Deep Learning Differentiates Papilledema, NAION, and Healthy Eyes With Unsegmented 3D OCT Volumes

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Neuro-ophthalmology

Background:

• Differentiating optic nerve head (ONH) swelling causes like non-arteritic anterior ischemic optic neuropathy (NAION) and papilledema is clinically significant.
• Previous deep learning (DL) models have shown success with fundus photos but not fully utilized 3D optical coherence tomography (OCT) volumes.
• Specific ONH and peripapillary retina (PPR) OCT features can distinguish NAION from papilledema.

Purpose of the Study:

• To develop and validate a DL approach using 3D OCT volumes to reliably differentiate NAION, papilledema, and healthy eyes.
• To evaluate the diagnostic performance of DL models utilizing the entire OCT volume, PPR, and ONH regions separately.

Main Methods:

• A retrospective review of 4619 spectral domain ONH OCT scans from 1539 eyes (NAION, papilledema, healthy) for internal validation.
• External validation was performed on an additional 1663 scans from 742 eyes.
• Three ResNet 3D-18 models were fine-tuned: entire OCT volume, PPR only, and ONH only.

Main Results:

• The model using the entire OCT scan achieved 94.9% accuracy internally and 90.1% externally, with high AUC-ROC (0.986 internal, 0.977 external) and F1 scores.
• The PPR-only model showed strong performance (94.2% accuracy internally, AUC-ROC 0.966).
• The ONH-only model achieved 85.0% accuracy internally with an AUC-ROC of 0.965.

Conclusions:

• A DL model utilizing the whole ONH OCT scan is a robust tool for differentiating causes of swollen ONH.
• Both ONH and PPR regions contribute to differentiating these optic nerve disorders.
• Automated DL approaches show significant potential in assisting the diagnosis of acquired optic disc swelling.