Research On Retina Image Generation And Segmentation Algorithm Based On Deep Learning

High-precision images of the characteristics of fundus retinal vascular disease can greatly improve the efficiency of disease diagnosis.In fact,producing a high-precision retinal vascular image from a color fundus image requires a professional doctor to invest a lot of time and energy in manual marking.Manual labeling has a high error rate and strong subjectivity.If deep learning technology is used for automatic labeling,it can not only reflect the objectivity of labeling blood vessels,but also improve the work efficiency of doctors.However,the insufficient sample size of the marked fundus images in the retina data set makes the generalization ability of the deep learning model poor,which affects the effect of training and segmentation.To this end,this paper studies the image generation algorithm,expands the sample size,and studies the corresponding deep learning segmentation model to improve the segmentation performance.(1)For the problems of the insufficient sample size of labeled fundus images in retinal data collection,blurred generated retinal images and unstable network training process.This thesis proposes an image generation algorithm based on Deep Convolution Generative Adversarial Networks(DCGAN)Multi-scale Generative Adversarial Networks(Multi-s GAN).The algorithm improves the discriminator and loss function in the DCGAN network model.It incorporates the idea of style transformation and adds a feature extraction network to generate a large quantity of clear and morphologically diverse retinal images.(2)For the complex and small vessels in retinal images that lead to low segmentation accuracy and incorrect segmentation,this thesis proposes an improved image segmentation algorithm IR2U-Net based on the R2U-NET model of U-Net.(Improved R2U-NET).The algorithm first replaces simple short-circuit connections of the R2U-Net network model with residual paths,uses a composite objective function,to complete training and testing on a publicly available fundus retinal dataset.The experimental results show that the algorithm has improved precision,accuracy,recall and IoU,respectively.(3)Conducting integration experiments to merge the data collection generated by the Multi-sGAN algorithm with the original data collection,as the enhanced training data collection into the IR2U-Net model for training.The experimental results show that the trained model has better generalization ability.