Answers are grounded in this paper and cite their sources.
Purpose: To develop the first three-dimensional (3D) optic nerve head (ONH) atlas using artificial intelligence (AI)-based registration and evaluate its use in (1) atlas-referenced retinal nerve fiber layer (RNFL) analysis for glaucoma diagnosis, and (2) strain-based assessment of glaucoma severity. Methods: Large-scale optical coherence tomography (OCT) datasets were registered using REFLECTIVITY-generated tissue segmentations. A healthy atlas (n = 460) and glaucoma atlases for mild (n = 852), moderate (n = 640), and severe (n = 546) disease were built using an AI-driven algorithm incorporating intensity- and structure-based constraints with deformation regularization. Atlas quality was evaluated using interlayer contrast and agreement between warped scans and the template. Atlas-referenced RNFL thickness was obtained by defining Bruch's membrane opening (BMO)-centered reference points along a fixed 3-mm-diameter circle in atlas space and mapping these locations to each subject using the registration deformation field. These measurements were then compared with conventional subject-specific RNFL measurements. Effective strain relative to each atlas was computed by morphing subjects to the atlas template. Atlas-referenced RNFL profiles were evaluated for healthy versus early glaucoma classification, and strain volumes were used for multiclass glaucoma severity classification. Results: The healthy atlas captured normative ONH anatomy, and the glaucoma atlases showed stage-dependent changes (cup deepening, prelaminar thinning, rim narrowing). Registration accuracy was high (normalized cross-correlation = 0.86 ± 0.05; Dice coefficient = 0.90 ± 0.02), with strong inter-layer contrast (0.40 ± 0.03). Atlas-referenced RNFL thickness improved separation between healthy and mild glaucoma (Cohen's d = 0.77 vs. 0.72). Classification using atlas-referenced profiles outperformed conventional fixed-circle measurements (area under the curve [AUC] = 0.91 ± 0.02 vs. 0.89 ± 0.02; DeLong P = 0.003). Mean effective strain differed significantly across severity groups (ANOVA P < 0.05), and convolutional neural networks (CNNs) trained on strain maps achieved strong classification (AUC = 0.79). Conclusions: The ONH atlas provides a population-derived structural reference enabling anatomically consistent RNFL mapping and 3D characterization of structural deviation. Atlas-referenced RNFL measurements improved diagnostic discrimination, and geometric strain maps demonstrated the potential of the framework for characterizing disease-associated anatomical differences.
No takes yet. Share an insight, caveat, or question.
Chuangsuwanich et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: