ROC Analysis of Classifiers in Automatic Detection of Diabetic Retinopathy using Shape Features of Fundus Images

Abstract—ROC Analysis of Classifiers in Automatic Detection of Diabetic Retinopathy using Shape Features of Fundus Images. Data < Final Year Projects > Mining Techniques help in discovering useful information from the available data. Classification, one of the data mining techniques finds its application in many areas, making rapid advancements in the field of biology and medicine. Diabetic Retinopathy, a threatening retinal disease places high necessity for computational approaches to automatically detect the disease. Shape related features were extracted from the masked retinal images. Sixty two classification algorithms were run on the extracted features and the performance of the classifiers were evaluated using cross validation with varying folds of 3, 5, 10 and 30. The results were compared using area under the ROC curve (AUC). Ten classification algorithms yielded area greater than 0.9 for various folds. AdaBoostM1 with Decision Stump provided the best performance with area under the curve of 0.996, specificity of 100%, sensitivity of 93.33% and accuracy of 96.67%.