Automated Diagnosis and Grading of Diabetic Retinopathy Using Optical Coherence Tomography

We determine the feasibility and accuracy of a computer-assisted diagnostic (CAD) system to diagnose and grade nonproliferative diabetic retinopathy (NPDR) from optical coherence tomography (OCT) images.

A cross-sectional, single-center study was done of type II diabetics who presented for routine screening and/or monitoring exams. Inclusion criteria were age 18 or older, diagnosis of diabetes mellitus type II, and clear media allowing for OCT imaging. Exclusion criteria were inability to image the macula, posterior staphylomas, proliferative diabetic retinopathy, and concurrent retinovascular disease. All patients underwent a full dilated eye exam and spectral-domain OCT of a 6 × 6 mm area of the macula in both eyes. These images then were analyzed by a novel CAD system that segments the retina into 12 layers; quantifies the reflectivity, curvature, and thickness of each layer; and ultimately uses this information to train a neural network that classifies images as either normal or having NPDR, and then further grades the level of retinopathy. A first dataset was tested by "leave-one-subject-out" (LOSO) methods and by 2- and 4-fold cross-validation. The system then was tested on a second, independent dataset.

Using LOSO experiments on a dataset of images from 80 patients, the proposed CAD system distinguished normal from NPDR subjects with 93.8% accuracy (sensitivity = 92.5%, specificity = 95%) and achieved 97.4% correct classification between subclinical and mild/moderate DR. When tested on an independent dataset of 40 patients, the proposed system distinguished between normal and NPDR subjects with 92.5% accuracy and between subclinical and mild/moderate NPDR with 95% accuracy.

A CAD system for automated diagnosis of NPDR based on macular OCT images from type II diabetics is feasible, reliable, and accurate.