Predicting Humphrey 24-2 Visual Field Sensitivity from OCT: A Comparison of Thickness Maps and Printed Summary Parameters

Optical coherence tomography (OCT) measures structural damage in glaucoma objectively and quantitatively. The extent and location of that damage correspond to the severity and position of visual field loss. One OCT session yields two representations of the same scan: thickness maps and the summary parameters the device prints. How well either representation captures that structure–function relationship is unclear, and the two have not been compared within one cohort. We compared them by predicting the 52 non-blind-spot sensitivities of the Humphrey 24-2 visual field from 277 paired OCT and visual field examinations in 144 patients. A convolutional neural network was trained on the thickness maps, and 52 pointwise XGBoost regressors on the summary parameters. The two predictions were combined by late fusion. Fusion reached a pooled root mean square error of 8.069 dB, against 8.660 dB for the parameter branch and 8.540 dB for the image branch. The gain over the parameter branch held across five backbones, three seeds, and a held-out test set. A five-backbone image ensemble also outperformed the parameter branch by 0.852 dB per eye. Fusing the summary parameters into that ensemble, however, did not detectably improve on it. The fusion gain over the parameter branch is, moreover, smaller than the pointwise test–retest variability of standard automated perimetry. In a matched comparison, removing spatial detail within sectors cost the five-backbone ensemble less than 0.2 dB per eye in cross-validation, whereas networks given the same sector values outperformed gradient boosting. The advantage of the map-based pipeline thus lay in the model rather than in spatial detail, at least in cross-validation; the small held-out set confirmed neither result.