAN’s unsupervised learning style,GAN is very likely to achieve true intelligent learning,which will become one of the main research fields in the field of artificial intelligence in the future.As a hot artificial intelligence algorithm,deep learning has an extraordinary impact in various fields.Compared with traditional machine learning algorithms,deep learning can easily exceed the machine learning and other intelligent algorithms in various evaluation indicators,especially in computer vision and image processing,the development of deep learning has made these fields an unprecedented breakthrough.With the continuous development of deep learning and excellent performance in image processing,deep learning is becoming the mainstream intelligent auxiliary diagnosis and intelligent diagnosis algorithm in the field of medical image processing.Retinal disease detection is the only non-invasive method of detection in humans.However,obtaining a vascular binary image from a color fundus image requires a professional doctor to spend a lot of time and effort on labeling the color fundus image,and high-precision fundus vascular segmentation can greatly improve the diagnosis of disease,so the use of advanced intelligent algorithms is of great significance and value to perform fundus vascular segmentation and improve the accuracy of vessel segmentation.Since the number of fundus vascular images marked with the existing fundus image data is very small,the generalization model of the deep lexarning model trained by using these data is relatively poor.An intuitive idea is to apply the two popular techniques of generative confrontation network and deep learning to the field of blood vessel segmentation.The generative adversarial network is responsible for generating new fundus images to augment the dataset,and then the augmented dataset is used for training.The deep learning model is trained to obtain a deep learning model with better generalization performance.The main work of this paper includes two aspects.One is to propose a segmentation network(DRAU-Net)with DenseBlock,ResBlock and Attention mechanism added to U-Net.It is proved by experiments that the DRAU-Net is compared with Several other classic networks have great advantages in the fundus segmentation task,and the effect of blood vessel segmentation has reached the world’s leading level and exceeds the discriminating ability of professional doctors.The second is to use the above proposed network structure as a generator to design a new GAN network(DRAGAN),using DRAGAN to generate realistic color fundus images on the existing binary blood vessel segmentation map,using the generated data and real data simultaneously.And the generated data and real data are used to train a DRAU-Net model to enhance the generalization capabilities of the network.By evaluating the trained DRAU-Net model on the test set,the generalization ability of DRAU-Net is discriminated from the evaluation index,and it is also indirectly proved that the color fundus image generated by DRAGAN can be used for data amplification.A way to enhance the generalization capabilities of the model in the absence of data has been provided to enhance the generalization capabilities of the model in the absence of data.