Research On Learning-based Retinal Image Segmentation And Diabetes Classification

In recent years,with the development of computer vision technology in the field of medicine,medical aided diagnosis has become a hot research direction.More and more image segmentation algorithms and image classification algorithms have been applied in the field of medical research,especially in the direction of retinal image,including retinal image vascular segmentation and the diagnosis of diabetic retinopathy.Retinal image has an important reference value for the analysis of cardiovascular diseases such as diabetes and hypertension,and the processing of image details is very important for doctors to analyze image diagnosis and the communication between doctors and patients.The main research contents and innovative achievements of this paper include:(1)Due to retinal vascular characteristics of image is not obvious,blood vessels are prone to missing,in the process of reconstruction in reaction to the phenomenon,this paper puts forward the improved SRCNN for retinal image super-resolution reconstruction,first using convolution kernel small lightweight module instead of feature mapping part of the ordinary convolution,then series lightweight module to improve the ability of network to extract the features,However,it may lead to the problem of gradient disappearance or explosion of the network.Therefore,the lightweight module is fused with Re Zero residuals to solve this problem and at the same time,the reconstruction results of blood vessel details are clearer.(2)Due to the complex structure of the blood vessels in the retinal image,the phenomenon of vascular discontinuity will occur during the segmentation.This paper proposed to improve U-Net to segment the blood vessels in the retinal image.Firstly,the dilated convolution with shortcut was added into the U-Net network,and the dilated rate was 1,2 and 4 for feature extraction.The network can reuse the image features while enlarging the receptive field,so as to better recognize the vessels of different widths,especially the vessels with less obvious terminal features.At the same time,the image attention module is also used to replace the jump connection in U-Net,which can extract the features of retinal images at different scales into the expansion path for use,so as to better distinguish the blood vessels and background of retinal images.(3)Due to the complexity of the area of diabetic retinopathy and the variety of characteristic information such as the shape and size of the lesions,it is more difficult to classify the lesions of retinal images.In this paper,Ghost Net lightweight network was used to classify diabetic retinopathy,and residuals were used to connect the ghost module at each scale.Enable the network to ensure the validity of features at every scale.At the same time,this model not only has a small number of parameters,but also has a satisfactory classification accuracy.