Abstract
Purpose: :
To develop a novel, efficient screening system to automatically analyze fundus photographs to detect diabetic retinopathy (DR) and grade its severity.
Methods: :
We developed a three-stage algorithm to automatically detect and grade the severity of DR from color fundus images. The first stage extracts the foreground, comprising candidate DR lesions, from the background, comprising the disc and major vasculature using a novel Region of Intersection (RoI) algorithm. The second stage, a classifier, rejects false positives, identifies hard exudates and cotton wool spots as bright lesions, and identifies hemorrhages and microaneurysms as dark lesions. Support Vector Machine (SVM) and Gaussian mixture model (GMM) classifiers were evaluated singly and in combination. Classifiers were trained on 40 images from the public DIARETDB1 dataset and 70 images from a local dataset consisting of images from 20 subjects (40 eyes, 7 images per eye). The third stage analyzes the number and type of lesions per eye to generate a severity grade. The first stage was tested against 539 images from the public DRIVE, DIARETDB0, and DIARETDB1 datasets and the local dataset. The second stage was tested on the remaining 49 images from the DIARETDB1 database and 210 local dataset images. The third stage was tested on the local dataset, which was manually graded for DR severity.
Results: :
The first-stage RoI algorithm detected the disc and major vasculature with 100% accuracy. The best second stage results for bright lesions using DIARETDB1 images were obtained with a combination of SVM and GMM classifiers having 91% sensitivity, 92% specificity, 79% positive prediction value (PPV), 96% negative prediction value (NPV) and 97% area under ROC curve (AUC). For the dark lesions the best results were for the SVM classifier, having 96% sensitivity, 89% specificity, 99% PPV, 76% NPV and 93% AUC. For the local data set, the results for bright lesions were 90% sensitivity, 97% specificity, 90% PPV, 98% NPV and 95% AUC; for Dark lesions the results were 67% sensitivity, 98% specificity, 85% PPV, 94% NPV and 94% AUC. The third stage graded DR as none, mild, moderate and severe with 95% accuracy.
Conclusions: :
An automated screening system can reliably screen fundus images to detect and grade DR. This may improve the efficiency of care delivery.
Keywords: detection • diabetic retinopathy • optic disc