Decision tree CART algorithm for diabetic retinopathy classification

Abstract

Diabetic retinopathy is a retinal disease that affects diabetes patients and the major cause of blindness for age population. It is an asymptomatic disease, which involves changes to blood vessels that can cause them to bleed or leak fluid, causing distortion of vision. Therefore, the blood vessels extraction is very important to help ophthalmologists to recognise this disease at the first stage in order to prevent an eventual loss of vision. Consequently, in this paper, we propose an automatic system for diabetic retinopathy detection from color fundus images. The proposed approach is based on the segmentation of blood vessels and extracts the geometric features, which are used in the early detection of diabetic retinopathy. The Hessian matrix, ISODATA algorithm and active contour are used for the segmentation of the blood vessels, we have used. Finally, we have applied the decision tree CART algorithm to classify images into normal (NO-DR) or DR. The proposed system was tested on the DRIVE and Messidor databases and achieved an average sensitivity, specificity and accuracy of 89%, 99% and 96%, respectively for the segmentation of retinal vessels and 91%, 100% and 93%, respectively for the classification of diabetic retinopathy. Finally, the obtained results indicate that our approach is effective in diabetic retinopathy detection with better accuracy over existing methods, which can help ophthalmologists in early diagnosis.