Prediction of Visual Field Progression with Optic Disc Photographs and OCT with A Deep Learning Model

Abstract

Purpose : To test the hypothesis that baseline or serial optic disc photographs (ODP) along with baseline retinal nerve fiber layer (RNFL) OCT data can predict subsequent visual field (VF) progression with deep learning (DL).

Methods : 3,079 eyes (1,765 patients) with ≥3 ODP, 5 or more 24-2 VFs, and ≥3 years of follow-up were included. Serial VF mean deviation (MD) rates of change were estimated with linear regression starting with 5 visits adding visits sequentially until final visit. VF progression was defined as a negative slope with p<0.05 at two consecutive visits and at final visit. A convolutional neural network (NN) with a ResNet152 backbone and pre-trained on ImageNet with an additional Siamese NN for adding second and third ODPs was designed. Global and clock hour RNFL thickness, baseline MD, and age labels were concatenated into a feature vector before being interconnected in the fully connected layer with the ResNet outputs. Area under ROC curve (AUC) was used to evaluate model’s performance for prediction of VF progression.

Results : Average (SD) follow-up time and baseline VF MD for stable and progressing eyes were 7.8 (4.7) and 10.3 (5.1) years and –3.2 (4.9) and –4.0 (5.0) dB, respectively. VF progression was detected in 900 eyes (29%). Mean (SD) time to progression (TTP) in the progressing group was 6.8 (4.2) years. Mean (SD) time interval between the first, second, and third ODP and TTP was 7.5 (5.0), 4.0 (2.8) and 2.0 (1.4) years, respectively. AUC (95% CI) for the model with only baseline ODP was 0.659 (0.588-0.729), whereas the model with baseline ODP and OCT performed significantly better (AUC=0.813; 95% CI: 0.757-0.869, p<.001). After including the first 2 or 3 ODPs and baseline OCT data, AUC increased to 0.860 (0.811-0.909) and 0.894 (0.852-0.935), respectively; both were significantly higher than prior less complex models (p<.027 for all) (Table and Figure 1). The most complex model had the highest AUC (0.911) for prediction of fast progression (MD rates <–1 dB/year).

Conclusions : Our DL model using baseline OCT and serial DPs is able to predict VF progression years ahead of time with high accuracy. This model may be implemented in the clinical setting for earlier detection of glaucoma progression once validated.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.