Attention-Enhanced Deep Features with Heterogeneous Ensemble Learning for Glaucoma Detection
Abdullah Al Shafi, Nishat Sadaf Lira, Abrar Hasan, Kazi Saeed Alam, Swapnil Kundu Argha
cs.CV · cs.AI · cs.LG
Quick verdict
Claims CBAM+InceptionV3 features with ensemble learning reach 0.90 accuracy on EDC; pipeline sound but gains tiny, untested; fine-tuned CNN baseline missing.
Summary
The paper proposes a glaucoma detection pipeline that extracts features from retinal fundus images with InceptionV3 augmented by CBAM attention modules inserted after each mixed layer (mixed0–mixed9), followed by GAP and a 1024-neuron ReLU layer. These features are classified by eight ML models (MLP, SVM, KNN, LR, RF, AdaBoost, XGBoost, LightGBM) plus two voting ensembles: Single-Level Ensemble (SLE) and a two-level Double-Level Ensemble (DLE). SMOTE+Tomek Links balances the training split only. On the EDC dataset (~2,000 images), handcrafted CDR features perform poorly (accuracy 0.54–0.63), InceptionV3 features reach 0.86–0.89, and InceptionV3+CBAM features peak at 0.90 accuracy with SLE and DLE. On the imbalanced BEH dataset (634 images), SMOTE+TL raises glaucoma recall sharply (e.g., MLP 0.39→0.89, SVM 0.13→0.87) while test sets keep their original distribution. Grad-CAM on one example pair suggests CBAM concentrates attention on the optic disc. Main limitations: the 0.01–0.02 margins carry no statistical tests, confidence intervals, or repeated runs; no end-to-end fine-tuned CNN baseline is compared; Grad-CAM evidence is a single example; no code or hyperparameters are released; and validation covers only two small Kaggle datasets. The contribution is a competent combinatorial integration of standard components rather than a methodological advance.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 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
- Inserting CBAM into InceptionV3 yields features that outperform both handcrafted CDR features and plain InceptionV3 features across eight classifiers and two ensembles (Table I: ensemble accuracy 0.90 vs 0.89 for plain InceptionV3 on EDC).
- SLE and DLE voting ensembles give the best overall EDC results with CBAM features: 0.90 accuracy, 0.91 precision, 0.90 recall, 0.90 F1.
- SMOTE+Tomek Links applied only to training data substantially improves glaucoma sensitivity on imbalanced BEH, e.g., glaucoma recall rises from 0.39 to 0.89 (MLP) and 0.13 to 0.87 (SVM) (Table II).
- Grad-CAM shows the CBAM model concentrates activation on the optic disc/cup region, offering qualitative evidence of clinically plausible attention.
Strengths
Systematic feature-representation ablation
Table I compares CDR, InceptionV3, and InceptionV3+CBAM features across eight classifiers plus SLE/DLE on EDC, isolating each representation's contribution under an identical protocol.
Leakage-aware imbalance handling
SMOTE+TL is applied to training splits only while validation/test keep original distributions (Sec. IV-A); Table II reports class-wise metrics instead of hiding majority-class bias behind overall accuracy.
Quantified minority-class failure and fix
Efficiency and interpretability angles
Weaknesses
No statistical support for headline margins
Missing end-to-end fine-tuning baseline
Suspiciously weak CDR baseline
Single-example interpretability evidence
Reproducibility gaps and an unexplained anomaly
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv preprint (2609.06699, cs.CV, Sep 2026; KUET / Daffodil International University / Green University, Bangladesh) proposes a binary glaucoma detection pipeline for retinal fundus images. Four claims structure the paper: (1) CBAM-enhanced InceptionV3 features outperform both handcrafted CDR features and plain InceptionV3 features; (2) heterogeneous voting ensembles (SLE, DLE) improve robustness; (3) SMOTE+Tomek Links alleviates class imbalance; (4) Grad-CAM confirms attention on clinically relevant regions. Evaluation covers two Kaggle datasets: EDC (~2,000 glaucoma/normal images) and BEH (634 images, imbalanced), split 80/10/10.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 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
What accuracy does this glaucoma detection method achieve?
On the EDC dataset (~2,000 fundus images), InceptionV3+CBAM features classified with the SLE or DLE voting ensembles reach 0.90 accuracy and about 0.90 F1. Plain InceptionV3 features reach 0.86–0.89 with the same classifiers, and handcrafted cup-to-disc-ratio features only 0.54–0.63.
Which datasets were used in the paper?
Two public Kaggle datasets: the Eye-Disease-Classification set (EDC, glaucoma vs normal, about 2,000 images) and the BEH glaucoma set (634 images, with far fewer glaucoma than normal images).
Is code available for this paper?
No. The paper links the two Kaggle datasets but does not announce any code release, and it omits classifier hyperparameters and split seeds, so exact reproduction requires full reimplementation.
How does the paper handle class imbalance?
SMOTE combined with Tomek Links is applied only to the training split; validation and test sets keep their original distributions. On the imbalanced BEH dataset this raises glaucoma recall from as low as 0.13 (SVM, imbalanced) to roughly 0.85–0.91 across models, though KNN's overall accuracy actually drops.
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.
- Figure image content (Grad-CAM heatmaps, workflow diagrams, Fig. 8 scatter) cannot be assessed from the text extraction; only captions, axis labels and tick values are present, so visual claims are judged from the authors' descriptions.
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.