Abstract
Purpose :
Lengthy and unreliable visual field (VF) testing presents a burden to both glaucoma patient and clinician. We retrospectively evaluated the ability to predict VF loss from unsegmented optical coherence tomography (OCT) scans using deep learning (DL) technology.
Methods :
Data used in this study consist of 1643 matched OCT-VF pairs encompassing 998 eyes of 542 patients that visited the glaucoma clinic of the University Hospitals Leuven between 2015-2019. Inclusion criteria were defined as having had a SPECTRALIS® OCT scan with the Glaucoma Module Premium Edition (scanning laser ophthalmoscopy (SLO), 24 radial scans and three circumpapillary rings), as well as a reliable Humphrey Field Analyzer (HFA) 3 exam with the strategy 24-2 SITA Standard (52 test points), performed on the same day. Data was split into train/validation/test sets on patient level, following a 60/20/20 key.
The convolutional neural network used across all experiments was an Xception model pretrained on ImageNet. The 52 output nodes feature a linear activation function to allow for regression. The model was trained using mean squared error loss and Adam optimizer. Four models were trained (using the 3.5mm, 4.1mm, 4.7mm circumpapillary scans and SLO images), which were also compared with a combined model. Analysis on the 24 radial scans was beyond the scope of this study. Performance was evaluated using Pearson's r and mean absolute error (MAE). 95% confidence intervals were obtained through bootstrap sampling.
Results :
The circumpapillary scan with the largest radius (4.7mm) achieved the best performance among all individual models (r=0.77 [0.72-0.82], MAE=5.08 [4.68-5.50]). Models trained on circumpapillary OCT scans significantly outperformed the model trained using SLO images (r=0.65 [0.58-0.71], MAE=5.79 [5.23-6.38]). Model combination resulted in an r=0.78 [0.74-0.83] and MAE=4.86 [4.43-5.29]. Performance was comparable on the test set (r=0.79 [0.75-0.82], MAE=4.76 [4.40-5.15]).
Conclusions :
This study is the first to report on a DL system that predicts individual VF threshold values from unsegmented OCT scans. The average correlation of 0.79 exceeds the performance of prior work that leveraged retinal layer thickness info using non-DL techniques. Fast, consistent VF prediction from OCT could become an ersatz solution for visual function estimation in patients that are unable to complete HFA exams.
This is a 2021 ARVO Annual Meeting abstract.