Abstract
Purpose :
To develop and validate deep learning models that estimate visual field test results such as visual field index (VFI) and mean deviation (MD) from raw optical coherence tomography (OCT) data of healthy and glaucomatous eyes.
Methods :
Macular and optic nerve head (ONH) OCT scans (Cirrus HD-OCT; Zeiss, Dublin, CA) were acquired on 115 healthy subjects and 464 glaucoma patients at multiple visits resulting in 504 volumes of healthy and 3651 volumes of glaucomatous eyes. A 3D convolutional neural network (CNN) was trained to infer VFI and MD obtained from 24-2 Humphrey visual field tests conducted at the same visit. The volumes were downsampled to 64x64x128 due to GPU memory limitations but no further pre-processing was performed. The data set was split into 3329 eyes for training, 422 eyes for validation and 404 eyes for testing, ensuring that eyes belonging to the same patient were kept in the same set. Network training was posed as a regression problem with mean square error as the loss function. The 5-fold cross-validation accuracy of the system was evaluated for root mean square error (RMSE) and Pearson’s correlation (PC). We also computed class activation maps (CAM) to identify regions within OCT volumes that the CNN found important for VFI performance. The system was compared to classical machine learning (ML) methods (Random Forests and others) trained on a set of 23 measurements: peripapillary RNFL thickness at 12 clock-hours and four quadrants, mean retinal nerve fiber layer (RNFL) thickness, rim and disc area, horizontal/vertical cup-to-disc ratio, cup volume and age at visit.
Results :
The best results were achieved on ONH scans when inferring VFI with a PC of 0.88±0.035 for the CNN and a significantly lower (p < 0.01) PC of 0.79±0.029 for the Random Forest classifier. Prediction of MD was equally accurate with a PC of 0.88±0.023 on ONH scans. The corresponding RMSE were 12.2±1.55% for VFI and 4.1±0.35dB for MD. PC on the macular scans (using the CNN) were 0.86±0.012 for VFI and 0.85±0.014 for MD. Interestingly, CAMs showed intra-retinal layers outside of the RNFL highlighted, indicating significant contribution to inferring VFI (Fig.1).
Conclusions :
A deep learning approach trained on ONH scans can infer VFI and MD directly from raw OCT volumes. This technique significantly outperformed all classical ML methods trained on classical OCT features.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.