Reliability-Weighted Multimodal Deep Learning for Glaucoma Detection Across Heterogeneous Clinical Sites
Ophthalmology Science · 2026-11 · 101460
DOI: 10.1016/j.xops.2026.101460 ↗Kurz gefasst
Fachartikel in der Zeitschrift Ophthalmology Science, 2026-11. Verfasst von Akram Belghith und 11 weiteren Autor:innen. Die Arbeit nennt 18 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Purpose
To develop and evaluate a multimodal deep learning framework for glaucoma detection that dynamically weights three imaging modalities obtained during a single widefield optical coherence tomography (OCT) session and to assess the contributions of reliability-weighted fusion and domain-adversarial training.
Design
Cross-sectional diagnostic-accuracy study.
Participants
The training set comprised 4,932 healthy eyes and 207 eyes with optic neuropathy consistent with glaucoma (ON-G) from 10 optometry practices. The held-out test set comprised 233 ON-G eyes and 396 healthy eyes from independent cohorts.
Methods
Models using the retinal nerve fiber layer (RNFL) thickness map, RNFL circular B-scan, and fundus photograph individually and in combination were evaluated. The trimodal model used a ConvNeXt-Base encoder, temperature-scaled reliability-weighted fusion, and a domain-adversarial neural network. Equal-weight and no-domain-adaptation ablations were also evaluated. Main Outcome Measures Age- and sex-adjusted AUROC with bootstrap 95% confidence intervals, sensitivity at 95% and 80% specificity, and area under the precision-recall curve.
Results
Among unimodal models, the RNFL circular B-scan achieved the highest AUROC (0.969; 95% CI, 0.957–0.980), followed by the RNFL thickness map (0.959; 95% CI, 0.944–0.972) and fundus photograph (0.860; 95% CI, 0.829–0.891). The trimodal reliability-weighted model achieved an AUROC of 0.994 (95% CI, 0.989–0.998) and 98.3% sensitivity at 95% specificity. Equal-weight fusion reduced AUROC to 0.980, and removing domain adaptation yielded an AUROC of 0.981.
Conclusions
Reliability-weighted fusion of three imaging outputs from a single widefield OCT session improved glaucoma classification relative to individual modalities and equal-weight fusion.
Zitieren
APA
Belghith, A., Hood, D. C., Tsamis, E., Gonzalez, R., Bowd, C., Jalil, J., Fazio, M. A., Girkin, C. A., Owsley, C., Weinreb, R. N., Christopher, M., & Zangwill, L. M. (2026). Reliability-Weighted Multimodal Deep Learning for Glaucoma Detection Across Heterogeneous Clinical Sites. Ophthalmology Science, 101460. https://doi.org/10.1016/j.xops.2026.101460
Vancouver
Belghith A, Hood DC, Tsamis E, Gonzalez R, Bowd C, Jalil J, et al. Reliability-Weighted Multimodal Deep Learning for Glaucoma Detection Across Heterogeneous Clinical Sites. Ophthalmology Science. 2026:101460. doi:10.1016/j.xops.2026.101460
BibTeX
@article{Belghith2026,
author = {Belghith, Akram and Hood, Donald C. and Tsamis, Emmanouil and Gonzalez, Ruben and Bowd, Christopher and Jalil, Jalil and Fazio, Massimo A. and Girkin, Christopher A. and Owsley, Cynthia and Weinreb, Robert N. and Christopher, Mark and Zangwill, Linda M.},
title = {Reliability-Weighted Multimodal Deep Learning for Glaucoma Detection Across Heterogeneous Clinical Sites},
journal = {Ophthalmology Science},
year = {2026},
pages = {101460},
doi = {10.1016/j.xops.2026.101460}
}
EUVIMED