Related Experiment Video
Updated: Sep 19, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep Learning Differentiates Papilledema, NAION, and Healthy Eyes With Unsegmented 3D OCT Volumes
David Szanto1, Jui-Kai Wang2, Brian Woods3
1From the Department of Ophthalmology (D.S., M.J.K.), Icahn School of Medicine at Mount Sinai, New York, Texas, USA.
Objective:
Deep learning (DL) has been used to differentiate papilledema from healthy eyes and optic disc elevation on fundus photos. As we described optic nerve head (ONH) and peripapillary retina (PPR) optical coherence tomography (OCT) features that distinguish non-arteritic anterior ischemic optic neuropathy (NAION) from papilledema, we hypothesized that a DL approach using the full 3D OCT volume could reliably differentiate NAION, papilledema and healthy eyes.
Design:
This retrospective review analyzed OCT scans from eyes with acute NAION, papilledema, and healthy eyes from randomized and nonrandomized clinical trials.
Participants:
We investigated a total of 4619 raw spectral domain ONH volume scans from 1539 eyes, including 1138 from eyes with idiopathic intracranial hypertension (IIH, Frisén grade ≥ 1), 648 from eyes with acute NAION, and 2833 scans from healthy eyes. We performed external validation on an additional 1663 scans from 742 eyes across these groups.
Methods:
We fine-tuned 3 ResNet 3D-18 models: one with the entire OCT volume, one with the PPR, and one with the optic nerve head excluding the PPR. We then evaluated the models on an external validation set.
Main Outcome Measures:
The primary outcome measures were accuracy, area under the Receiver Operating Characteristic curve (AUC-ROC), and weighted precision, recall, and F1 scores.
Results:
Our model classified the 3 conditions using the entire scan with an internal validation accuracy of 94.9%, macro-average AUC-ROC of 0.986 with weighted F1 scores ranging from 0.93 to 0.95. In external validation, the entire scan model had an accuracy of 90.1% with a macro-average AUC-ROC of 0.977 and weighted F1-score range of 0.89 to 0.94. The PPR alone model attained an accuracy of 94.2%, with a macro-average AUC-ROC of 0.966 and weighted F1-score range of 0.81 to 0.88. The ONH alone model reached an accuracy of 85.0% with an AUC-ROC of 0.965 and weighted F1-score range of 0.84 to 0.89.
Conclusion:
Our findings demonstrate that the model using the whole ONH OCT scan is a robust diagnostic tool for differentiating causes of swollen ONH. Changes in the PPR due to ONH swelling as well as ONH alone can also differentiate the disorders. The results reinforce the potential of automated approaches in assisting in the diagnosis of acquired optic disc swelling.
