Fundus Image-based Glaucoma Screening via Retinal Knowledge-Oriented Dynamic Multi-Level Feature Integration

Chi Liu; Yuzhuo Zhou; Sheng Shen; Zongyuan Ge; Fengshi Jing; Shiran Zhang; Yu Jiang; Anli Wang

cs.CVApr 14, 2026

Fundus Image-based Glaucoma Screening via Retinal Knowledge-Oriented Dynamic Multi-Level Feature Integration

Authors: Chi Liu, Yuzhuo Zhou, Sheng Shen, Zongyuan Ge, Fengshi Jing, Shiran Zhang, Yu Jiang, Anli Wang,

Organizations: Faculty of Data Science, City University of Macau, Avenida Padre Tomás Pereira, Taipa, 999078, Macao, China · State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangdong Provincial Key Laboratory of Ophthalmology and Vision Science, Guangdong Provincial Clinical Research Center for Ocular Diseases, Guangzhou, 510000, Guangdong, China · Hainan International College, Minzu University of China, Li’an International Education Innovation Pilot Zone, Lingshui, 572423, Hainan, China · School of Information Technology, Faculty of Business and Hospitality, Torrens University Australia, 196 Flinders St, 3000, VIC, Australia · AIM for Health Lab, Faculty of IT, Monash University, Australia, 25 Exhibition Walk, Clayton, 3800, VIC, Australia · Department of Ophthalmology, Guangzhou First People’s Hospital, the Second Affiliated Hospital of South China University of Technology, No. 1, Panfu Road, Yuexiu District, Guangzhou, 510000, Guangdong, China · The Third Xiangya Hospital of Central South University, 138 Tongzipo Road, Changsha, 410013, Hunan, China

Abstract

While deep learning has advanced automated glaucoma screening via color fundus photography, existing purely data-driven models often overfit to confounding imaging artifacts and struggle to capture unpredictable pathological cues located beyond predefined anatomical boundaries. To address these limitations, we propose a retinal knowledge-oriented framework that synergizes dynamic multi-scale feature learning with domain-specific anatomical priors. The proposed architecture adopts a tri-branch structure to jointly model the global retinal context, the structural characteristics of the optic cup and disc, and dynamically cropped pathological regions. Specifically, we devise a Dynamic Window Mechanism (DWM) that adaptively discovers diagnostically informative patches via image-level supervision. Furthermore, we introduce a Knowledge-Enhanced Convolutional Block Attention Module (KE-CBAM) that explicitly incorporates retinal priors from a pre-trained foundation model RETFound to guide spatial attention, preventing the network from assigning spurious weights to irrelevant background noise. Extensive evaluations on the large-scale AIROGS dataset demonstrate that our method achieves a state-of-the-art AUC of $98.5\%$ and an accuracy of $94.6\%$ . More importantly, the integration of anatomical priors effectively mitigates the inherent class imbalance, significantly improving the detection of referable glaucoma cases. Additional validations on the SMDG-19 benchmark further confirm its superior cross-domain generalization, indicating that our contribution provides a robust, interpretable, and scalable solution for real-world clinical glaucoma diagnosis. Our code is available at https://github.com/magiczhuo/Glaucoma-Detection

Figures & tables

Figure 1: Conceptual illustration of the proposed retinal knowledge-oriented glaucoma screening framework. Existing deep learning models rely primarily on purely data-driven attention mechanisms, which suffer from limited anatomical awareness, structural uncertainty in optic disc–cup regions, and poor generalization across heterogeneous imaging devices. To address these challenges, the proposed framework integrates retinal priors extracted from the RETFound foundation model to guide attention learning. Knowledge-guided attention enables the model to jointly capture global retinal context, structural features of the optic disc–cup complex, and dynamically localized pathological regions. By combining retinal knowledge injection with multi-scale feature integration, the proposed framework achieves robust and interpretable glaucoma screening across diverse clinical datasets.

Category	Reference / Method	Input Focus	Domain Prior	Attention Type	Key Limitations
Clinical & Traditional	Caiado et al. [ 5 ] (Laguna ONhE)	Local (ONH)	Colorimetric	None	Sensitive to imaging variability; subjective interpretation.
Clinical & Traditional	Joshi et al. [ 19 ] , Cheng et al. [ 6 ]	Local (OC/OD)	Handcrafted	None	Fixed handcrafted features; limited generalization.
Deep Learning (Local-based)	Li et al. [ 23 ] (Region-based CNN)	Local (OC/OD)	Spatial	Disc-based	Loses global context; OD/OC variability sensitive.
Deep Learning (Local-based)	Xu et al. [ 37 ]	Local (OC/OD)	Structural	None	Fixed ROIs; misses peripheral cues.
DL (Global-based)	Huang et al. [ 15 ] (CNN-ViT & Deformable)	Global	None	Generic (Multi-Head)	Prone to artifact-driven attention.
Foundation Models	Jalili et al. [ 17 ] (GPT-4V)	Local	Prompt-based Clinical	Generic (Transformer)	Limited global context; inconsistent outputs.

Table 1: Comparison of representative existing methods and the proposed framework, highlighting the research gap in glaucoma screening.

Figure 2: Overview of the proposed retinal knowledge-oriented glaucoma screening framework. The architecture adopts a tri-branch structure consisting of a global context branch, an ROI structural branch, and a dynamic window localization branch. Each branch employs a ResNet152 backbone enhanced with the proposed KE-CBAM, which integrates retinal anatomical priors extracted from the RETFound foundation model. The multi-scale feature representations are fused to produce the final glaucoma screening prediction.

Figure 3: Illustration of the DWM. High-response regions are identified from the global feature map using response score estimation. The top-scoring patches are cropped and resized to form adaptive local inputs, enabling the model to capture subtle pathological cues beyond predefined anatomical regions.

Figure 4: Workflow of the proposed KE-CBAM. Retinal anatomical priors extracted from the RETFound foundation model are projected into the backbone feature space and fused with CBAM-generated channel–spatial features through a cross-modal attention mechanism. This process enables the network to incorporate domain-specific retinal knowledge into the feature weighting process, improving attention localization toward clinically relevant structures.

Figure 5: ROI 800 represents the enlarged region of interest around the optic disc. ROI 800 CLAHE denotes the contrast-enhanced input image using Contrast Limited Adaptive Histogram Equalization (CLAHE). Red boxes indicate the ROI region, yellow boxes illustrate the enlarged view, and blue boxes highlight the contrast enhancement effect.

Model (3cls)	Evaluation Metrics
Model (3cls)	AP	AUC	Acc	F1	Sen	Spe
Patch5Model (No AM)	0.778 $\pm$ 0.014 [0.761, 0.796]	0.958 $\pm$ 0.006 [0.951, 0.966]	0.924 $\pm$ 0.008 [0.914, 0.934]	0.726 $\pm$ 0.027 [0.692, 0.760]	0.704 $\pm$ 0.034 [0.662, 0.747]	0.927 $\pm$ 0.013 [0.911, 0.944]
Branch3CBAM	0.790 $\pm$ 0.014 [0.773, 0.807]	0.963 $\pm$ 0.004 [0.958, 0.968]	0.928 $\pm$ 0.004 [0.923, 0.932]	0.770 $\pm$ 0.017 [0.749, 0.792]	0.767 $\pm$ 0.024 [0.738, 0.796]	0.944 $\pm$ 0.006 [0.937, 0.951]
Branch3KECBAM	0.806 $\pm$ 0.003 [0.803, 0.809]	0.970 $\pm$ 0.001 [0.969, 0.971]	0.934 $\pm$ 0.001 [0.933, 0.935]	0.783 $\pm$ 0.006 [0.776, 0.790]	0.782 $\pm$ 0.008 [0.772, 0.792]	0.949 $\pm$ 0.002 [0.947, 0.951]

Table 2: Performance evaluation for tri-class and binary classification settings, with 5-fold means, standard deviations, and 95% confidence intervals (0.95CI). The tri-class setting includes Negative , Positive , and Suspect . The binary setting defines Non-referable as Negative , Referable as ( Positive and Suspect )

Panel of Figure 6 — Figure 6: Row-normalized confusion matrices of the proposed framework under (a) tri-class (Negative/Suspect/Positive) and (b) binary (non-referable vs referable) settings.

Panel of Figure 7 — Figure 7: Representative examples of optic disc ROI extraction under diverse real-world imaging conditions. Top row: (a) standard high-quality image, (b) severe out-of-focus with blurred boundaries, (c) low-resolution with local artifacts, (d) severely underexposed with dim illumination, (e) low-contrast with weak vessel patterns. Bottom row (f-j): corresponding ROI crops demonstrating consistent optic disc localization across all conditions.

Part I: Performance comparison on EyePACS Dataset
Author	Method	Acc (%)	Sen (%)	Spe (%)	AUC (%)
Li and Liu (2022) [ 24 ]	U-Nets-DenseNet	83.40	82.30	84.60	91.50
Zhao et al. (2023) [ 44 ]	ResNet50-MaxViT	91.60	92.50	90.70	97.00
Alam et al. (2023) [ 1 ]	SA-GoogleNet	90.00	90.00	90.10	96.10
Joshi et al. (2024) [ 20 ]	VisionDeep-AI	91.40	93.00	89.90	97.00
Almeida et al. (2024) [ 2 ]	AVS-DenseNet	77.80	85.90	69.80	85.00

Table 3: Comprehensive performance evaluation: Part I compares our proposed method with recent state-of-the-art methods on the EyePACS dataset. Part II evaluates the generalization capability across seven public datasets from the SMDG-19 [ 21 ] in baseline comparisons [ 30 ] .

Model	Params (M)	FLOPs (G)	Time (ms)	FPS	F1 (%)
3branch (Baseline)	60.75	171.73	$20.93\pm 3.50$	$49.13\pm 12.70$	$84.2\pm 3.9$
3branch-cbam	67.34	172.06	$38.51\pm 0.18$	$25.97\pm 0.15$	$89.03\pm 0.20$
3branch-rcbam	83.59	200.66	$62.56\pm 0.23$	$15.99\pm 0.07$	$89.70\pm 0.36$
Proposed vs. baseline	+37.7%	+16.8%	+198.8%	-67.4%	+5.6%

Table 4: Model complexity and computational efficiency comparison across three architectural variants. Values are reported as mean

\pm

standard deviation, with 95% confidence intervals shown in brackets for time and FPS. FLOPs are measured with input size

299\times 299\times 3

Degradation	Model	Level	Acc.	Sens.	F1	AUC
Blur	baseline	0	$0.9373\pm 0.0057$	$0.8176\pm 0.0734$	$0.8633\pm 0.0194$	$0.9804\pm 0.0045$
	Proposed	0	$0.9421\pm 0.0016$	$0.9254\pm 0.0061$	$0.8864\pm 0.0023$	$0.9833\pm 0.0002$
	baseline	1	$0.9361\pm 0.0067$	$0.8148\pm 0.0729$	$0.8607\pm 0.0213$	$0.9784\pm 0.0050$
	Proposed	1	$0.9413\pm 0.0014$	$0.9201\pm 0.0074$	$0.8844\pm 0.0020$	$0.9821\pm 0.0001$
	baseline	2	$0.9319\pm 0.0072$	$0.7939\pm 0.0757$	$0.8495\pm 0.0233$	$0.9784\pm 0.0050$
	Proposed	2	$0.9387\pm 0.0013$	$0.9107\pm 0.0077$	$0.8788\pm 0.0019$	$0.9804\pm 0.0002$

Table 5: Performance comparison between the Patch5Model and proposed Branch3KECBAM under three degradation levels (lvl = 0, 1, 2) for blur, JPEG compression, and Gaussian noise. Values are reported as mean

\pm

standard deviation across 5 folds.

Figure 8: Examples of misclassified referable and non‑referable cases. Yellow bounding boxes indicate regions that may confuse the model due to imaging artifacts or structural ambiguity.

Panel of Figure 9 — Figure 9: Top-6 windows selected by DWM on representative fundus images. Large-scale patches ( L1-L3 : $224\times 224$ ) and fine-grained local patches ( S1-S3 : $112\times 112$ ) consistently focus on the optic disc, neuroretinal rim, and major vascular regions.

Panel of Figure 10 — Figure 10: Saliency maps generated via Grad-CAM++ for representative glaucoma cases: Each row displays both high-quality (a, c, e, g) and low-quality (b, d, f, h) samples of each case. Columns within each group represent: Full Original, KE-CBAM CAM Map and Baseline CAM Map. The consistent focus on the optic disc and cup margins across diverse cases validates the robustness of our knowledge-enhanced attention mechanism.

Panel of Figure 11 — Figure 11: T-SNE visualization of feature embeddings learned by different model variants. The proposed KE-CBAM produces clearer inter-class separation and more compact clusters.

Glaucoma Detection

See all 18 papers

Attention-Enhanced Deep Features with Heterogeneous Ensemble Learning for Glaucoma Detection
Sep 6, 2026Abdullah Al Shafi, Nishat Sadaf Lira, Abrar Hasan +2Glaucoma Detection Ensemble Learning
Department of CSE, Khulna University of Engineering & Technology, Bangladesh · Department of CSE, Daffodil International University, Bangladesh · Department of SWE, Green University of Bangladesh
GlaKG: A Biomarker-Centric Fundus Knowledge Graph for Explainable Glaucoma Diagnosis and Risk Assessment
Jul 6, 2026Cheng Huang, Jia Zhang, Yi Jiang +7Medical Image Analysis Medical Diagnosis
University of Texas Southwestern Medical Center · 2Southern Methodist University · 3Nanyang Technological University +1
An Automated Multimodal Glaucoma Detection Framework Using ViT and a Stacking-Based Ensemble
Jul 2, 2026Ishrat Jahan, Muhammad E. H Chowdhury, Murugappan Murugappan +7Stacked Ensemble Learning Multimodal Classification
Department of Computer Science and Engineering, Shahjalal University of Science and Technology, Sylhet, Bangladesh · Department of Electrical Engineering, Qatar University, Doha, Qatar · Department of Electronics and Communication Engineering, Kuwait College of Science and Technology, Block 4, Doha, Kuwait +5

Sep 6, 2026J/K move · Enter open · S save

Attention-Enhanced Deep Features with Heterogeneous Ensemble Learning for Glaucoma Detection

Abdullah Al Shafi, Nishat Sadaf Lira, Abrar Hasan, Kazi Saeed Alam+1

Department of CSE, Khulna University of Engineering & Technology, Bangladesh · Department of CSE, Daffodil International University, Bangladesh · Department of SWE, Green University of Bangladesh

Glaucoma Detection Ensemble Learning Deep Ensembles

Open paper AskPDF arXiv

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.

Abstract

Glaucoma Detection

Explore similar work

Attention-Enhanced Deep Features with Heterogeneous Ensemble Learning for Glaucoma Detection

GlaKG: A Biomarker-Centric Fundus Knowledge Graph for Explainable Glaucoma Diagnosis and Risk Assessment

An Automated Multimodal Glaucoma Detection Framework Using ViT and a Stacking-Based Ensemble