Geometric Model for Semi-Automatic Foveal Localization Based on the Optic Nerve Head-Fovea Axis

To evaluate the accuracy of a semi-automatic foveal localization method based on optic nerve head (ONH)-fovea geometry.

This cross-sectional validation study included 46 eyes. For each image, the foveal center was manually selected before optic nerve head (ONH) segmentation to provide a partially blinded reference standard for the semi-automatic method. The primary endpoint was the median optimized localization error, tested against a prespecified 212 μm Euclidean threshold. Secondary outcomes included distance-based success rate, Bland-Altman directional bias, optimized angle analysis, and overlap-based grading using paired 150 μm central-foveal rings. Grade 3 indicated optimal localization, with the model-derived point within 50 μm of the manual reference; Grade 2 indicated at least 50% ring overlap; Grade 1 indicated at least 33% ring overlap; and Grade 0 indicated less than 33% overlap or error exceeding 212 μm.

The median optimized localization error was 100 μm, with a bootstrap 95% confidence interval of 78-144 μm. The interquartile range was 56-211 μm, and the mean ± standard deviation was 126 ± 90 μm. The median error was significantly below the 212 μm clinical threshold (p < 0.001). Overall, 36 of 46 eyes (78.3%; 95% Wilson CI, 64-88%) had an error below 212 μm. Using the 150 μm grading ring, 33 of 46 eyes (71.7%) achieved Grade ≥1, and 24 eyes (52.2%) achieved Grade ≥2. The optimized angle was 6.37° ± 0.45°.

In normal eyes, the semi-automatic ONH-fovea geometric model localized the fovea in 30° infrared images with clinically acceptable accuracy in most eyes. Evaluation in eyes with macular disease, where direct foveal localization may be challenging, is warranted as a next step.