Detecting Glaucoma Across Multi-ethnic Myopic and Non-Myopic Populations Using an Uncertainty-Aware Vision Transformer: A Multicentre Model Development and Validation Study
Raghavan Lavanya, Yangqin Feng, Ten Cheer Quek, Quan V. Hoang, Linda Yi-Chieh Poon, Jost B. Jonas, Ya Xing Wang, Vinay Nangia, et al.
cs.CV · cs.AI
Quick verdict
Credible ViT glaucoma screening model with strong external validation (16 datasets, AUROC 86–100%); headline metrics aided by excluding uncertain cases.
Summary
The authors present RetiGON, a two-stage deep learning system for detecting glaucoma from colour fundus photographs: a MobileNetV2-UNet performs quality control and optic disc localisation, and a ViT-B/16 classifier with entropy-based uncertainty scoring (trained with cost-sensitive loss and SWAD) classifies the disc. Training used 56,483 fundus images from Singapore population-based (SEED) and clinic cohorts, 57.1% myopic and 14.4% highly myopic, with glaucoma labels standardised via clinical, OCT, and perimetry data; glaucoma suspects were excluded. On the internal test set the AUROC was 97.8% (98.7% after excluding <10% uncertainty-flagged cases; sensitivity 94.5%, specificity 97.3%). Across 16 external datasets from eight countries, AUROCs ranged 86.4–99.6%; in high-myopia subgroups, 85.5–93.3% externally and 97.8% internally. In an exploratory study of 50 highly myopic eyes, the model beat eight specialists and two graders reading photographs alone (92.0% vs 70.6% mean accuracy, p=0.008) and matched specialists given full clinical data (p=0.064). Limitations: headline metrics depend on excluding uncertain cases, cohorts are predominantly Asian, advanced myopic maculopathy is absent, the clinical comparison is small, and no code or weights are public (commercially licensed).
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A ViT-B/16 model trained on standardized multimodal-label, myopia-rich data detects glaucoma from CFP with internal AUROC 98.7% (97.8% unfiltered).
- The model generalizes across 16 external datasets from eight countries with AUROCs of 86.4–99.6%, including HM subgroups at 85.5–93.3% externally.
- In highly myopic eyes, RetiGON outperforms ophthalmologists and trained graders reading CFP alone (92.0% vs 70.6% mean accuracy, p=0.008) and matches specialists with full clinical data (p=0.064).
- Entropy-based uncertainty flagging identifies unreliable predictions enriched for high myopia and challenging discs, enabling a clinician-referral workflow.
Strengths
Standardized multimodal ground truth at training
Labels combine slit-lamp exam, tonometry, gonioscopy, CCT, biometry, OCT and Humphrey VF criteria (GHT plus a cluster of ≥3 non-edge points <5% on pattern deviation), with ISGEO Category 2 as fallback and specialist review; glaucoma suspects and ocular hypertension excluded — a clear upgrade over the photo-only labels of prior work the paper critiques.
Unusually broad external validation
Sixteen datasets across eight countries spanning population-based (BES, CIEMS, BMES), nine public repositories and four hospital cohorts; the reference-standard stratification (Group A 95.2% vs B 93.8% vs C 93.1%) converts label heterogeneity from a caveat into an analysed variable.
Myopia-inclusive training design with subgroup reporting
Uncertainty mechanism supported by analysis, not just claimed
Well-designed paired reader study
Weaknesses
Headline metrics are post-filter numbers
Internal numerical inconsistencies
Closed model, commercial conflicts
External threshold policy unspecified
Coverage gaps in the hardest cases
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This multicentre development-and-validation study trains RetiGON on 56,483 colour fundus photographs (CFPs) from 16,364 Singapore Eye Research Institute participants (SEED population cohorts plus three SNEC clinical cohorts; 57.1% myopic, 14.4% high myopia; 3,635 glaucoma images, 6.4% prevalence), split 70:10:20 with participant-level separation. The pipeline is two-stage: MobileNetV2-UNet quality control/disc localisation, then an ImageNet-pretrained ViT-B/16 with cost-sensitive loss, SWAD weight averaging, and entropy-based uncertainty scoring; the <10% most uncertain predictions are referred for clinician review. Central claims: (1) standardized multimodal ground truth (VF criteria + OCT + clinical exam, ISGEO Category 2 fallback) yields internal AUROC 98.7% (95% CI 98.2–99.1); (2) generalization across 16 external datasets from eight countries (AUROC 86.4–99.6%); (3) robust high-myopia (HM) performance (internal 97.8%; external 85.5–93.3%); (4) superiority over human CFP-only reading in HM eyes (92.0% vs 70.6% mean accuracy, p=0.008) with parity to specialists given full clinical data (p=0.064).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is the RetiGON AI model for detecting glaucoma from fundus photos?
On its internal Singapore test set the model reached an AUROC of 97.8% before filtering and 98.7% (sensitivity 94.5%, specificity 97.3%) after excluding under 10% of predictions flagged as uncertain. Across 16 external datasets from eight countries, AUROCs ranged from 86.4% to 99.6%.
Does the model work in highly myopic eyes, where glaucoma is hardest to spot?
Yes, with some degradation. Internal AUROC in high myopia was 97.8%; external HM datasets scored 93.3% (Taiwan), 91.8% (Thailand), 86.5% (Beijing Eye Study) and 85.5% (Seoul). In a 50-eye study it read photos better than specialists and graders (92.0% vs 70.6% mean accuracy, p=0.008), but missed early-stage glaucoma in about a third of misclassified Seoul eyes.
Is the RetiGON code or model publicly available?
No. The algorithm has been commercially licensed (authors are co-founders of Eye.AI Pte Ltd), so code and weights are not public. The authors say controlled docker-based access may be granted for academic research on reasonable request, subject to institutional approval.
How does the AI compare with human eye doctors at diagnosing glaucoma?
Reading fundus photos alone, the model (92.0% accuracy) clearly beat eight glaucoma specialists and two trained graders (mean 70.6%, p=0.008). When specialists were given full clinical data — pressure, OCT, visual fields — their mean accuracy rose to 84.2%, statistically comparable to the model (p=0.064).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (architecture, saliency maps, reader-study plots) are described in captions but not analyzable from extracted text.
- Supplementary tables (per-dataset demographics, ablations, per-dataset sensitivity/specificity in Tables 2–3) are referenced but absent from the extraction.
- arXiv preprint — peer-review status unknown; internal inconsistencies may be pre-publication artifacts.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.