Deep Learning with Disc Photos or OCT Scans in Glaucoma Detection

Area of Science:

• Ophthalmology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Glaucoma diagnosis relies on detecting functional visual field (VF) impairment.
• Deep learning (DL) models show promise for automated glaucoma detection.
• Comparing DL models using different imaging modalities is crucial for diagnostic accuracy.

Purpose of the Study:

• To compare the accuracy of a DL model using OCT RNFLT maps versus disc photos (DPs) for glaucoma detection.
• To assess the diagnostic performance of these DL models across diverse demographic groups (race, sex, ethnicity).

Main Methods:

• Retrospective cohort study utilizing OCT and DP datasets from 2011-2022.
• Development of DL models trained on OCT RNFLT maps and DPs for glaucoma detection based on VF impairment.
• Included 16,936 image sets with high-quality OCT scans and reliable VF tests.

Main Results:

• The OCT-based DL model achieved a significantly higher area under the curve (AUC) of 0.90 compared to the DP-based model (AUC = 0.86, P < 0.005).
• Superior performance of the OCT model was consistent across racial, sex, and ethnic demographic groups.
• Both models showed variations in accuracy across demographic groups, indicating potential biases.

Conclusions:

• DL models utilizing OCT RNFLT maps demonstrate superior accuracy for glaucoma detection compared to DP-based models.
• The objective and quantitative nature of RNFLT measurements likely contributes to the improved performance.
• Observed demographic disparities emphasize the critical need for diverse and equitable datasets in developing fair AI diagnostic tools.