Abstract The purpose is to evaluate whether numeric optical coherence tomography (OCT) data can predict Humphrey Visual Field Analyzer (HFA) 30–2 mean deviation (MD) using deep learning (DL). In this retrospective study, 1200 eyes (432 glaucoma and 768 normal eyes) that underwent spectral-domain OCT (12 × 9.0-mm high-density scanning) and Humphrey Field Analyzer (HFA) 30–2 testing on the same day were analyzed. Pixel-wise retinal thickness numeric values were directly exported from OCT and input into eight deep learning models, including five convolutional neural networks (ResNet50, VGG16, InceptionV3, EfficientNetB0, and DenseNet121) and three vision transformers (ViT, DeiT, and BEiT). Two distinct prediction tasks were performed: (1) a regression task to predict continuous HFA MD values and (2) a classification task to discriminate VF deterioration at predefined MD thresholds. For the regression task, model performance was assessed using standard regression metrics, and agreement between predicted and measured MD values was additionally evaluated using the Bland–Altman analysis. For the classification task, model performance was evaluated using threshold-based discrimination metrics. Both tasks were evaluated using fivefold cross-validation. The mean age was 63.3 ± 21.5 years, axial length 25.43 ± 1.33 mm, and baseline MD − 5.32 ± 4.33 dB. Among all models, InceptionV3 achieved the best performance, with a mean absolute error of 3.36 dB and a coefficient of determination of 0.53. In classification analyses at clinically relevant MD thresholds, CNN-based models achieved high discrimination performance across early, moderate, and severe visual field loss. We developed an accurate and explainable DL system directly leveraging raw numeric OCT thickness data to predict HFA 30–2 MD values; this predictive DL approach may enhance diagnosis of glaucoma.
Yasuda et al. (Wed,) studied this question.