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

Source-linked AI summary

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

arXiv:2609.06699v1Sep 6, 2026cs.CVcs.AIcs.LG

TL;DR

Glaucoma detection is challenged by irreversible optic-nerve damage, imperfect feature refinement, class imbalance, and dependence on individual classifiers. The paper combines InceptionV3 features with CBAM, heterogeneous ensembles, and SMOTE+TL, then compares feature types across two public datasets. Deep features outperform handcrafted CDR features, while attention-enhanced representations achieve the strongest overall performance and Grad-CAM focuses on clinically relevant retinal regions.

Problem
Glaucoma causes irreversible optic-nerve damage, while automated detection remains challenged by feature refinement, class imbalance, and reliance on individual classifiers.
Method
The framework extracts InceptionV3 features, refines them with CBAM, combines heterogeneous classifiers through SLE and DLE, and applies SMOTE+TL for imbalance.
Results
Deep feature methods outperform handcrafted CDR-based methods, while attention-enhanced deep representations achieve the strongest overall performance across the evaluated datasets.
Takeaways & Limitations
Grad-CAM provides qualitative evidence that the proposed model focuses on clinically relevant retinal regions, supporting interpretability of its predictions.
Takeaways & Limitations
The experiments used relatively small public datasets, which may limit model generalizability.

Abstract

from arXiv · show

Glaucoma is a progressive optic neuropathy characterized by irreversible damage to the optic nerve, making timely diagnosis critical to prevent permanent vision loss. Although deep learning has demonstrated promising performance in automated glaucoma detection, existing approaches often overlook feature refinement, suffer from class imbalance, and rely on individual classifiers that limit prediction robustness. To address these challenges, this paper proposes a hybrid glaucoma detection framework that integrates attention-enhanced deep feature extraction with heterogeneous ensemble learning. Specifically, deep representations are extracted using InceptionV3 and subsequently refined by incorporating the Convolutional Block Attention Module (CBAM) to enhance discriminative retinal features. To improve classification robustness, the extracted features are classified using multiple machine learning models together with Single-Level Ensemble (SLE) and Double-Level Ensemble (DLE) strategies, while SMOTE combined with Tomek Links (SMOTE+TL) is employed to alleviate class imbalance. Furthermore, a systematic comparison of handcrafted, deep, and attention-enhanced deep feature representations is conducted. Experimental evaluation on two public retinal fundus datasets demonstrates that deep feature-based methods consistently outperform handcrafted feature-based methods, while the proposed attention-enhanced framework achieves the best overall performance. Furthermore, Grad-CAM visualizations confirm that the proposed model focuses on clinically relevant retinal regions, providing interpretable evidence on the model's prediction process.

I. INTRODUCTION

Glaucoma can cause irreversible optic-nerve damage, while conventional diagnosis is time-consuming and automated systems can reduce clinical workload. This work combines attention-enhanced features, heterogeneous ensembles, imbalance handling, and Grad-CAM interpretability for automated detection.

Glaucoma progressively damages the optic nerve and can cause irreversible blindness, making early diagnosis important.
Traditional diagnosis uses expert assessment and tests including perimetry, dilated examination, CDR measurement, pachymetry, and tonometry.
Automated systems typically extract retinal features before classifying glaucoma presence or absence using indicators such as optic-disc morphology, optic-cup morphology, vessels, and CDR.
The proposed method integrates CBAM with InceptionV3 and compares handcrafted CDR features with deep-learning features for glaucoma detection.
Heterogeneous SLE and DLE ensembles improve classification stability and performance, while SMOTE+TL addresses class imbalance where applicable.
Grad-CAM visualization highlights diagnostically relevant retinal regions to provide interpretability for model predictions.

III. DATASET DESCRIPTION

The study uses two Kaggle-sourced benchmark datasets containing glaucoma and normal fundus images. EDC has around 2,000 images after selecting two classes, whereas BEH contains 634 images and is highly imbalanced.

Two benchmark datasets sourced from Kaggle were used for the experiments.
The EDC dataset originally has four classes, but this study uses its glaucoma and normal classes, totaling around 2,000 fundus images.
The BEH dataset contains 634 eye images categorized as glaucoma or normal and is highly imbalanced.

IV. PROPOSED GLAUCOMA DETECTION METHOD

The proposed architecture processes fundus images through preprocessing, handcrafted and automated feature extraction, and classification. Classification uses traditional machine-learning models together with advanced ensemble strategies.

The workflow begins with fundus-image preprocessing before feature extraction and classification.
The method extracts both handcrafted and automated features to support glaucoma classification.
Traditional machine-learning classifiers and advanced ensemble strategies perform the final classification.
The workflow is presented as an overall pipeline from the fundus image to glaucoma detection.

A. Preprocessing

Preprocessing addresses class imbalance, image dimensions, contrast, and pixel scaling, while CDR extraction uses multiple image-processing operations. CBAM is also illustrated as refining InceptionV3-derived features.

Balancing: SMOTE+TL is applied only to the BEH training set to increase minority-class representation and improve class separability.Validation and test sets retain their original class distributions for unbiased evaluation.
Resizing: Input fundus images are resized to 299x299 pixels to match the InceptionV3 input specification.
Histogram equalization: Histogram equalization is applied to improve fundus-image contrast.
CDR Calculation: CDR calculation uses color-channel separation, enhancement, vessel processing, vessel removal, and ellipse fitting to estimate the optic-disc boundary.
Deep Learning with Attention Module: CBAM refines InceptionV3 feature maps through sequential channel and spatial attention, emphasizing informative regions and suppressing less useful ones.

B. Feature Extraction

The paper compares handcrafted CDR-based and InceptionV3 deep features, then refines convolutional representations with CBAM to emphasize informative retinal regions. The feature-extraction pipeline also includes ensemble-related figures and a final 1024-neuron representation.

Handcrafted Features: CDR features are computed from detected optic-disc and optic-cup boundaries as the ratio of their major radii.The pipeline enhances color channels, removes vessels, estimates the optic disc ellipse, detects the optic cup, and then calculates vertical CDR.
Deep Features: InceptionV3 captures retinal patterns at different spatial scales through parallel convolution filters in its inception modules.This provides the deep feature baseline for fundus-image representation.
Evaluation and Ensemble Context: The feature-extraction evaluation is reported in Table I, while the paper separately presents single- and double-level ensemble constructions.The supplied passages identify the EDC performance table and the SLE and DLE figures but do not provide their numerical results.
Attention-Enhanced Features: CBAM sequentially applies channel and spatial attention to InceptionV3 feature maps, emphasizing informative regions and suppressing less useful ones.The paper describes the intermediate map F, channel-refined map F′, and final attention-refined map F′′; element-wise multiplication combines maps with attention weights.
Attention-Enhanced Features: CBAM blocks are inserted after mixed0 through mixed9, followed by global average pooling and a 1024-neuron ReLU fully connected layer.These components produce the final feature representation.

1) Traditional ML Models:

The classification stage evaluates heterogeneous conventional machine-learning models and combines them through single-level and double-level voting ensembles. These ensembles are intended to reduce dependence on one decision function and improve prediction stability.

Conventional Classifiers: The evaluated classifiers include MLP, SVM, KNN, LR, RF, AdaBoost, XGBoost, and LightGBM.All listed models are applied to the extracted feature representations.
Ensemble Motivation: Heterogeneous classifier combinations are used because different models can learn different decision boundaries from the same feature representation.The paper motivates ensemble learning as a way to reduce dependence on a single decision function.
Single-Level Ensemble: SLE combines all machine-learning methods in one voting classifier.The construction is illustrated in Figure 5.
Double-Level Ensemble: DLE first forms two voting classifiers and then ensembles them into a third classifier at a second level.VC1 contains MLP, SVM, KNN, and LR, whereas VC2 contains RF, AdaBoost, XGBoost, and LightGBM.

D. Visual Explanation using Grad-CAM

Grad-CAM is used to interpret model predictions by identifying influential fundus-image regions through gradients from the final convolutional layer. The experimental setup uses held-out validation and test divisions and reports several classification metrics plus inference time.

Visual Explanation: Grad-CAM weights convolutional feature maps with final-layer gradients to generate a heatmap of influential retinal regions.This provides a visual explanation of which fundus locations contribute most strongly to the model prediction.
Experimental Setup: Each dataset is split into 80% training, 10% validation, and 10% testing subsets.The reported evaluation includes macro precision, recall, F1-Score, accuracy, and inference time.
Evaluation Metrics: Table II reports overall accuracy and class-wise precision, recall, and F1-Score for classifiers using InceptionV3+CBAM features on the BEH dataset.The caption specifies that accuracy is overall, while the other three metrics are reported per class.
Visual Explanation: The Grad-CAM figure contrasts focus regions for a misclassified InceptionV3 extractor with a correctly classified InceptionV3+CBAM extractor.The comparison is designed to examine how attention-enhanced features change the model’s visual focus.

B. Result Analysis on EDC Dataset

On the EDC dataset, attention-enhanced deep features and ensemble learning outperform conventional CDR-based representations, while Grad-CAM indicates more clinically focused feature extraction. Classifier choice also involves a predictive-performance versus inference-time trade-off.

Feature and ensemble performance: InceptionV3+CBAM features deliver the strongest EDC results, with SLE and DLE reaching 0.90 accuracy and improved precision, recall, and F1-Score.InceptionV3 alone yields accuracy rates of 0.86–0.88 and F1-Scores of 0.87–0.89, while attention-based features improve the ensemble results.
Feature and ensemble performance: Attention-based deep features with ensemble learning produce better glaucoma identification than standard CDR assessment methods.CDR feature-trained models show low accuracy and weak precision-recall performance, whereas deep representations improve classifier performance.
Interpretability: Grad-CAM shows InceptionV3+CBAM concentrating on the optic disc and surrounding cup region, unlike InceptionV3’s diffuse background and peripheral activation.The localized activation provides qualitative evidence that CBAM emphasizes diagnostically meaningful retinal structures and suppresses irrelevant background information.
Efficiency trade-offs: XGBoost achieves the highest F1 with relatively low inference time, while AdaBoost incurs the highest inference time for only marginal F1 improvement.KNN and LR are fastest but produce the lowest accuracy, illustrating the trade-off between predictive performance and computational efficiency.

C. Result Analysis on BEH Dataset

On the imbalanced BEH dataset, SMOTE+TL improves recognition of the minority glaucoma class while preserving relatively stable normal-class performance. The paper concludes that attention-enhanced representations and heterogeneous ensembles improve robustness and sensitivity, although small public datasets constrain generalizability.

Class imbalance: SMOTE+TL improves glaucoma recognition on BEH by increasing minority-class representation, while normal-class performance remains relatively stable.The reported improvement primarily benefits glaucoma sensitivity under class imbalance.
Framework conclusion: The proposed framework combines CBAM-enhanced InceptionV3 features, heterogeneous SLE and DLE ensembles, and SMOTE+TL to improve classification robustness and glaucoma sensitivity.The conclusion identifies attention-enhanced representations, ensemble strategies, and class balancing as complementary components of the framework.
Limitations: Experiments on relatively small public datasets may limit the model’s generalizability, motivating validation on larger datasets and improved computational efficiency.This scope boundary is stated as a limitation and future-work direction.