Pointwise Visual Field Estimation From Optical Coherence Tomography in Glaucoma Using Deep Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Standard automated perimetry is the gold standard for monitoring glaucoma, but it suffers from intrasubject variability.
• Accurate visual field (VF) assessment is crucial for effective glaucoma management.

Purpose of the Study:

• To develop and validate a deep learning (DL) regression model for estimating VF sensitivity from optical coherence tomography (OCT) scans.
• To assess the accuracy of DL models in predicting pointwise and overall VF sensitivity compared to standard perimetry.

Main Methods:

• A customized DL regression model with an Xception backbone was trained and validated.
• The model utilized unsegmented OCT scans from multiple imaging modalities.
• Data from patients with glaucoma examinations, including Humphrey Field Analyzer (HFA) 24-2 SITA Standard (SS) VF tests and SPECTRALIS OCT, were used.

Main Results:

• The DL model achieved a mean absolute error (MAE) of 2.89 dB for mean deviation (MD) estimation, a 54% reduction from baseline.
• For 52 VF threshold values, the model yielded an MAE of 4.82 dB, a 38% reduction from baseline.
• The DL model explained 75% of the variance in MD and 58% in pointwise sensitivity estimation.

Conclusions:

• Deep learning can accurately estimate global and pointwise VF sensitivities.
• These estimations fall within the 90% test-retest confidence intervals of standard VF tests.
• DL-based VF prediction from OCT scans offers a promising alternative for patients with unreliable VF exams.