Classification Of Ophthalmic Diseases Based On Optical Coherence Tomography

With the improvement of life quality,people pay more attention to the health of their eyes and will go to the hospital for a comprehensive examination regularly.Optical coherence tomography(OCT)is the most common technique for ophthalmic examination at present.The classification of OCT images in ophthalmology is one of the important processes of doctors’ diagnosis.Automated classification research are important for assisting doctors’ classification,relieving doctors’ work pressure and improving diagnosis efficiency.In recent years,with the continuous improvement of computer computing power and the wide application of deep neural network in the field of image processing,many scholars have also applied deep neural networks to the task of OCT image classification in ophthalmology.This paper focuses on the research and improvement of OCT image classification algorithm in ophthalmology.The main research contents are as follows.Firstly,the dataset used in this paper is composed of multiple channels,and it needs to be preprocessed for data sorting.Which mainly includes standardized image data processing,classification and labeling according to categories,denoising of images containing noise in the dataset,and finally re-screening and expanding the data.Thus,OCT image dataset that meets the requirements of this project is produced to provide the basis for subsequent training of the network model.Secondly,the existing image classification neural networks such as VGG16,Goog Le Net,Mobile Net and Res Net50 are used to train the dataset.The feature extraction capabilities of different neural networks are compared,and the classification capabilities of different networks are analyzed through visual confusion matrix.Select the better Res Net50 network as the base network for the subsequent improvement.Finally,the Res Net50 network is studied and improved.In order to improve the classification accuracy of the network,the channel attention mechanism is introduced,and the feature extraction ability of the network is improved by strengthening the useful feature channel.As the target area of ophthalmic images is only a small part of the whole image,it has the characteristics of localization.The gated attention mechanism is introduced to better integrate deep features and shallow features,obtain more detailed information of finer scale,and further improve the feature extraction ability of the network and the accuracy of classification.In addition,the fully connected layer is improved and the activation function,loss function,learning rate,and batch size are optimally selected to further improve the performance of the network.In this paper,by fusing two attention mechanisms and improving the fully connected layer,and optimize the activation function,learning rate,etc.The experimental results show that the improved network model is superior to other models in the task of OCT image classification in terms of the accuracy rate,recall rate and specificity of each category,and the overall classification accuracy is also at a high level.So,the improved network model has good classification performance and can basically meet the task requirements of assisting physicians in classification.