Using deep leaning models to detect ophthalmic diseases: A comparative study.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Diabetic retinopathy, age-related macular degeneration, and glaucomatous optic neuropathy are leading causes of vision impairment.
• Accurate and timely detection is crucial for preventing vision loss.
• Automated systems offer potential for improving screening efficiency and accessibility.

Purpose of the Study:

• To compare the diagnostic agreement of a deep learning system with non-physician graders and ophthalmologists.
• To evaluate the performance across different experience levels of clinicians.
• To assess detection accuracy for referable diabetic retinopathy, age-related macular degeneration, and glaucomatous optic neuropathy.

Main Methods:

• A deep learning system was developed and validated using over 210,000 fundus photographs.
• A test set of 300 images was graded by five non-physician graders, 47 general ophthalmologists, and compared against a specialist reference standard.
• Area Under the Receiver Operator Characteristic Curve (AUC) was used to quantify agreement and compare performance.

Main Results:

• For referable diabetic retinopathy, the deep learning system achieved an AUC of 0.990, comparable to human graders.
• For age-related macular degeneration, the deep learning system's AUC was 0.945, similar to ophthalmologists.
• For glaucomatous optic neuropathy, the deep learning system (AUC 0.994) outperformed non-physician graders and was comparable to experienced ophthalmologists.

Conclusions:

• The deep learning system demonstrates comparable accuracy to human experts for diabetic retinopathy and age-related macular degeneration detection.
• The AI system shows superior performance over non-physician graders in identifying referable glaucomatous optic neuropathy.
• Deep learning holds significant promise for augmenting ophthalmic diagnostic capabilities.