Predicting Visual Fields From Optical Coherence Tomography via an Ensemble of Deep Representation Learners

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Spectral domain optical coherence tomography (SD-OCT) provides detailed retinal imaging.
• Retinal nerve fiber layer thickness (RNFLT) is a key indicator of glaucoma and other optic neuropathies.
• Predicting visual function (VF) from OCT data can aid in early diagnosis and monitoring.

Purpose of the Study:

• To develop and validate a deep learning (DL) method for predicting visual function.
• To utilize SD-OCT derived RNFLT measurements and images for visual function prediction.
• To compare DL model performance against traditional methods.

Main Methods:

• Two DL ensemble models were developed: one using RNFLT profile only, and another incorporating both RNFLT and SD-OCT images.
• Models were trained on data from healthy and glaucomatous participants.
• Performance was evaluated on an independent test-retest dataset using pointwise prediction mean error (ME), mean absolute error (MAE), and correlation with the best available estimate (BAE) of true VF.

Main Results:

• The DL model incorporating both RNFLT and SD-OCT images demonstrated excellent accuracy (ME 0.5 dB, MAE 2.3 dB).
• This model significantly outperformed other tested methods, including statistical and single-input DL approaches.
• Predictions from OCT images approached the accuracy of single real visual field tests in estimating the BAE VF.

Conclusions:

• Deep learning models integrating SD-OCT images and RNFLT measurements offer a powerful tool for predicting visual function.
• This approach shows promise in enhancing diagnostic capabilities for conditions affecting visual pathways.
• The developed method provides an accurate and efficient alternative or adjunct to conventional visual field testing.