The Research On The Key Technology Of Intelligent Graded Diagnosis For Ocular Fundus Lesions

The visual system is an important part of the nervous system.It enables humans to have visual functions and enables us to communicate with nature.Recently,with people’s living standards improved and the aging of the global population increased,the incidence of diabetic retinopathy has also increased a lot.Diabetic retinopathy is a complication of diabetes.The clinical manifestations are blurred vision and decreased vision.If there is no effective diagnosis and treatment for a long time,you may eventually face the risk of blindness.The global economic loss caused by decreased vision or blindness is about 3 trillion US dollars every year.Corresponding to the rising incidence is the shortage of medical resources.Whether in developed or developing countries,the growth rate of medical resources is far below the annual growth rate of visually impaired patients.In this case,it is particularly important to use artificial intelligence to assist doctors in disease diagnosis.On the basis of careful study of the medical characteristics of macular edema and many detection methods,we propose an enhanced multi-feature fusion network(EMFN)suitable for the graded diagnosis of macular edema.In this model,in order to highlight the details of the fundus image,we have done a variety of processing for the input image,and selected the green channel of the original image,morphological features,CLAHE,and image curvature as input,and then these four features are fused in the high-level features generated in the network.At the same time,in order to further improve the detection accuracy of the model,we also made further adjustments to the network structure of the model,introducing an attention mechanism,allowing the model to select important features for learning.During the process of training EMFN,we found that the size of the training data set seriously affects the detection accuracy of the model,but due to the particularity of medical data,the cost of acquiring a large amount of medical data is very high.Compared with other fields,the amount of open source fundus image datasets is very small.In response to this situation,we propose a fundus image generation model based on a generative adversarial network.The model is mainly composed of three parts,generator,discriminator and VGG network.The VGG network is mainly used to extract the high-level semantic information of the original image and the generated image,and then use it as a content loss function to guide the model to better retain the semantic information of the original image.Besides,for the purpose of improving the detail of the edges of the generated image,we also introduce a smoothed input image,and guide the model to generate the fundus images with clear edges by letting the discriminator distinguish the smoothed fundus image from the original fundus image.