Effect of Signal-Strength Filtering on 3D Convolutional Neural Network-Based Visual Field Estimation From Macular OCT

To determine whether excluding low-signal-strength optical coherence tomography (OCT) scans from training improves deep learning-based visual field (VF) estimation from macular OCT volumes.

We retrospectively analyzed 79,803 paired OCT and VF examinations from 8511 patients at five institutions. A three-dimensional convolutional neural network (3DCNN) estimated 24-2 and 10-2 pointwise sensitivities and mean deviation (MD) from NIDEK OCT volumes. Using patient-wise 10-fold cross-validation, we compared training on all scans with filtering at signal strength indices (SSIs) ≥ 7; training sets differed in size and SSI composition, and both models were tested on identical held-out samples stratified by SSI. Outcomes were pointwise mean absolute error (MAE) and absolute MD error.

Mean SSI was 8.03 ± 1.65; 85.5% of scans had SSI ≥ 7. In the SSI < 7 subgroup, the all-SSI-trained model yielded lower errors for all endpoints, with reductions of 0.439 dB (24-2 MAE), 0.473 dB (10-2 MAE), 0.417 dB (24-2 MD error), and 0.453 dB (10-2 MD error; all adjusted P < 0.001). In the SSI ≥ 7 subgroup, no significant differences were observed. Error increased as SSI decreased, without a clear threshold at SSI = 7.

In this EfficientNet 3D-based segmentation-free 3DCNN framework, using OCT scans across the full SSI range was associated with lower error in low-SSI held-out scans, with no significant penalty in higher SSI held-out scans in the primary comparison within this dataset.

Including lower-signal-strength NIDEK macular OCT scans in development may improve robustness to lower-SSI inputs in this segmentation-free OCT-based VF framework.