A Deep Learning Algorithm to Quantify Neuroretinal Rim Loss From Optic Disc Photographs.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Accurate quantification of neuroretinal damage is crucial for glaucoma diagnosis and management.
• Current methods for assessing optic disc damage can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a deep learning (DL) algorithm for quantifying glaucomatous neuroretinal damage.
• To utilize minimum rim width relative to Bruch membrane opening (BMO-MRW) from spectral-domain optical coherence tomography (SDOCT) as a reference standard.
• To assess the DL algorithm's accuracy in detecting glaucoma from fundus photographs.

Main Methods:

• A convolutional neural network (DL) was trained on 9282 pairs of optic disc photographs and SDOCT scans.
• The DL model predicted SDOCT BMO-MRW values from fundus photographs.
• Performance was evaluated by comparing DL predictions to SDOCT measurements and assessing the area under the receiver operating curve (AUC) for glaucoma detection.

Main Results:

• DL predictions of global BMO-MRW highly correlated with SDOCT measurements (Pearson's r = 0.88).
• The mean absolute error of DL predictions was 27.8 μm.
• The DL algorithm achieved high AUCs (0.945) for discriminating between glaucomatous and healthy eyes, comparable to SDOCT (0.933).

Conclusions:

• A DL network can accurately quantify neuroretinal damage on fundus photographs using SDOCT BMO-MRW.
• The developed algorithm demonstrates high accuracy for glaucoma detection.
• This AI approach may reduce the need for manual grading of optic disc photographs in clinical practice.