Diabetic retinopathy screening using deep neural network.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy is a leading cause of vision loss.
• Early detection through screening programs is crucial for preventing blindness.
• Current screening methods can be resource-intensive.

Purpose of the Study:

• To evaluate the efficacy of a deep neural network (DNN) in detecting referable diabetic retinopathy.
• To assess DNN performance using both a local (New Zealand) and an international diabetic retinal image database.

Main Methods:

• A retrospective audit was conducted using 485 retinal images from the Otago database and 1200 from the Messidor international database.
• A DNN was trained and tested to identify referable diabetic retinopathy, defined as moderate or worse retinopathy or exudates near the fovea.
• Performance was evaluated using receiver operating characteristic (ROC) curves, sensitivity, and specificity.

Main Results:

• The DNN achieved an Area Under the ROC Curve (AUC) of 0.901 (84.6% sensitivity, 79.7% specificity) for the Otago dataset.
• For the Messidor dataset, the DNN achieved a higher AUC of 0.980 (96.0% sensitivity, 90.0% specificity).
• These results indicate high accuracy in detecting referable diabetic retinopathy.

Conclusions:

• Deep neural networks demonstrate significant potential for accurate diabetic retinopathy detection in screening programs.
• The DNN's performance was comparable or superior to established benchmarks.
• Integration into community screening is feasible, pending further validation for diabetic macular edema detection.