Research On Improved CGAN And Its Application In Retina OCT Image Analysis

Conditional Generative Adversarial Nets(cGAN)aims to solve the problem of uncontrollable training of Generative Adversarial Nets(GAN).It is a derivative model which adds supervisory information to GAN and composed of a generator and a discriminator.The adversarial and unified training mode of generator and discriminator makes this model have the following advantages:(1)less demand for the quantity of training data.(2)high quality of generated images.(3)the model of generator and discriminator can be combined with other neural network structures.In this paper,two cGAN based models were proposed,aiming at the two tasks of retinal OCT image generation and retinal OCT image segmentation,and effective improvements were made to them.The imbalance of training data seriously affects the performance of medical image classification algorithm.To expand the disease data set,we proposed an end-to-end framework for OCT image generation based on cGAN.The new structural similarity index(SSIM)loss is introduced so that three kinds of retinal disease images are able to be generated from images with no disease.The generated images assume the natural structure of the retina and thus are visually appealing.The method is further validated by testing the classification performance trained by the generated images,and the experimental results further verify the effectiveness of this method.Image segmentation has always been the key task of medical image processing.In this work,we made some improvements to cGAN,and the proposed method can be used to the automatic segmentation of pigment epithelium detachment(PED)and Subretinal fluid(SRF).The improvements are described below.First,the residual block is added to the generator network structure.Secondly,the segmentation problem is converted into pixel level classification problem,we combine the LI loss and Focal loss,solving the problem that retina background area has more pixels than lesion area.We compared the results of this method with those of other methods.In the experiment,Dice coefficient,true positive rate,false positive rate were used to evaluate the results.Experimental results show that the cGAN model with improved generator structure and loss function can accurately segment PED and SRF regions on the OCT dataset of Al Challenger,and the Dice coefficient of the two diseases increases by 2.54%and 4.87%respectively compared with the method before improvement,and the final segmentation results are close to the ground truth.