Multicenter OCT-Based Visual Field Representations via Segmentation-Free 3D CNNs: Forecasting, Longitudinal...

To evaluate segmentation-free three-dimensional convolutional neural networks (3DCNNs) that derive visual field (VF)-shaped representations from macular optical coherence tomography (OCT) volumes (OCT-based estimated VF; OCT-VF) and to assess their performance in forecasting future Humphrey field analyzer (HFA) measurements and characterizing longitudinal variability in clinical datasets.

The 3DCNN models were trained on 129,007 heterogeneous real-world paired OCT-HFA datasets from 13,366 patients (24,313 eyes) across five institutions using 10-fold cross-validation. Forecasting accuracy for the last HFA visit was evaluated using HFA-only forecasts and hybrid models that combined HFA-anchored intercept with longitudinal slopes derived from both OCT-VF and HFA (AOS-AVG). Longitudinal variability was assessed as residual deviations from linear trends using jackknife resampling.

The AOS-AVG yielded the lowest forecasting errors across endpoints and significantly reduced mean absolute error compared with HFA-only forecasts (pointwise thresholds: 24-2, 2.45 vs. 2.68 dB; 10-2, 2.44 vs. 2.65 dB; both P < 0.001). OCT-VF showed significantly lower residual variability than HFA (MD: 24-2, 0.79 vs. 1.13 dB; 10-2, 0.87 vs. 1.07 dB; both P < 0.001). Eyes classified as progression-positive by OCT-VF but negative by HFA exhibited greater perimetric variability and steeper HFA MD decline than true-negative eyes.

Hybrid forecasting approaches that incorporate OCT-VF-derived slope information can improve prediction of future visual-field outcomes. The lower variability of OCT-VF representations may also enable detection of progression-related signals that remain obscured by perimetric noise in routine longitudinal monitoring.

OCT-VF may serve as a complementary tool for precision disease monitoring alongside current automated perimetry.