AI-Driven 3D Structural Analysis of the Optic Nerve Head to Predict Visual Field Loss from Optic Disc Drusen

Abstract

Purpose : To assess whether AI algorithms exploiting 3D structural features of the optic nerve head (ONH) could predict visual field loss from optic disc drusen (ODD).

Methods : Optical coherence tomography (OCT) raster scans of the ONH were acquired for both eyes of 64 subjects with confirmed optic disc drusen (ODD). For each scan, neural tissues, connective tissues, and ODD regions were segmented in 3D using previously described algorithms (Neurology, 2023, 100(2):e192-e202; and the software Reflectivity, Abyss Processing Pte Ltd; Figure 1A). To predict visual field loss (specifically, mean deviation [MD]) from ONH structure, we used 2 approaches: (1) a machine learning algorithm (Random Forest) that used the drusen volume (in mm³) as input along with 40 other structural parameters derived from the segmentation (e.g. minimum-rim-width and lamina cribrosa depth); (2) a geometric deep learning algorithm (PointNet) that took 3D point clouds of the ONH as inputs (Figure 1B), together with local tissue boundary and thickness information. Training and testing sets were split equally. To assess performance, we reported R² values on the testing sets (predicted vs ground-truth MD) as mean ± standard deviation following a Monte Carlo cross-validation process with 100 iterations.

Results : Using Random Forest, the drusen volume could predict MD values but with a relatively weak fit between predicted and ground-truth MD values (R² = 0.13±0.13). When the model was augmented with an additional 40 structural parameters, a significant increase in performance was observed (R² = 0.48±0.09; Figure 1C). PointNet, which also exploited 3D structural information, demonstrated a similar level of performance (R² = 0.46±0.06).

Conclusions : Drusen volume alone is a weak predictor of MD. Integrating a comprehensive 3D analysis of ONH structures, as has been proposed for glaucoma, can significantly improve predictive reliability. With further validations, our AI algorithms could be particularly beneficial to clinicians managing ODD, especially in scenarios where familiarity with this uncommon condition is limited.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. (A) 3D automated segmentation of optic nerve head (ONH) tissues, including optic disc drusen regions. (B) ONH 3D point cloud that serves as input to our PointNet algorithm. (C) AI-predicted mean deviation versus ground truth.