Research On Image Recognition Technology Of Diabetic Retinopathy Based On Deep Learning

With the continuous improvement of people’s living standards,Diabetic retinopathy is threatening people’s visual function and becoming a disease in daily life.At present,the diagnosis of diabetic retinopathy in hospitals mainly depends on ophthalmologists with sufficient clinical experience.Obviously,different doctors are subjective about the diagnostic results,it is also time-consuming and laborious for patients to repeatedly examine.And there are some issues of poor fundus image quality and inadequate feature extraction in the existing research work.In order to improve the diagnostic accuracy and efficiency,the research content of this paper is as follows to address the problems in diabetic retinopathy image recognition technology:(1)In view of the poor image quality and the fact that the existing methods will overenhance or optimize the fundus image,resulting in the loss of important information and the inability to ensure the clarity of the original image,a UNet-type network is proposed,which combines residual dense block and Att Op Blk attention module.Residual dense block is used to extract local dense features,and Att Op Blk attention is used to enhance local detail features and improve the contrast,color and details of the image.This method successfully solves the problems that traditional methods can’t maintain brightness and contrast of the original image,and enhances the effectiveness and clarity of the image.(2)To address the problem of DR classification with inadequate feature extraction and poor accuracy in the existing methods,an AIDnet network is proposed,which combines the generating countermeasure network and the dual attention mechanism.This method uses the improved ACGAN to generate images,which makes the network pay equal attention to each kind of features and improves the feature extraction effect of the network.In this paper,by fusing features of different scales on the channel dimension in parallel through the Inceptionv3 network,while using a dual-attention mechanism to achieve the fusion of channel and spatial features,the classification accuracy of DR is improved and meets the needs of clinical medicine to some extent.(3)In order to solve the problems of low segmentation precision and large overlapping error between prediction and real results in the existing methods,a network model based on Deep Labv3+ combined with depth supervision and conditional random field is proposed.This method can better learn the semantic features of optic disc and macular region through depth supervision,thus improving the accuracy of segmentation;At the same time,the dependence between pixels is considered by conditional random field,and the segmentation results are adjusted more finely,so that the model can be better represented in complex image scenes.The experimental results indicate that the method proposed can effectively improve the automatic diagnosis and treatment of diabetic retinopathy in this paper,which has some application value and clinical significance.