Interpretable AI plus Handheld, Portable Retinal Photographs: A Low-Cost Glaucoma Screening Solution for West Africa
Charis Y. N. Chiang, Tarela Sarimiye, Adeyinka Ashaye, Martin Buist, Michael A. Hauser, Olusola Olawoye, Michaël J. A. Girard
cs.CV · cs.AI
Quick verdict
Interpretable AI screens glaucoma from handheld fundus photos in Nigeria (Volk AUC 0.85 vs Canon 0.93); promising, but ground truth included tabletop images.
Summary
The study develops and evaluates an interpretable deep-learning pipeline that screens for glaucoma from single fundus photographs taken with a low-cost handheld camera (Volk Viva) in a Nigerian community cohort, benchmarking it against the clinical tabletop Canon CR-2-AF on the same eyes. From 681 participants, the paired gradable cohort comprised 1,091 eyes (287 glaucoma, 410 suspects, 394 non-glaucoma). Device-specific component models perform vessel segmentation, cup/disc segmentation, and detection of twelve optic nerve head features; a late-fusion classifier integrates images, masks, features, cup-to-disc ratios, and age into three-class predictions, with Grad-CAM, SHAP, and ablation analyses and a per-image clinical report. Dice coefficients reached 0.98/0.95 (Volk) and 0.94/0.96 (Canon) for vessels and cup/disc; classification macro-AUC was 0.85±0.01 (Volk) versus 0.93±0.01 (Canon). Ablations showed masks mattered most for Canon, while image and tabular inputs dominated Volk. Limitations include ground-truth circularity (Canon images contributed to the reference diagnosis, favoring that device), a single cohort without external validation, very small segmentation training sets (60 images/device), absent or rare features with zero SHAP weights, and no code release or inferential comparison between devices.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
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
- An interpretable modular pipeline can classify eyes as non-glaucoma, glaucoma suspect, or glaucoma from a single portable fundus photograph, achieving macro-AUC 0.85±0.01 (Volk Viva) versus 0.93±0.01 (Canon CR-2-AF) on the same paired 1,091-eye cohort.
- Device-specific component models reach high segmentation agreement (vessel Dice 0.98 Volk / 0.94 Canon; cup-disc Dice 0.95 / 0.96) and moderate ONH feature-detection AUC (0.83±0.03 / 0.87±0.04).
- Modality reliance differs by device: ablating masks cuts Canon macro-AUC by 19.5% but Volk by only 0.3%, while the RGB image (−7.2%) and tabular features (−8.5%) matter most for Volk.
- Low-cost handheld imaging combined with interpretable AI can plausibly enable community-level glaucoma screening in resource-limited West African settings, pending external validation.
Strengths
Paired same-eye device comparison on a real community cohort
1,091 eyes from 589 participants were imaged on both Volk Viva and Canon CR-2-AF, with identical subject-level partitions shared across devices (Methods, Feature Detector section) — a genuinely fair design for device comparison, free of subject-level leakage.
Interpretability is structural, not cosmetic
Intermediate outputs (vessel masks, cup/disc masks, twelve ONH features, vCDR/hCDR) are independently verifiable clinical quantities; the Figure 5 report combines Grad-CAM focus maps, class confidences, and SHAP contributions normalized to 100%, letting a non-specialist check the model rather than trust it.
Honest, device-divergent ablation finding
Transparent bookkeeping of data losses
Deployment-relevant workflow detail
Weaknesses
Ground-truth circularity confounds the headline comparison
No inferential statistics anywhere
Missing baselines for the modular design
Small and curated segmentation training/evaluation sets
The twelve-feature capability overstates what the data support
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper presents a modular, interpretable deep-learning pipeline for three-class glaucoma screening (non-glaucoma / glaucoma suspect / glaucoma) from single fundus photographs, developed and evaluated on a community-based cohort of 681 participants (1,362 eyes) recruited through outreaches in Oyo State, Nigeria. Two central claims: (1) a low-cost handheld camera (Volk Viva) supports glaucoma screening AI with performance reasonably comparable to the clinical tabletop Canon CR-2-AF — classification macro-AUC 0.85±0.01 vs 0.93±0.01 on the same paired cohort of 1,091 eyes (Table 2); (2) the pipeline's intermediate outputs (vessel and cup/disc segmentations, twelve optic-nerve-head features, CDR measurements, Grad-CAM, SHAP-based report) make decisions auditable by non-specialist clinicians. Both are plausible and clinically consequential, but the first is weakened by a ground-truth circularity the authors themselves acknowledge, and the second is supported qualitatively (Figure 5 sample report) rather than by any clinician evaluation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 AI glaucoma screening with a handheld fundus camera?
In this Nigerian community study of 1,091 eyes, the three-class classifier (glaucoma / glaucoma suspect / non-glaucoma) built on handheld Volk Viva photos reached a macro AUC of 0.85±0.01, versus 0.93±0.01 for the clinic tabletop Canon CR-2-AF on the same eyes. Note the comparison partly favors the tabletop camera because its photos also contributed to the ground-truth diagnoses.
Why did the portable camera score lower than the tabletop camera in this study?
Two reasons given: portable images were blurrier with fewer visible fine structures, and — more importantly — the Canon tabletop photographs were part of the clinical work-up used to establish each eye's diagnosis, giving that device a methodological head start that the authors themselves acknowledge.
What does the AI report show for each eye scan?
Each scan produces a one-page report containing image gradability, segmented retinal vessels and optic cup/disc boundaries, estimated cup-to-disc ratios, the twelve detected optic-nerve-head features, the predicted class with confidence scores, a Grad-CAM map showing where the model looked, and SHAP-based percentages of how much each feature contributed to the decision.
Is the code or dataset from this glaucoma AI study publicly available?
No. The paper contains no code or data availability statement, and the cohort is private clinical data from Nigerian community outreaches. The authors state that external validation in independent West African cohorts is needed before deployment.
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 (ROC curves, confusion matrices, sample report, Grad-CAM overlays) are captioned but their content cannot be analyzed from the extracted text.
- Accuracy, class-balanced accuracy, and per-class precision/recall are promised in Methods but absent from the extracted text; judgments about per-class performance (notably the 410-eye suspect class) are therefore not possible.
- The source of AUC standard deviations (folds vs bootstrap) is inferred from the Figure 4 caption ('mean of confusion matrix across the five folds'), not explicitly stated for all metrics.
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.