Research On Classification Of Retinal Diseases Based On SE-Block

Retinal diseases,especially diabetic retinopathy(DR)and age-related macular degeneration(AMD),are still the leading causes of low vision or even blindness.If the signs and symptoms of retinal diseases can be identified at an early stage and targeted drug treatment can be used,it can prevent loss of vision and even cure.However,hundreds of millions of people worldwide suffer from retinal diseases each year,and in the face of extensive and complex diagnostic work,a large number of professional and experienced ophthalmologists are needed.But some underdeveloped areas do not have an excellent medical environment.In recent years,there have been endless ways to use deep learning to help medical staff analyze and process medical data to treat diseases.Among them,convolutional neural network(CNN)is the most widely used in medical image analysis.Compared with the traditional algorithms used for image classification,Convolutional Neural Networks(CNN)can not only automatically extract features in images,but also improve the effect and speed.These advantages make it a new research focus in the field of image classification.Therefore,the application of convolutional neural networks in deep learning to the auxiliary diagnosis of retinal diseases has certain research significance.In this paper,both retinal diseases and convolutional neural network(CNN)are studied in depth.Based on the open source retinal OCT image dataset,a convolutional network model for classification of retinal diseases was proposed on four classification tasks(CNV,DRUSEN,DME,and NORMAL).The goal is to achieve the level of judgment of a professional ophthalmologist to assist the doctor in diagnosing retinal diseases.The paper works as follows: Firstly,a traditional convolutional network model is trained based on the open source retina OCT image dataset,and its training parameters are saved.Then,the SE and Bottleneck layers arecombined into a new module SE-Block,and a new convolutional neural network is designed based on SE-Block.Among them,SE-Layer based on the Attention idea to learn the importance of each feature channel according to the network loss,and then assign different weights to different channels according to the degree of importance,which improves the classification ability of the network.Bottleneck-Layer significantly reduces the parameters and calculations of the network while reducing the accuracy of classification,and reduces the over-fitting.Finally,based on the saved network parameters,the new network model is trained by parameter migration method,which accelerates the convergence of the network.Based on SE-Block's convolutional network model,this paper achieved 97.5% classification accuracy,98.6% sensitivity and 97.8% specificity on the open source retinal OCT image dataset.At the end of the paper,the method of this paper is compared with the methods used in recent years.The advantages of this method are reflected by the comparison of experimental results.