Related Experiment Video
Updated: Sep 15, 2026
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
Diagnostic Likelihood Ratios for Glaucoma Using Deep Learning-Predicted Retinal Nerve Fiber Layer Thickness from
Neda Nilforoushan1, Douglas R da Costa1, Rafael Scherer1
1From the Bascom Palmer Eye Institute, University of Miami School of Medicine, Miami, Florida.
Purpose:
To translate continuous retinal nerve fiber layer (RNFL) thickness values predicted by a deep learning algorithm from fundus photographs into clinically interpretable diagnostic likelihood ratios for glaucoma.
Design:
Cross-sectional study.
Subjects:
A total of 232 participants with glaucomatous optic neuropathy and 127 healthy controls.
Methods:
A deep learning algorithm trained on spectral-domain OCT (SD-OCT) measurements (machine-to-machine, M2M) was applied to optic disc photographs to predict global RNFL thickness. Participant-level analyses were conducted by selecting the eye with the lower predicted RNFL thickness per participant. Diagnostic performance of M2M-predicted RNFL thickness was compared with SD-OCT measurements using receiver operating characteristic (ROC) analysis. Likelihood ratios were then estimated for continuous predicted RNFL thickness values using a ROC-based framework, allowing calculation of individualized post-test probabilities of glaucoma without dichotomizing results.
Main Outcome Measures:
Diagnostic performance (area under the ROC curve [AUC]) and likelihood ratios derived from continuous M2M-predicted RNFL thickness values.
Results:
Machine-to-machine-predicted RNFL thickness was significantly lower in glaucomatous eyes than in controls (71.3 ± 10.6 μm vs. 95.5 ± 7.2 μm; P < 0.001). Diagnostic discrimination was high for both M2M-predicted RNFL thickness and SD-OCT-measured global RNFL thickness, with AUCs of 0.986 (95% confidence interval [CI], 0.963-0.998) and 0.986 (95% CI, 0.971-0.996), respectively. The two AUCs did not differ significantly (P = 0.987), and M2M predictions retained diagnostic information. Likelihood ratios varied markedly across the range of predicted RNFL thickness values: thinner values were associated with large likelihood ratios that substantially increased post-test probability of glaucoma, whereas thicker values produced likelihood ratios well below unity, effectively reducing disease probability. Intermediate RNFL values were associated with likelihood ratios near 1, indicating limited diagnostic impact.
Conclusions:
Continuous RNFL thickness values predicted by a deep learning algorithm from fundus photographs can be translated into clinically meaningful likelihood ratios that enable individualized probabilistic diagnosis of glaucoma. This framework moves beyond binary classification by preserving the diagnostic information contained in continuous artificial intelligence-derived measurements and provides a practical pathway for integrating deep learning outputs into clinical decision-making.
Financial Disclosures:
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
