Microaneurysms Detection Of Diabetic Retinopathy In Color Fundus Images

Diabetic retinopathy(DR)is one of the major complications of diabetes.If not diagnosed and treated in time,it may lead to visual impairment and blindness.With regular eye examinations and timely treatment,the risk of vision loss in DR patients can be reduced.The analysis of DR through computer-aided technology can assist ophthalmologists to make a diagnosis more efficiently,and it is of vital importance to carry out a large-scale regular screening of DR.Microaneurysms(MAs)are the earliest clinical symptoms of DR.The accurate and reliable detection of microaneurysms is of great significance in the early diagnosis of DR.Methods for microaneurysms detection based on machine learning using different features were proposed.The main work and innovations are as follows:(1)Blood vessels enhancement and segmentation.In color fundus images,retinal blood vessel appears similar with microaneurysms on color and brightness.In order to prevent its interference on microaneurysms detection,blood vessel was enhanced by applying an improved method based on the eigenvalue analysis of the Hessian matrix,further segmented and eliminated.Compared with the Hessian matrix-based enhancement methods commonly used before,the method used in this study is more robust to low-intensity changes of images,getting more uniform enhancement and better results of blood vessel segmentation.(2)Microaneurysms candidate regions extraction.Due to the tiny structure of microaneurysms,it is difficult for microaneurysms detection in the whole image.The method of microaneurysms candidate regions processed into small patches was applied.Analysis of contrast characteristics was performed on the image after blood vessel eliminated,and then connected component analysis based on shape characteristics of microaneurysms was used to extract candidate regions from fundus images.Finally,the image was divided into small patches according to candidate positions for subsequent detection.(3)Microaneurysms detection using two combinations modes of features and classifiers.Firstly,histogram of oriented gradient(HOG)feature and support vector machine were used for microaneurysms detection.Secondly,the application of directional local contrast(DLC)for microaneurysm candidate classification was proposed,and combined with overall features.Three machine learning classification algorithms were applied then for candidate classification,comparing their performance,and different voting rules were used for multi-classifier fusion to reach the final classification results.Experiments show that the multi-classifier fusion based on average voting or weighted voting rules performs better than single classifier.Compared with the method using histogram of oriented gradient feature and support vector machine,the result of the method using directional local contrast feature and multi-classifier fusion was better,of which area under curve(AUC)of receiver operating characteristic(ROC)was 0.901 and score of Free-response ROC(FROC)was 0.459.Lastly,compared with advanced algorithms in the past three years,the proposed microaneurysms detection methods have achieved good performance,reached the advanced level of existing algorithms,and require less computation time.