Classification Of Cataract Fundus Images Based On Deep Learning

Cataract fundus image is an important basis for routine medical diagnosis.Often basing on their own experiences,experienced doctors observe the fundus images to determine the extent of cataracts.The symptoms of cataract are often divided into normal,mild,moderate and severe four degrees.According to the requirements of medical cataract rapid diagnosis and low cost,a computer-based cataract automatic diagnosis system is designed.According to the patient's fundus images,the patient's health report can be quickly obtained and treated promptly.In order to study the classification of cataract fundus images,this paper mainly has done the following work:1.In the image preprocessing stage,histogram equalization and maximum entropy transformation are respectively used to enhance the fundus and vascular information of the images,in order to prepare for the feature extraction in the next stage.2.In the feature extraction stage,the spoke features are extracted from the image after the previous histogram equalization.The blood vessels extracted from the maximum entropy transformation of the image.The CNN features are extracted from the RGB images using the deep learning model.These three features as reference group,are used for classification and identification of different designed classifiers.3.Using the three characteristics of the cataract fundus,SVM,Random Forest and Xgboost algorithm classifiers are constructed respectively,and at the same time the Softmax function of deep learning classification is used to classify.After analyzing the four classification results,it is found that the accuracies of two-class classification and four-class classification based on deep learning are 93%and 90.81%respectively,which are better than the classification accuracies of other classifiers.In addition,in terms of feature presentation,CNN features also performed better than the vascular features,and the spoke features is the worst.4.Using Android Studio as a development tool,all the research work on image preprocessing,feature extraction and algorithm classification is integrated to automatically classify cataract fundus images on the mobile terminal.The purpose of this paper is to extract the features of the images by means of deep learning networks and combine them with other algorithms to make the classification research and strive to improve the accuracy of cataract classification and recognition through the combination of various features and methods.At the same time,the mobilization of cataract fundus image classifier is realized to improve the diagnostic efficiency of doctors,and great progress has been made in the practical application of research results.