Predicting visual fields from optical coherence tomography via an ensemble of deep representation learners.

Researchers

Journal

Modalities

Models

Abstract

To develop and validate a deep learning (DL) method of predicting visual function from spectral domain optical coherence tomography (SDOCT) derived retinal nerve fiber layer thickness (RNFLT) measurements and corresponding SDOCT images.Development and evaluation of diagnostic technology.Two DL ensemble models to predict pointwise VF sensitivity from SDOCT images (model 1 – RNFLT profile only; model 2 – RNFLT profile plus SDOCT image), and two reference models were developed. All models were tested in an independent test-retest dataset comprising 2181 SDOCT/VF pairs; the median of ∼10 VFs per eye was taken as the best available estimate (BAE) of the true VF. The performance of single VFs predicting the BAE VF was also evaluated.Training dataset: 954 eyes of 220 healthy and 332 glaucomatous participants. Test dataset: 144 eyes of 72 glaucomatous participants.Pointwise prediction mean error (ME), mean absolute error (MAE) and correlation of predictions with the BAE VF sensitivity.The median mean deviation was -4.17 (-14.22 – 0.88) dB. Model 2 had excellent accuracy (ME 0.5, standard deviation [SD] 0.8, dB) and overall performance (MAE 2.3, SD 3.1, dB), and significantly (paired t-test) outperformed the other methods. For single VFs predicting the BAE VF, the pointwise MAE was 1.5 (SD 0.7) dB. The association between SDOCT and single VF predictions of the BAE pointwise VF sensitivities was R2 = 0.78 and R2 = 0.88, respectively.Our method outperformed standard statistical and DL approaches. Predictions of BAEs from OCT images approached the accuracy of single real VF estimates of the BAE.Copyright © 2022. Published by Elsevier Inc.