Related Experiment Video
Updated: Jan 3, 2026
10:46
Doppler Optical Coherence Tomography of Retinal Circulation
Published on: September 18, 2012
19.2K
Human Versus Machine: Comparing a Deep Learning Algorithm to Human Gradings for Detecting Glaucoma on Fundus
Alessandro A Jammal1, Atalie C Thompson2, Eduardo B Mariottoni2
1Vision, Imaging and Performance Laboratory (VIP), Duke Eye Center and Department of Ophthalmology, Duke University, Durham, North Carolina, USA; Department of Ophthalmology, State University of Campinas, Campinas, Brazil.
American Journal of Ophthalmology
|November 16, 2019
Summary
A deep learning algorithm showed comparable or superior performance to human graders in detecting glaucoma-related vision loss by analyzing retinal nerve fiber layer damage, suggesting its potential for population screening.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Glaucoma diagnosis relies on detecting retinal nerve fiber layer (RNFL) damage.
- Accurate RNFL assessment is crucial for early glaucoma detection and management.
- Automated methods are being explored to enhance diagnostic efficiency and accuracy.
Purpose of the Study:
- To compare the diagnostic performance of a machine-to-machine (M2M) deep learning (DL) algorithm with human gradings.
- To evaluate the DL algorithm's ability to quantify RNFL damage from fundus photographs.
- To assess the algorithm's effectiveness in identifying glaucomatous optical neuropathy (GON).
Main Methods:
- An M2M DL algorithm was trained using RNFL thickness parameters from spectral-domain optical coherence tomography.
- The algorithm analyzed 490 fundus photos, comparing its predictions to glaucoma specialist gradings.
- Performance was evaluated using Spearman correlations with standard automated perimetry (SAP) and area under the receiver operating characteristic curves (AUC).
Main Results:
- The M2M DL algorithm's predicted RNFL thickness showed a stronger correlation with SAP mean deviation (rho=0.54) than human gradings (rho=0.48).
- The DL algorithm achieved a significantly higher partial AUC (0.529) compared to human graders (0.411) in discriminating glaucomatous visual field defects.
- These findings indicate superior or equivalent performance of the DL algorithm.
Conclusions:
- The M2M DL algorithm demonstrates performance on par with or exceeding human graders in detecting glaucomatous visual field loss.
- This deep learning approach shows potential for replacing human graders in large-scale glaucoma screening programs.
- Automated RNFL analysis using DL offers a promising tool for public health initiatives in glaucoma detection.
