Performance of a Deep-Learning Algorithm vs Manual Grading for Detecting Diabetic Retinopathy in India

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Over 60 million individuals in India have diabetes, placing them at high risk for diabetic retinopathy (DR).
• Early detection of DR through automated interpretation of retinal fundus photographs can support large-scale screening programs.

Purpose of the Study:

• To prospectively validate the performance of an automated DR detection system in a real-world clinical setting across two distinct sites in India.

Main Methods:

• A prospective observational study involving 3049 patients with diabetes across two Indian eye care centers.
• The automated DR grading system was compared against manual grading by on-site specialists, with adjudication by a panel for disagreements.

Main Results:

• The automated DR system demonstrated high performance, with sensitivity ranging from 88.9% to 92.1% and specificity from 92.2% to 95.2% across the two study sites.
• Performance metrics, including area under the curve (0.963-0.980), indicated that the automated system was equal to or exceeded manual grading accuracy.

Conclusions:

• The automated DR grading system is effective and generalizes well to the Indian patient population in a prospective setting.
• This validates the feasibility of employing automated DR grading systems to enhance the reach and efficiency of diabetic retinopathy screening programs in India.