OCT-based Visual Field Estimation Using Segmentation-free 3D CNN Shows Lower Variability than Subjective Standard Automated Perimetry

Purpose: To train and evaluate a segmentation-free 3D convolutional neural network (3DCNN) model for estimating visual field (VF) from optical coherence tomography (OCT) images and to compare the residual variability of OCT-based estimated VF (OCT-VF) with that of Humphrey Field Analyzer (HFA) measurements in a diverse clinical population. Design: Retrospective cross-sectional study. Participants: 5,351 patients (9,564 eyes) who underwent macular OCT imaging and Humphrey Field Analyzer (HFA) tests (24-2 or 10-2 test patterns) at a university hospital from 2006 to 2023. The dataset included 47,653 paired OCT-VF data points, including various ocular conditions. Methods: We trained a segmentation-free 3DCNN model based on the EfficientNet3D-b0 architecture on a comprehensive OCT dataset to estimate VF. We evaluated the model's performance using Pearson's correlation coefficient and Bland‒Altman analysis. We assessed residual variability using a jackknife resampling approach and compared OCT-VF and HFA datasets using generalized estimating equations (GEE), adjusting the number of VF tests, follow-up duration, age, and clustering by eye and patient. Main Outcome Measures: Correlations between estimated and measured VF thresholds and mean deviations (MDs), and residual variability of OCT-VF and HFA. Results: We observed strong correlations between the estimated and measured VF parameters (Pearson's r: 24-2 thresholds 0.893, MD 0.932; 10-2 thresholds 0.902, MD 0.945; all p < 0.001). Bland‒Altman analysis showed good agreement between the estimated and measured MD, with a slight proportional bias. GEE analysis demonstrated significantly lower residual variability for OCT-VF than for HFA (24-2 thresholds: 1.10 vs. 2.48 dB; 10-2 thresholds: 1.20 vs. 2.48 dB; all p < 0.001, Bonferroni-corrected), with lower variability across all test points, severities, and ages, thus highlighting the robustness of the segmentation-free 3DCNN approach in a heterogeneous clinical sample. Conclusions: A segmentation-free 3DCNN model objectively estimated VF from OCT images with high accuracy and significantly lower residual variability than subjective HFA measurements in a heterogeneous clinical sample, including patients with glaucoma and individuals with other ocular diseases. The improved reliability, lower variability, and objective nature of OCT-VF highlight its value for enhancing VF assessment and monitoring of various ocular conditions, potentially facilitating earlier detection of progression and more efficient disease management.