Boundary and Entropy-driven Adversarial Learning for Fundus Image Segmentation

Chi-Wing Fu; Kang Li; Lequan Yu; Pheng-Ann Heng; Shujun Wang; Xin Yang

arxiv: 1906.11143 · v2 · pith:5PGTOSG5new · submitted 2019-06-26 · 💻 cs.CV

Boundary and Entropy-driven Adversarial Learning for Fundus Image Segmentation

Shujun Wang , Lequan Yu , Kang Li , Xin Yang , Chi-Wing Fu , Pheng-Ann Heng This is my paper

Pith reviewed 2026-05-25 15:37 UTC · model grok-4.3

classification 💻 cs.CV

keywords unsupervised domain adaptationfundus image segmentationadversarial learningoptic disc cup segmentationboundary predictionentropy mapglaucoma screening

0 comments

The pith

Adversarial alignment of boundary predictions and entropy maps improves optic disc and cup segmentation across fundus image datasets.

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

The paper introduces an unsupervised domain adaptation approach to address differences between retinal fundus image datasets when segmenting the optic disc and cup. It applies adversarial learning so that boundary predictions and uncertainty maps in the target domain become similar to those in the source domain. This targets sharper boundaries and fewer high-uncertainty outputs in regions that are hard to segment. If the approach works, models can achieve higher accuracy on new unlabeled datasets without requiring additional annotations.

Core claim

The Boundary and Entropy-driven Adversarial Learning framework uses adversarial learning to encourage the boundary prediction and mask probability entropy map of the target domain to be similar to the source ones, generating more accurate boundaries and suppressing the high uncertainty predictions of OD and OC segmentation.

What carries the argument

Boundary and Entropy-driven Adversarial Learning framework that aligns boundary predictions and entropy maps between domains through adversarial training.

If this is right

More accurate optic disc and cup boundaries on target domain images that differ from the training set.
Reduced high-uncertainty predictions in ambiguous regions during OD and OC segmentation.
Higher overall segmentation performance than prior unsupervised domain adaptation techniques on the tested retinal datasets.

Where Pith is reading between the lines

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

The same alignment of boundary and uncertainty signals could extend to other medical segmentation tasks that suffer from scanner or population shifts.
Models trained this way may require fewer target-domain labels in clinical screening pipelines for glaucoma.
Further experiments could check whether the gains hold when the source and target datasets differ more strongly in resolution or patient demographics.

Load-bearing premise

Aligning boundary predictions and entropy maps of the target domain to the source domain via adversarial learning will produce improved segmentation accuracy on unlabeled target data.

What would settle it

A test on a held-out target fundus dataset in which the method shows no gain in boundary precision or reduction in uncertain predictions compared with standard adversarial domain adaptation baselines.

Figures

Figures reproduced from arXiv: 1906.11143 by Chi-Wing Fu, Kang Li, Lequan Yu, Pheng-Ann Heng, Shujun Wang, Xin Yang.

**Figure 1.** Figure 1: Comparison of the OD and OC predictions and the entropy maps of OD. The middle two columns show results on source and target domain images of the model trained without domain adaptation. The right most two columns show the results of our method on the same target domain image. Red color in the entropy maps ((b) and (d)) indicates high entropy values. Very recently, unsupervised domain adaptation methods ha… view at source ↗

**Figure 2.** Figure 2: Overview of our BEAL framework for unsupervised domain adaptation. The backbone is based on the DeepLabv3+ [2] architecture with Atrous Spatial Pyramid Pooling (ASPP) component followed by boundary and mask branches. We then apply Shannon Entropy (E) to obtain the entropy maps. Finally, we add two discriminators to apply adversarial learning on the boundary and entropy maps. structured (i.e., relative posi… view at source ↗

**Figure 3.** Figure 3: Qualitative results of pOSAL [12] and our method on the RIM-ONE-r3 dataset [5]. Our method can improve the segmentation results with accurate boundary, and generate clear prediction entropy maps. Green and blue lines represent the disc and cup contours, respectively. The entropy values are rescaled to [0,1] for better visualization. Quantitative analysis. We use the dice coefficients (DI) of OD and OC to q… view at source ↗

read the original abstract

Accurate segmentation of the optic disc (OD) and cup (OC)in fundus images from different datasets is critical for glaucoma disease screening. The cross-domain discrepancy (domain shift) hinders the generalization of deep neural networks to work on different domain datasets.In this work, we present an unsupervised domain adaptation framework,called Boundary and Entropy-driven Adversarial Learning (BEAL), to improve the OD and OC segmentation performance, especially on the ambiguous boundary regions. In particular, our proposed BEAL frame-work utilizes the adversarial learning to encourage the boundary prediction and mask probability entropy map (uncertainty map) of the target domain to be similar to the source ones, generating more accurate boundaries and suppressing the high uncertainty predictions of OD and OC segmentation. We evaluate the proposed BEAL framework on two public retinal fundus image datasets (Drishti-GS and RIM-ONE-r3), and the experiment results demonstrate that our method outperforms the state-of-the-art unsupervised domain adaptation methods. Codes will be available at https://github.com/EmmaW8/BEAL.

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.

Desk Editor's Note private letter to a colleague

BEAL adds adversarial alignment on boundary predictions and entropy maps to standard UDA for OD/OC fundus segmentation and reports gains versus prior methods on two small datasets, but the abstract contains no metrics so the size and reliability of the improvement remain unknown.

read the letter

The main takeaway is that this paper proposes BEAL, which layers two adversarial losses—one on boundary maps and one on entropy maps—onto an unsupervised domain adaptation pipeline to sharpen optic disc and cup boundaries on unlabeled target fundus images. The core idea is to push the target domain's boundary output and uncertainty map to resemble those from the source domain, which is a reasonable way to target the ambiguous regions that matter for glaucoma screening. The combination itself looks new relative to the domain adaptation papers they cite, and the focus on entropy to suppress high-uncertainty predictions is a sensible addition for medical images where boundary uncertainty is common. They evaluate on Drishti-GS and RIM-ONE-r3 and state that the method beats existing UDA baselines, with code promised on GitHub. That practical framing and the targeted losses are the parts that could be useful to others working on similar narrow adaptation tasks. The soft spots are straightforward. The abstract supplies no Dice scores, no tables, no ablation results, and no indication of how large the gains actually are, so the central claim cannot be checked from the summary alone. The datasets are small and public, which is typical but leaves open questions about variance across splits or larger domain gaps. The assumption that aligning those two maps will reliably improve segmentation accuracy is plausible but not automatic; it depends on whether the adversarial signals actually propagate to better pixel-level predictions. This work is aimed at researchers doing domain adaptation in retinal imaging rather than a broad audience. A reader already working on UDA for medical segmentation might pick up the boundary-plus-entropy trick if the full experiments hold up. It is coherent enough on its own terms to deserve a serious referee who can examine the actual numbers, ablations, and implementation details rather than a desk reject.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes Boundary and Entropy-driven Adversarial Learning (BEAL), an unsupervised domain adaptation framework for optic disc (OD) and optic cup (OC) segmentation in fundus images. It employs adversarial learning to encourage the boundary predictions and mask probability entropy maps (uncertainty maps) of the target domain to match those of the source domain, with the goal of producing more accurate boundaries and suppressing high-uncertainty predictions. The framework is evaluated on the Drishti-GS and RIM-ONE-r3 datasets and reported to outperform prior state-of-the-art UDA methods.

Significance. If the quantitative results and implementation details hold, the work offers a focused extension of adversarial domain adaptation that targets boundary precision and predictive uncertainty—two factors directly relevant to glaucoma screening. The emphasis on entropy-map alignment provides a concrete mechanism for handling ambiguous regions that standard feature-level adaptation often overlooks.

major comments (2)

[Abstract] Abstract: the claim of outperformance on Drishti-GS and RIM-ONE-r3 is stated without any numerical metrics, baseline comparisons, or error bars, preventing verification of the central empirical claim.
[Method] Method description: the assumption that adversarial alignment of boundary predictions and entropy maps will necessarily improve segmentation accuracy on unlabeled target data is presented without an ablation study isolating the contribution of each alignment term or a theoretical argument showing why this alignment reduces domain shift more effectively than existing boundary-aware or uncertainty-aware UDA losses.

minor comments (1)

[Abstract] The GitHub link for code release is a positive step toward reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. Below we provide point-by-point responses to the major comments. We will revise the paper to strengthen the presentation where the concerns are valid.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of outperformance on Drishti-GS and RIM-ONE-r3 is stated without any numerical metrics, baseline comparisons, or error bars, preventing verification of the central empirical claim.

Authors: We agree that the abstract would be strengthened by including concrete metrics. In the revised version we will add the key Dice scores for OD and OC on both target datasets together with the main baseline comparisons (e.g., AdaptSegNet, ADDA) so that the outperformance claim can be directly verified from the abstract. revision: yes
Referee: [Method] Method description: the assumption that adversarial alignment of boundary predictions and entropy maps will necessarily improve segmentation accuracy on unlabeled target data is presented without an ablation study isolating the contribution of each alignment term or a theoretical argument showing why this alignment reduces domain shift more effectively than existing boundary-aware or uncertainty-aware UDA losses.

Authors: The current manuscript presents the overall framework and reports end-to-end gains but does not contain a dedicated ablation isolating the boundary and entropy terms nor a formal theoretical argument. We will therefore add an ablation study (removing each adversarial term in turn) and expand the discussion section with additional analysis of why boundary and entropy alignment are complementary to existing feature-level UDA losses, supported by the observed qualitative improvements on ambiguous regions. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an unsupervised domain adaptation framework (BEAL) that applies adversarial learning to align boundary predictions and entropy maps between source and target domains for OD/OC segmentation. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim rests on the independent design of the adversarial objectives and empirical results on Drishti-GS and RIM-ONE-r3 datasets, without any reduction of outputs to inputs by construction or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review is abstract-only; no explicit free parameters, invented entities, or detailed axioms are extractable. The core premise rests on the existence of domain shift and the effectiveness of adversarial alignment.

axioms (2)

domain assumption Cross-domain discrepancy (domain shift) hinders generalization of deep neural networks for OD/OC segmentation.
Explicitly stated in the abstract as the central motivation.
domain assumption Adversarial alignment of boundary predictions and entropy maps will improve target-domain segmentation.
This is the load-bearing mechanism proposed in the abstract.

pith-pipeline@v0.9.0 · 5725 in / 1303 out tokens · 28435 ms · 2026-05-25T15:37:11.393737+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 1 internal anchor

[1]

In: MLMI

Chen, C., Dou, Q., Chen, H., Heng, P.A.: Semantic-aware generative adversarial nets for unsupervised domain adaptation in chest X-ray segmentation. In: MLMI. pp. 143–151. Springer, Cham (2018)

work page 2018
[2]

In: ECCV

Chen, L.C., Zhu, Y., Papandreou, G., et al.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801–818 (2018) Title Suppressed Due to Excessive Length 9

work page 2018
[3]

In: IJCAI

Dou, Q., Ouyang, C., Chen, C., Chen, H., Heng, P.A.: Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversar- ial loss. In: IJCAI. pp. 691–697 (2018)

work page 2018
[4]

IEEE TMI 37(7), 1597–1605 (2018)

Fu, H., Cheng, J., Xu, Y., et al.: Joint optic disc and cup segmentation based on multi-label deep network and polar transformation. IEEE TMI 37(7), 1597–1605 (2018)

work page 2018
[5]

In: 24th interna- tional symposium on computer-based medical systems (CBMS)

Fumero, F., Alay´ on, S., Sanchez, J.L., Sigut, J., Gonzalez-Hernandez, M.: RIM- ONE: An open retinal image database for optic nerve evaluation. In: 24th interna- tional symposium on computer-based medical systems (CBMS). pp. 1–6 (2011)

work page 2011
[6]

FCNs in the Wild: Pixel-level Adversarial and Constraint-based Adaptation

Hoﬀman, J., Wang, D., Yu, F., Darrell, T.: FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[7]

In: ISBI

Javanmardi, M., Tasdizen, T.: Domain adaptation for biomedical image segmen- tation using adversarial training. In: ISBI. pp. 554–558. IEEE (2018)

work page 2018
[8]

In: IPMI

Kamnitsas, K., Baumgartner, C., et al.: Unsupervised domain adaptation in brain lesion segmentation with adversarial networks. In: IPMI. pp. 597–609. Springer, Cham. (2017)

work page 2017
[9]

JSM Biomedical Imaging Data Papers 2(1), 1004 (2015)

Sivaswamy, J., Krishnadas, S., Chakravarty, A., et al.: A comprehensive retinal image dataset for the assessment of glaucoma from the optic nerve head analysis. JSM Biomedical Imaging Data Papers 2(1), 1004 (2015)

work page 2015
[10]

In: CVPR

Tsai, Y.H., Hung, W.C., Schulter, S., et al.: Learning to adapt structured output space for semantic segmentation. In: CVPR. pp. 7472–7481 (2018)

work page 2018
[11]

In: CVPR

Vu, T.H., Jain, H., Bucher, M., Cord, M., P´ erez, P.: ADVENT: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR. pp. 2517–2526 (2019)

work page 2019
[12]

IEEE TMI p

Wang, S., Yu, L., Yang, X., Fu, C.W., Heng, P.A.: Patch-based output space adver- sarial learning for joint optic disc and cup segmentation. IEEE TMI p. to appear (2019)

work page 2019
[13]

In: MICCAI

Zhang, Y., Miao, S., Mansi, T., Liao, R.: Task driven generative modeling for unsupervised domain adaptation: Application to X-ray image segmentation. In: MICCAI. pp. 599–607. Springer (2018)

work page 2018
[14]

In: ICCV

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV. pp. 2223–2232 (2017)

work page 2017

[1] [1]

In: MLMI

Chen, C., Dou, Q., Chen, H., Heng, P.A.: Semantic-aware generative adversarial nets for unsupervised domain adaptation in chest X-ray segmentation. In: MLMI. pp. 143–151. Springer, Cham (2018)

work page 2018

[2] [2]

In: ECCV

Chen, L.C., Zhu, Y., Papandreou, G., et al.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV. pp. 801–818 (2018) Title Suppressed Due to Excessive Length 9

work page 2018

[3] [3]

In: IJCAI

Dou, Q., Ouyang, C., Chen, C., Chen, H., Heng, P.A.: Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversar- ial loss. In: IJCAI. pp. 691–697 (2018)

work page 2018

[4] [4]

IEEE TMI 37(7), 1597–1605 (2018)

Fu, H., Cheng, J., Xu, Y., et al.: Joint optic disc and cup segmentation based on multi-label deep network and polar transformation. IEEE TMI 37(7), 1597–1605 (2018)

work page 2018

[5] [5]

In: 24th interna- tional symposium on computer-based medical systems (CBMS)

Fumero, F., Alay´ on, S., Sanchez, J.L., Sigut, J., Gonzalez-Hernandez, M.: RIM- ONE: An open retinal image database for optic nerve evaluation. In: 24th interna- tional symposium on computer-based medical systems (CBMS). pp. 1–6 (2011)

work page 2011

[6] [6]

FCNs in the Wild: Pixel-level Adversarial and Constraint-based Adaptation

Hoﬀman, J., Wang, D., Yu, F., Darrell, T.: FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[7] [7]

In: ISBI

Javanmardi, M., Tasdizen, T.: Domain adaptation for biomedical image segmen- tation using adversarial training. In: ISBI. pp. 554–558. IEEE (2018)

work page 2018

[8] [8]

In: IPMI

Kamnitsas, K., Baumgartner, C., et al.: Unsupervised domain adaptation in brain lesion segmentation with adversarial networks. In: IPMI. pp. 597–609. Springer, Cham. (2017)

work page 2017

[9] [9]

JSM Biomedical Imaging Data Papers 2(1), 1004 (2015)

Sivaswamy, J., Krishnadas, S., Chakravarty, A., et al.: A comprehensive retinal image dataset for the assessment of glaucoma from the optic nerve head analysis. JSM Biomedical Imaging Data Papers 2(1), 1004 (2015)

work page 2015

[10] [10]

In: CVPR

Tsai, Y.H., Hung, W.C., Schulter, S., et al.: Learning to adapt structured output space for semantic segmentation. In: CVPR. pp. 7472–7481 (2018)

work page 2018

[11] [11]

In: CVPR

Vu, T.H., Jain, H., Bucher, M., Cord, M., P´ erez, P.: ADVENT: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR. pp. 2517–2526 (2019)

work page 2019

[12] [12]

IEEE TMI p

Wang, S., Yu, L., Yang, X., Fu, C.W., Heng, P.A.: Patch-based output space adver- sarial learning for joint optic disc and cup segmentation. IEEE TMI p. to appear (2019)

work page 2019

[13] [13]

In: MICCAI

Zhang, Y., Miao, S., Mansi, T., Liao, R.: Task driven generative modeling for unsupervised domain adaptation: Application to X-ray image segmentation. In: MICCAI. pp. 599–607. Springer (2018)

work page 2018

[14] [14]

In: ICCV

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV. pp. 2223–2232 (2017)

work page 2017