pith. machine review for the scientific record. sign in

arxiv: 2604.12351 · v1 · submitted 2026-04-14 · 💻 cs.CV

Recognition: unknown

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

Yuzhuo Zhou , Chi Liu , Sheng Shen , Zongyuan Ge , Fengshi Jing , Shiran Zhang , Yu Jiang , Anli Wang
show 4 more authors
Wenjian Liu Feilong Yang Tianqing Zhu Xiaotong Han
Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords glaucoma screeningfundus imagesretinal priorsdynamic feature integrationattention mechanismcross-domain generalizationoptic disc and cupdeep learning
0
0 comments X

The pith

A tri-branch framework fuses retinal anatomical priors with dynamic lesion localization to raise glaucoma screening robustness in fundus images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing data-driven deep learning models for glaucoma detection from color fundus photographs lack explicit retinal anatomical knowledge and rely on fixed regions, which reduces their reliability when imaging conditions or patient populations change. The paper sets out to show that embedding priors from a pre-trained foundation model into a dynamic multi-level system can overcome these limits. The approach uses three complementary branches for global context, optic disc and cup structure, and adaptively selected pathological areas, guided by a knowledge-enhanced attention module and a dynamic window that identifies useful regions on the fly. A sympathetic reader cares because large-scale automated screening requires models that maintain high performance without retraining on every new clinical dataset. If the integration works, screening systems can become more accurate and generalizable by letting known retinal structure steer feature learning instead of relying solely on image statistics.

Core claim

The central claim is that a retinal knowledge-oriented framework that combines dynamic multi-scale feature learning with domain-specific retinal priors in a tri-branch architecture, using a Dynamic Window Mechanism to locate informative regions and a Knowledge-Enhanced Convolutional Attention Module to inject retinal priors, produces more robust glaucoma classification than purely data-driven baselines.

What carries the argument

The Knowledge-Enhanced Convolutional Attention Module that injects retinal priors into attention learning together with the Dynamic Window Mechanism that adaptively selects diagnostically relevant regions.

If this is right

  • The tri-branch design captures global retinal context, optic disc and cup structure, and localized pathological cues at the same time.
  • Adaptive region selection via the Dynamic Window Mechanism allows pathological signals outside fixed anatomical zones to contribute to the decision.
  • Retinal priors steer attention learning so the model focuses on clinically meaningful patterns rather than dataset-specific noise.
  • Performance reaches an AUC of 98.5 percent and accuracy of 94.6 percent on the large AIROGS collection while maintaining strong results on multiple external benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prior-guided dynamic localization could be tested on other retinal diseases where lesions appear in variable locations.
  • If the priors transfer reliably, the method reduces the amount of labeled data needed for new screening tasks.
  • The approach implies that medical imaging models may benefit more from explicit anatomical constraints than from ever-larger purely statistical training sets.

Load-bearing premise

Retinal priors taken from a pre-trained foundation model will remain accurate and helpful when the model encounters new fundus datasets whose imaging conditions and patient demographics differ from the pre-training data.

What would settle it

On a held-out fundus dataset with different cameras and demographics, the full model shows no gain in AUC or accuracy over a plain convolutional network trained from scratch, or expert review finds that the attention maps highlight areas unrelated to actual glaucomatous damage.

Figures

Figures reproduced from arXiv: 2604.12351 by Anli Wang, Chi Liu, Feilong Yang, Fengshi Jing, Sheng Shen, Shiran Zhang, Tianqing Zhu, Wenjian Liu, Xiaotong Han, Yu Jiang, Yuzhuo Zhou, Zongyuan Ge.

Figure 1
Figure 1. 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 … view at source ↗
Figure 2
Figure 2. 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 Knowledge-Enhanced Convolutional Block Attention Module (KE-CBAM), which integrates retinal anatomical pr… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Dynamic Window Mechanism (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 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Workflow of the proposed Knowledge-Enhanced Convolutional Block Attention Module (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 wei… view at source ↗
Figure 5
Figure 5. 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 His￾togram Equalization (CLAHE). Red boxes indicate the ROI region, yellow boxes illustrate the enlarged view, and blue boxes highlight the contrast enhancement effect. samples, where the latter two categories collectively repre￾sent referable gla… view at source ↗
Figure 6
Figure 6. Figure 6: The Tri-class ROC curve of the proposed framework, demonstrating high discriminative capability across Referable (Glaucoma, Suspect) and Non-Referable (Normal). results under a tri-class setting consisting of Negative, Suspect, and Positive glaucoma. The performance of the proposed Branch3KECBAM is compared with the baseline Branch3CBAM in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Row-normalized confusion matrices of the proposed framework under (a) tri-class (Negative/Suspect/Positive) and (b) binary (non-referable vs referable) settings [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. 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. Non-referable Cases (Negative Samples) Orig. KE-CBAM Baseline (a) Neg-Sample 1 Orig. KE-CBAM Baseline (b) Neg-Sample 2 (c) Neg-Sample 3 (d) Neg-Sample 4 Referable Cases (Positive & Suspect Samples) Orig. KE-CBAM Baseline (e) Pos-Sam… view at source ↗
Figure 9
Figure 9. Figure 9: 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 mechanis… view at source ↗
Figure 10
Figure 10. Figure 10: T-SNE visualization of feature embeddings learned by different model variants. The proposed KE-CBAM produces clearer inter-class separation and more compact clusters. concentrates its attention around the OC and OD. Notably, this performance persists under varying image quality con￾ditions. For the low-quality samples in all categories (b, d, f, h), where suffer from low resolution, insufficient explo￾rat… view at source ↗
read the original abstract

Automated diagnosis based on color fundus photography is essential for large-scale glaucoma screening. However, existing deep learning models are typically data-driven and lack explicit integration of retinal anatomical knowledge, which limits their robustness across heterogeneous clinical datasets. Moreover, pathological cues in fundus images may appear beyond predefined anatomical regions, making fixed-region feature extraction insufficient for reliable diagnosis. To address these challenges, we propose a retinal knowledge-oriented glaucoma screening framework that integrates dynamic multi-scale feature learning with domain-specific retinal priors. The framework adopts a tri-branch structure to capture complementary retinal representations, including global retinal context, structural features of the optic disc/cup, and dynamically localized pathological regions. A Dynamic Window Mechanism is devised to adaptively identify diagnostically informative regions, while a Knowledge-Enhanced Convolutional Attention Module incorporates retinal priors extracted from a pre-trained foundation model to guide attention learning. Extensive experiments on the large-scale AIROGS dataset demonstrate that the proposed method outperforms diverse baselines, achieving an AUC of 98.5% and an accuracy of 94.6%. Additional evaluations on multiple datasets from the SMDG-19 benchmark further confirm its strong cross-domain generalization capability, indicating that knowledge-guided attention combined with adaptive lesion localization can significantly improve the robustness of automated glaucoma screening systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a retinal knowledge-oriented glaucoma screening framework for fundus images. It uses a tri-branch architecture to capture global retinal context, optic disc/cup structural features, and dynamically localized pathological regions. A Dynamic Window Mechanism adaptively identifies informative regions, and a Knowledge-Enhanced Convolutional Attention Module injects retinal priors from a pre-trained foundation model to guide attention. Experiments report an AUC of 98.5% and accuracy of 94.6% on the large-scale AIROGS dataset, with additional evaluations on SMDG-19 benchmark datasets claimed to demonstrate cross-domain generalization.

Significance. If the performance numbers are reproducible and the knowledge integration demonstrably improves robustness, the work could contribute to more reliable automated glaucoma screening in heterogeneous clinical settings. The tri-branch design with adaptive localization addresses a plausible limitation of fixed-region approaches, and the use of pre-trained retinal priors is a reasonable direction for injecting anatomical knowledge. However, the significance is limited by the absence of detailed validation for the core knowledge-transfer assumption.

major comments (2)
  1. [Abstract / Experiments] Abstract and Experiments (assumed §4): The central claim attributes the reported AUC 98.5% and cross-domain gains on SMDG-19 to the Knowledge-Enhanced Convolutional Attention Module, yet no quantitative check is provided that the extracted retinal priors remain diagnostically accurate under domain shift (e.g., no prior-to-ground-truth agreement scores, no ablation of the module on out-of-domain sets, and no comparison of prior quality before/after transfer). If the priors degrade, the gains could arise from the tri-branch architecture or Dynamic Window Mechanism alone.
  2. [Abstract] Abstract: No information is given on training protocol, data splits, statistical testing, or ablation studies. This makes it impossible to assess whether the reported numbers reflect a genuine advance or could be reproduced under standard practices.
minor comments (1)
  1. [Abstract] The abstract refers to 'diverse baselines' and 'multiple datasets from the SMDG-19 benchmark' without naming them or providing per-dataset metrics; adding this detail would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and have revised the manuscript accordingly to improve clarity and strengthen the validation of our claims.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments (assumed §4): The central claim attributes the reported AUC 98.5% and cross-domain gains on SMDG-19 to the Knowledge-Enhanced Convolutional Attention Module, yet no quantitative check is provided that the extracted retinal priors remain diagnostically accurate under domain shift (e.g., no prior-to-ground-truth agreement scores, no ablation of the module on out-of-domain sets, and no comparison of prior quality before/after transfer). If the priors degrade, the gains could arise from the tri-branch architecture or Dynamic Window Mechanism alone.

    Authors: We agree that explicit quantitative validation of the retinal priors' accuracy under domain shift (such as prior-to-ground-truth agreement or isolated module ablations on out-of-domain data) would provide stronger support for attributing gains specifically to the Knowledge-Enhanced Convolutional Attention Module. Our current experiments show strong overall performance and cross-domain generalization on SMDG-19 with the full framework, but we did not include these specific prior-quality checks. In the revised manuscript, we will add an ablation isolating the module on the SMDG-19 out-of-domain sets along with any feasible prior-quality comparisons. revision: yes

  2. Referee: [Abstract] Abstract: No information is given on training protocol, data splits, statistical testing, or ablation studies. This makes it impossible to assess whether the reported numbers reflect a genuine advance or could be reproduced under standard practices.

    Authors: The training protocol, data splits, statistical testing, and ablation studies are detailed in the Experiments section of the full manuscript. However, we acknowledge that the abstract lacks a summary of these elements, which limits immediate assessment. We have revised the abstract to briefly note the experimental setup, including cross-validation and the presence of ablation studies confirming component contributions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance on external benchmarks

full rationale

The paper proposes a tri-branch architecture with a Dynamic Window Mechanism and Knowledge-Enhanced Convolutional Attention Module that incorporates priors from a pre-trained foundation model. All reported results consist of empirical AUC and accuracy metrics evaluated on independent public datasets (AIROGS and SMDG-19 benchmark). No equations, derivations, fitted parameters renamed as predictions, or self-referential definitions appear in the provided text. The central claims are therefore falsifiable against external data and do not reduce to the model's own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the assumption that a pre-trained foundation model supplies reliable retinal priors and that the dynamic window reliably surfaces diagnostically relevant patches; no free parameters, axioms, or invented entities are explicitly declared in the abstract.

pith-pipeline@v0.9.0 · 5558 in / 1091 out tokens · 66687 ms · 2026-05-10T15:02:47.556931+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 1 canonical work pages

  1. [1]

    World Health Organization, Blindness and visual impairment, WHO FactSheet,2023.URL:https://www.who.int/news-room/fact-sheets/ detail/blindness-and-visual-impairment, accessed: 2026-02-12

    2023

  2. [2]

    H. A. Quigley, A. T. Broman, The number of people with glaucoma worldwide in 2010 and 2020, British journal of ophthalmology 90 (2006) 262–267

    2010

  3. [3]

    Y.-C.Tham,X.Li,T.Y.Wong,H.A.Quigley,T.Aung,C.-Y.Cheng, Global prevalence of glaucoma and projections of glaucoma burden through2040:asystematicreviewandmeta-analysis,Ophthalmology 121 (2014) 2081–2090

    2014

  4. [4]

    J. M. Tielsch, J. Katz, H. A. Quigley, N. R. Miller, A. Sommer, Intraobserver and interobserver agreement in measurement of optic disc characteristics, Ophthalmology 95 (1988) 350–356

    1988

  5. [5]

    M.Rosano,A.Furnari,L.Gulino,C.Santoro,G.M.Farinella,Image- based navigation in real-world environments via multiple mid-level representations: Fusion models, benchmark and efficient evaluation, Autonomous Robots 47 (2023) 1483–1502

    2023

  6. [6]

    S.-A. Yuan, Z. Wang, F.-L. He, S.-W. Zhang, Z.-Y. Zhao, Gfhanet: Globalfeaturehybridattentionnetworkforsalientobjectdetectionin side-scan sonar images, IEEE Access (2024)

    2024

  7. [7]

    Jalili, A

    J. Jalili, A. Jiravarnsirikul, C. Bowd, B. Chuter, A. Belghith, M. H. Goldbaum, S. L. Baxter, R. N. Weinreb, L. M. Zangwill, M. Christo- pher, Glaucomadetectionandfeatureidentificationviagpt-4vfundus image analysis, Ophthalmology Science 5 (2025) 100667

    2025

  8. [8]

    Huang, X

    W. Huang, X. Liao, H. Chen, Y. Hu, W. Jia, Q. Wang, Deep local-to-global feature learning for medical image super-resolution, Computerized Medical Imaging and Graphics 115 (2024) 102374

    2024

  9. [9]

    S. Woo, J. Park, J.-Y. Lee, I. S. Kweon, Cbam: Convolutional block attention module, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 3–19

    2018

  10. [10]

    J. B. Jonas, G. C. Gusek, G. O. Naumann, Optic disk morphometry in high myopia, Graefe’s archive for clinical and experimental ophthalmology 226 (1988) 587–590

    1988

  11. [11]

    D.F.Garway-Heath,D.Poinoosawmy,F.W.Fitzke,R.A.Hitchings, Mappingthevisualfieldtotheopticdiscinnormaltensionglaucoma eyes, Ophthalmology 107 (2000) 1809–1815

    2000

  12. [12]

    G. C. Caiado, G. A. Samico, G. V. da Silva Filho, S. H. Teixeira, T. S. Prata, C. P. B. Gracitelli, A. Paranhos Jr, Correlation of optic nerve hemoglobin levels with structural and functional parameters in glaucoma, Scientific Reports 15 (2025) 19190

    2025

  13. [13]

    G.D.Joshi,J.Sivaswamy,S.Krishnadas, Opticdiskandcupsegmen- tationfrommonocularcolorretinalimagesforglaucomaassessment, IEEE transactions on medical imaging 30 (2011) 1192–1205

    2011

  14. [14]

    Cheng, J

    J. Cheng, J. Liu, Y. Xu, F. Yin, D. W. K. Wong, N.-M. Tan, D. Tao, C.-Y. Cheng, T. Aung, T. Y. Wong, Superpixel classification based optic disc and optic cup segmentation for glaucoma screening, IEEE transactions on medical imaging 32 (2013) 1019–1032

    2013

  15. [15]

    F. Li, W. Xiang, L. Zhang, W. Pan, X. Zhang, M. Jiang, H. Zou, Jointopticdiskandcupsegmentationforglaucomascreeningusinga region-based deep learning network, Eye 37 (2023) 1080–1087

    2023

  16. [16]

    Y. Xu, J. Liu, J. Cheng, F. Yin, N. M. Tan, D. W. K. Wong, C. Y. Cheng, Y. C. Tham, T. Y. Wong, Efficient optic cup localization based on superpixel classification for glaucoma diagnosis in digital fundus images, in: Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), IEEE, 2012, pp. 49–52

    2012

  17. [17]

    Y. Zhou, M. A. Chia, S. K. Wagner, M. S. Ayhan, D. J. Williamson, R. R. Struyven, T. Liu, M. Xu, M. G. Lozano, P. Woodward-Court, et al., A foundation model for generalizable disease detection from retinal images, Nature 622 (2023) 156–163

    2023

  18. [18]

    Q. Hou, Y. Zhou, J. H. L. Goh, K. Zou, S. M. E. Yew, S. Srinivasan, M. Wang, T. W. S. Lo, X. Lei, S. K. Wagner, et al., Can a natural image-based foundation model outperform a retina-specific model in detecting ocular and systemic diseases?, Ophthalmology Science 6 (2026) 100923

    2026

  19. [19]

    H.Fu,J.Cheng,Y.Xu,D.W.K.Wong,J.Liu,X.Cao, Jointopticdisc and cup segmentation based on multi-label deep network and polar transformation, IEEE transactions on medical imaging 37 (2018) 1597–1605

    2018

  20. [20]

    Truhn, T

    C.DeVente,K.A.Vermeer,N.Jaccard,H.Wang,H.Sun,F.Khader, D. Truhn, T. Aimyshev, Y. Zhanibekuly, T.-D. Le, et al., Airogs: Artificialintelligenceforrobustglaucomascreeningchallenge, IEEE transactions on medical imaging 43 (2023) 542–557

    2023

  21. [21]

    doi:10.34740/KAGGLE/DS/2329670

    R.Kiefer,Smdg,astandardizedfundusglaucomadataset,2023.URL: https://www.kaggle.com/ds/2329670. doi:10.34740/KAGGLE/DS/2329670

    work page doi:10.34740/kaggle/ds/2329670 2023

  22. [22]

    P. Li, J. Liu, Early diagnosis and quantitative analysis of stages in retinopathy of prematurity based on deep convolutional neural networks, TranslationalVisionScience&Technology11(2022)17– 17

    2022

  23. [23]

    J. Zhao, H. Lei, H. Xie, P. Li, Y. Liu, G. Zhang, B. Lei, Dual-branch attention network and swin spatial pyramid pooling for retinopathy of prematurity classification, in: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), IEEE, 2023, pp. 1–4

    2023

  24. [24]

    M. Alam, E. J. Zhao, C. K. Lam, D. L. Rubin, Segmentation- assisted fully convolutional neural network enhances deep learning performance to identify proliferative diabetic retinopathy, Journal of clinical medicine 12 (2023) 385

    2023

  25. [25]

    R.C.Joshi,A.K.Sharma,M.K.Dutta,Visiondeep-ai:Deeplearning- basedretinalbloodvesselssegmentationandmulti-classclassification framework for eye diagnosis, Biomedical Signal Processing and Control 94 (2024) 106273

    2024

  26. [26]

    Almeida, J

    J. Almeida, J. Kubicek, M. Penhaker, M. Cerny, M. Augustynek, A. Varysova, A. Bansal, J. Timkovic, Enhancing rop plus form diag- nosis: an automatic blood vessel segmentation approach for newborn fundus images, Results in engineering 24 (2024) 103054

    2024

  27. [27]

    Xiong, F

    H. Xiong, F. Long, M. S. Alam, J. Sang, Multi-glaucnet: A multi- task model for optic disc segmentation, blood vessel segmentation and glaucoma detection, Biomedical Signal Processing and Control 99 (2025) 106850

    2025

  28. [28]

    C.Guan,H.Ai,W.Wang,R.P.Singh,S.Song, Diffmcg:Adiffusion model with mask-conditioned guiding module for medical image classification, Neural Networks (2026) 108690

    2026

  29. [29]

    S.S.U.Shah,M.Huzaifa,H.K.K.Tareen,M.N.Bajwa, Optiguard: Generalized, attention-driven & explainable glaucoma classification, in:202547thAnnualInternationalConferenceoftheIEEEEngineer- ing in Medicine and Biology Society (EMBC), IEEE, 2025, pp. 1–7. Yuzhuo Zhou, Chi Liu, Sheng Shen, et al.:Preprint submitted to ElsevierPage 13 of 13

    2025