Prediction of Central Visual Field Measures From Macular OCT Volume Scans With Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate visual field (VF) testing is crucial for diagnosing and monitoring glaucoma and other optic nerve head disorders.
• Macular optical coherence tomography (OCT) provides detailed structural information but does not directly measure VF function.

Purpose of the Study:

• To develop and validate a deep learning (DL) model capable of predicting central 10° visual field (VF) parameters from macular OCT volume scans.
• To assess the accuracy of DL predictions compared to traditional linear models.

Main Methods:

• A dataset of 1121 macular OCT volume scans and corresponding 10-2 VFs from 289 eyes was used.
• A 3D convolutional neural network (3D DenseNet121) was trained to predict VF mean deviation (MD), threshold sensitivities (TS), and total deviation (TD) values.
• Performance was evaluated using 10-fold cross-validation, comparing DL predictions against ground truth and baseline linear models via correlation and mean absolute error (MAE).

Main Results:

• The DL model achieved an average correlation of 0.74 for MD prediction with an MAE of 3.5 dB.
• For TS, DL showed significantly higher correlations (0.71) compared to baseline models (0.52) (P < 0.001), with lower MAE (6.5 dB vs. 7.5 dB).
• Similar improvements were observed for TD predictions, with DL correlations of 0.69 versus 0.48 for baseline models (P < 0.001) and MAE of 6.1 dB versus 7.8 dB.

Conclusions:

• Macular OCT volume scans can be utilized with deep learning to predict global and local central VF parameters with clinically relevant accuracy.
• This approach holds promise for confirming and supplementing clinical VF assessments, potentially improving diagnostic efficiency and patient care.