Abstract
Purpose: :
To evaluate the sensitivity of retinal analysis software to identify nonproliferative diabetic retinopathy and to differentiate diseased from normal retinas.
Methods: :
A retrospective study was performed on digital fundus photos of 14 eyes obtained from a database of known diabetics treated through Aravind Eye Hospitals. An automated retinal analysis program was developed at New York University’s Courant Institute of Mathematical Sciences to recognize lesions of nonproliferative diabetic retinopathy, including dot-blot hemorrhages, microaneurysms, and hard exudates. This adaptive learning program was given a training data set of 12 images, after which it was tested on 3 randomly selected images. Sensitivities were determined for the program in comparison with an ophthalmologist.
Results: :
The program correctly identified 234,439 (80%) background structure pixels, 2,555 (83%) hemorrhage pixels, and 4,322 (93%) exudate pixels. However, it occasionally mislabeled structures such as optic disc and retinal vasculature as diabetic lesions.
Conclusions: :
Automated retinal analysis can be a useful adjunct to clinical evaluation in diabetic retinopathy screening. Especially in medically underserved or remote areas, automated analysis may increase the volume of patients to be screened for diabetic retinopathy. With further post-processing, normal eyes should no longer be misidentified by the computer program. This will allow eye care professionals more time to devote to patients in underserved areas.
Keywords: diabetic retinopathy