Abstract
Purpose :
To develop and validate a three-dimensional (3D) deep learning (DL) model to predict visual field (VF) global indices, including visual field index (VFI), mean deviation (MD), and pattern standard deviation (PSD), from paired optical coherence tomography (OCT) volumetric optic disc and macula scans.
Methods :
We retrospectively collected 4,892 OCT scans with paired VF tests from patients with confirmed glaucoma in the Chinese University of Hong Kong Eye Center, and randomly split them into training (80%), tuning (5%), and primary validation (15%) sets on patient-level. We modified the ResNet structure with 3D convolutional layers and the global average pooling layer to build the DL model (Figure 1). The input was the paired volumetric OCT optic disc scan of size 64x256x64 pixels and macula scan of size 32x256x32 pixels after downsizing. The ground truths were the true VF global indices including VFI, MD, and PSD. The outputs were predicted VF global indices. We performed Pearson’s correlation coefficients (R) and Bland-Altman analysis to estimate the agreement between predicted and true VF indices. We used mean absolute error (MAE) to evaluate accuracies and differences in predictions. We also evaluated the performance stratified by glaucoma severities.
Results :
Table 1 shows the overall MAEs of VFI, MD, and PSD of 11.86%, 3.92 dB, and 3.06 dB, respectively. Pearson’s correlation coefficients showed that there was a strong correlation between the predicted and true MD (R = 0.70), a moderate correlation between the predicted and true VFI (R =0.69) and a weak correlation between the predicted and true PSD (R = 0.36). The Bland-Altman plots indicate no significant systemic and proportional bias in VFI and MD predictions (Figure 2). When stratified by glaucoma severities, The MAE of predicted VFI was lowest in mild glaucoma and highest in severe glaucoma, while the MAEs of predicted MD and PSD were lowest in moderate glaucoma and highest in severe glaucoma.
Conclusions :
The proposed 3D DL model could predict VF global indices from OCT volumetric optic disc and macula scans with good agreements and relatively small errors, especially for MD prediction in patients with mild glaucoma. With further refinement, it would possibly facilitate OCT as a screening tool for glaucoma early detection by illustrating structural changes and predicting functional changes from volumetric scans.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.