Research On Analysis Algorithms Of Retinal Image Based On Deep Learning

Retinal fundus image is the major basis for the diagnosis of diabetic retinopathy,macular disease and other eye diseases.With the popularization of portable fundus camera and the rapid development of artificial intelligence technology,there are many automatic screening systems for eye diseases.The success of these automatic diagnostic systems largely depends on the quality of the input image.Practical datasets can include a large number of poor quality retinal images caused by uneven illuminate,occlusion,patients movements and so on.Therefore,effective retinal image quality assessment algorithm is important for intelligent screening of eye diseases.Furthermore,the poor quality fundus image can be enhanced by image enhancement algorithm.For the patients with vitreous opacity,cataract and other eye diseases,the fundus image will be blurred and low contrast due to the damage of its refractive medium.These patients can not access high quality retinal images even if the parameters of fundus camera are adjusted repeatedly.In intelligent diagnosis system,image enhancement of poor quality fundus image is also one of the major problems.The contents and innovations of this paper are as follows:Firstly,a method of retinal image quality assessment based on deep learning is proposed.The proposed method introduces the human visual attention mechanism into CNN architecture.The main network adopts VGG-19 architecture.The transfer learning is used to initialize the network weight from Image Net.The attention net is based on foreground extraction by extracting the blood vessel and suspected regions of lesion and assigning higher weights to region of interest to enhance the learning of these important areas.The experimental results show that compared with the traditional retinal image quality assessment method,adding the attention net to the VGG-19 architecture which is in line with HVS and human visual attention mechanism can achieve better performance.Secondly,a method of retinal image enhancement based on generative adversarial networks is proposed.The method based on Cycle GAN achieves the transformation from poor quality fundus images to good quality fundus images.Cycle GAN architecture applied to image enhancement is first proposed.The network consists of two generators and two discriminators.In order to generate a higher resolution fundus image,the depth of the generator is increased to make it have a larger receptive field.The experimental results show that the proposed method can enhance the retinal image in poor quality caused by underexposure,uneven illuminance and blurring.At the same time,the method avoids the problem of color distortion in traditional image enhancement algorithms.Inaddition,by comparing the classification accuracy of eye disease screening system after image enhancement.The experimental results show that the proposed method can improve the accuracy of the screening system.