ET-Net: A Generic Edge-aTtention Guidance Network for Medical Image Segmentation

Hang Dai; Huazhu Fu; Jianbing Shen; Ling Shao; Yanwei Pang; Zhijie Zhang

arxiv: 1907.10936 · v1 · pith:HFLAPAMNnew · submitted 2019-07-25 · 💻 cs.CV

ET-Net: A Generic Edge-aTtention Guidance Network for Medical Image Segmentation

Zhijie Zhang , Huazhu Fu , Hang Dai , Jianbing Shen , Yanwei Pang , Ling Shao This is my paper

Pith reviewed 2026-05-24 16:15 UTC · model grok-4.3

classification 💻 cs.CV

keywords medical image segmentationedge attentionboundary guidanceoptic disc segmentationretinal vessel segmentationlung segmentationdeep network fusion

0 comments

The pith

Embedding edge-attention representations from early layers guides decoding and raises medical segmentation accuracy.

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

The paper introduces ET-Net to address the common neglect of edge information in medical image segmentation. It uses an edge guidance module to learn boundary-focused representations in early encoding stages and transfers them to multi-scale decoding layers via weighted fusion. A sympathetic reader would care because precise boundaries directly affect clinical utility in tasks such as retinal vessel or lung analysis, where small errors matter.

Core claim

ET-Net embeds edge-attention representations learned by an edge guidance module in early encoding layers, transfers them to the decoding stages, and fuses them with a weighted aggregation module, producing higher segmentation accuracy than prior methods on optic disc/cup, vessel, and lung tasks.

What carries the argument

The edge guidance module, which extracts edge-attention representations in early encoding layers for transfer and fusion into the decoder.

If this is right

Segmentation outputs preserve finer boundary detail on retinal and chest images.
The method achieves higher accuracy than prior state-of-the-art networks on optic disc/cup segmentation.
Vessel segmentation in retinal images improves without changing the core encoder-decoder backbone.
Lung segmentation in both X-ray and CT benefits from the same edge transfer mechanism.

Where Pith is reading between the lines

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

The same early-to-late edge transfer pattern could be tested on non-medical segmentation benchmarks to check domain generality.
Replacing the weighted aggregation with other fusion operators would isolate whether the specific weighting step is essential.
If the edge module proves robust, it could be inserted as a plug-in into existing U-Net variants with minimal retraining.

Load-bearing premise

The accuracy gains come from the specific edge boundary information captured and transferred rather than from the extra parameters or training procedure added by any auxiliary branch.

What would settle it

An ablation that removes the edge guidance module or replaces its output with random features while keeping parameter count similar, and still matches the reported accuracy on the four tasks.

Figures

Figures reproduced from arXiv: 1907.10936 by Hang Dai, Huazhu Fu, Jianbing Shen, Ling Shao, Yanwei Pang, Zhijie Zhang.

**Figure 1.** Figure 1: illustrates the architecture of our ET-Net, which is primarily based on an encoder-decoder network, with the EGM and WAM modules appended on the end. The ResNet-50 [8] is utilized as the encoder network, which comprises of four Encoding-Blocks (E-Blocks), one for each different feature map resolution. For each E-Block, the inputs first go through a feature extraction stream, which consists of a stack of 1×… view at source ↗

**Figure 2.** Figure 2: Illustration of the E-Block and Weighted Block. ‘U’, ‘+’ and ‘×’ denote upsampling, addition, and multiplication layers, respectively. to guide the process of segmentation in the decoding path; 2) it supervises the early convolutional layers using the edge detection loss. In our EGM, the outputs of E-Block 2 are upsampled to the same resolution as the outputs of E-Block 1, and then fed into the 1×1−3×3 co… view at source ↗

**Figure 3.** Figure 3: Visualization of segmentation results. From left to right: optic disc/cup, and vessel segmentation in retinal fundus images, lung segmentation in Chest X-Ray and CT images [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

read the original abstract

Segmentation is a fundamental task in medical image analysis. However, most existing methods focus on primary region extraction and ignore edge information, which is useful for obtaining accurate segmentation. In this paper, we propose a generic medical segmentation method, called Edge-aTtention guidance Network (ET-Net), which embeds edge-attention representations to guide the segmentation network. Specifically, an edge guidance module is utilized to learn the edge-attention representations in the early encoding layers, which are then transferred to the multi-scale decoding layers, fused using a weighted aggregation module. The experimental results on four segmentation tasks (i.e., optic disc/cup and vessel segmentation in retinal images, and lung segmentation in chest X-Ray and CT images) demonstrate that preserving edge-attention representations contributes to the final segmentation accuracy, and our proposed method outperforms current state-of-the-art segmentation methods. The source code of our method is available at https://github.com/ZzzJzzZ/ETNet.

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

ET-Net adds an early edge guidance module with weighted decoder fusion and reports gains on four medical segmentation tasks, but the results do not isolate whether the edge focus itself causes the lift or whether extra capacity and fusion do the work.

read the letter

The edge guidance module in ET-Net produces consistent improvements over prior segmentation methods on four medical imaging tasks, but the paper leaves open whether those gains come specifically from the edge attention or from the added network capacity and fusion. The new part is the way they extract edge features early and inject them into multiple decoder scales with learned weights. That assembly is not in the cited work, even if the pieces are familiar. They do well by evaluating on retinal and chest images and by putting the code online so the numbers can be checked. The soft spot is the missing ablation. Without a control that adds a non-edge branch with the same parameter budget, the claim that edge-attention representations are what drive the accuracy rests on the assumption that the module's boundary focus is causal. The abstract states the contribution but the evidence is the performance numbers alone. This is useful for practitioners who need better boundary accuracy in segmentation pipelines. A serious referee should see it because the empirical scope is reasonable and the code is available for reproduction.

Referee Report

2 major / 1 minor

Summary. The paper proposes ET-Net for medical image segmentation, which uses an edge guidance module to learn edge-attention representations in early encoding layers and transfers them to multi-scale decoding layers via a weighted aggregation module. Experiments on four tasks (optic disc/cup and vessel segmentation in retinal images; lung segmentation in chest X-ray and CT) claim that this preserves edge information to improve accuracy and that ET-Net outperforms current SOTA methods, with source code released at https://github.com/ZzzJzzZ/ETNet.

Significance. If the reported gains are causally due to the edge-attention mechanism, the method could provide a reusable way to incorporate boundary cues into encoder-decoder segmentation networks for medical tasks where edge precision matters. The public code release is a clear strength that aids reproducibility.

major comments (2)

[Abstract] Abstract: the central claim that 'preserving edge-attention representations contributes to the final segmentation accuracy' and that ET-Net outperforms SOTA rests on the assumption that the edge guidance module supplies boundary-specific information rather than generic capacity or fusion benefits; no ablation that matches parameter count while removing edge supervision (e.g., a dummy branch) is described, leaving the attribution unverified.
[Experiments] Experiments section: although consistent gains across four datasets are asserted, the abstract supplies no Dice/IoU tables, ablation results, or statistical tests, so the magnitude and reliability of the improvement cannot be assessed from the provided text.

minor comments (1)

The abstract would be strengthened by including at least one key quantitative result (e.g., mean Dice improvement) to support the performance claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address the major comments point by point below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that 'preserving edge-attention representations contributes to the final segmentation accuracy' and that ET-Net outperforms SOTA rests on the assumption that the edge guidance module supplies boundary-specific information rather than generic capacity or fusion benefits; no ablation that matches parameter count while removing edge supervision (e.g., a dummy branch) is described, leaving the attribution unverified.

Authors: We agree that an ablation controlling for parameter count via a dummy branch without edge supervision would more rigorously isolate the contribution of the edge-attention mechanism. The revised manuscript will include this control experiment, reporting performance differences to confirm that gains arise from boundary-specific guidance rather than added capacity or fusion operations alone. revision: yes
Referee: [Experiments] Experiments section: although consistent gains across four datasets are asserted, the abstract supplies no Dice/IoU tables, ablation results, or statistical tests, so the magnitude and reliability of the improvement cannot be assessed from the provided text.

Authors: The abstract is intentionally concise and summarizes findings at a high level; quantitative Dice/IoU scores, full ablation tables, and SOTA comparisons appear in the Experiments section of the manuscript. We will add a short statement of key metric improvements to the abstract for accessibility. Statistical significance tests (e.g., paired t-tests or Wilcoxon) across the four datasets will also be included in the revision to address reliability concerns. revision: partial

Circularity Check

0 steps flagged

No circularity: claims rest on empirical validation, not derivation reducing to inputs.

full rationale

The paper presents an empirical architecture (ET-Net with edge guidance and weighted aggregation modules) and reports performance gains on four segmentation tasks via comparisons to prior methods. No mathematical derivation chain, predictions, or first-principles results are claimed that could reduce to fitted parameters, self-definitions, or self-citation chains by construction. The abstract's assertion that edge-attention representations contribute to accuracy is framed as an experimental outcome, not an equation or fit that equates to its own inputs. This is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The central claim rests on the empirical performance of a deep network whose weights are fitted to the four medical datasets; the architecture itself contains many free hyperparameters whose values are not derived from first principles.

free parameters (3)

network depth and channel counts
Standard U-Net-style hyperparameters that are chosen by the authors and affect capacity.
edge guidance module internal parameters
Learned weights inside the edge attention branch that are fitted during training.
weighted aggregation coefficients
Fusion weights between edge and segmentation features that are either learned or set by hand.

axioms (2)

domain assumption Edge maps extracted from early encoder layers contain information that is complementary to the main segmentation features.
Invoked when the authors state that transferring these representations improves accuracy.
domain assumption The four chosen datasets are representative of the medical segmentation tasks the method targets.
Required to generalize the reported outperformance beyond the specific experiments.

pith-pipeline@v0.9.0 · 5707 in / 1571 out tokens · 21917 ms · 2026-05-24T16:15:18.200087+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

IEEE TMI (2010)

Aquino, A., Gegundez-Arias, M.E., Marin, D.: Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques. IEEE TMI (2010)

work page 2010
[2]

In: CVPR (2018)

Berman, M., Rannen Triki, A., Blaschko, M.B.: The lov´ asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: CVPR (2018)

work page 2018
[3]

In: CVPR (2016)

Chen, H., Qi, X., et al.: DCAN: deep contour-aware networks for accurate gland segmentation. In: CVPR (2016)

work page 2016
[4]

IEEE TMI (2013)

Cheng, J., Liu, J., et al.: Superpixel classiﬁcation based optic disc and optic cup segmentation for glaucoma screening. IEEE TMI (2013)

work page 2013
[5]

IEEE TMI (2018)

Fu, H., Cheng, J., et al.: Joint Optic Disc and Cup Segmentation Based on Multi- Label Deep Network and Polar Transformation. IEEE TMI (2018)

work page 2018
[6]

In: MICCAI (2016)

Fu, H., Xu, Y., et al.: DeepVessel: Retinal Vessel Segmentation via Deep Learning and Conditional Random Field. In: MICCAI (2016)

work page 2016
[7]

IEEE TMI (2019)

Gu, Z., Cheng, J., et al.: CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE TMI (2019)

work page 2019
[8]

In: CVPR (2016)

He, K., Zhang, X., et al.: Deep residual learning for image recognition. In: CVPR (2016)

work page 2016
[9]

QIMS (2014) Title Suppressed Due to Excessive Length 9

Jaeger, S., Candemir, S., et al.: Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. QIMS (2014) Title Suppressed Due to Excessive Length 9

work page 2014
[10]

Radiographics (2015)

Mansoor, A., Bagci, U., et al.: Segmentation and Image Analysis of Abnormal Lungs at CT: Current Approaches, Challenges, and Future Trends. Radiographics (2015)

work page 2015
[11]

CMPB (2018)

Moccia, S., Momi, E.D., et al.: Blood vessel segmentation algorithms review of methods, datasets and evaluation metrics. CMPB (2018)

work page 2018
[12]

In: MICCAI (2015)

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. In: MICCAI (2015)

work page 2015
[13]

TPAMI (2017)

Shelhamer, E., Long, J., Darrell, T.: Fully convolutional networks for semantic segmentation. TPAMI (2017)

work page 2017
[14]

In: IEEE ISBI (2014)

Sivaswamy, J., Krishnadas, S.R., et al.: Drishti-gs: Retinal image dataset for optic nerve head(onh) segmentation. In: IEEE ISBI (2014)

work page 2014
[15]

IEEE TMI (2004)

Staal, J., Abr` amoﬀ, M.D., et al.: Ridge-based vessel segmentation in color images of the retina. IEEE TMI (2004)

work page 2004
[16]

IEEE TMI (2003)

Tsai, A., Yezzi, A., et al.: A shape-based approach to the segmentation of medical imagery using level sets. IEEE TMI (2003)

work page 2003
[17]

IEEE TMI (2019)

Wang, S., Yu, L., et al.: Patch-based output space adversarial learning for joint optic disc and cup segmentation. IEEE TMI (2019)

work page 2019
[18]

arXiv:1904.09146 (2019)

Wang, W., Lai, Q., et al.: Salient object detection in the deep learning era: An in-depth survey. arXiv:1904.09146 (2019)

work page arXiv 1904
[19]

IEEE PAMI (2019)

Wang, W., Shen, J., Ling, H.: A deep network solution for attention and aesthetics aware photo cropping. IEEE PAMI (2019)

work page 2019

[1] [1]

IEEE TMI (2010)

Aquino, A., Gegundez-Arias, M.E., Marin, D.: Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques. IEEE TMI (2010)

work page 2010

[2] [2]

In: CVPR (2018)

Berman, M., Rannen Triki, A., Blaschko, M.B.: The lov´ asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: CVPR (2018)

work page 2018

[3] [3]

In: CVPR (2016)

Chen, H., Qi, X., et al.: DCAN: deep contour-aware networks for accurate gland segmentation. In: CVPR (2016)

work page 2016

[4] [4]

IEEE TMI (2013)

Cheng, J., Liu, J., et al.: Superpixel classiﬁcation based optic disc and optic cup segmentation for glaucoma screening. IEEE TMI (2013)

work page 2013

[5] [5]

IEEE TMI (2018)

Fu, H., Cheng, J., et al.: Joint Optic Disc and Cup Segmentation Based on Multi- Label Deep Network and Polar Transformation. IEEE TMI (2018)

work page 2018

[6] [6]

In: MICCAI (2016)

Fu, H., Xu, Y., et al.: DeepVessel: Retinal Vessel Segmentation via Deep Learning and Conditional Random Field. In: MICCAI (2016)

work page 2016

[7] [7]

IEEE TMI (2019)

Gu, Z., Cheng, J., et al.: CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE TMI (2019)

work page 2019

[8] [8]

In: CVPR (2016)

He, K., Zhang, X., et al.: Deep residual learning for image recognition. In: CVPR (2016)

work page 2016

[9] [9]

QIMS (2014) Title Suppressed Due to Excessive Length 9

Jaeger, S., Candemir, S., et al.: Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. QIMS (2014) Title Suppressed Due to Excessive Length 9

work page 2014

[10] [10]

Radiographics (2015)

Mansoor, A., Bagci, U., et al.: Segmentation and Image Analysis of Abnormal Lungs at CT: Current Approaches, Challenges, and Future Trends. Radiographics (2015)

work page 2015

[11] [11]

CMPB (2018)

Moccia, S., Momi, E.D., et al.: Blood vessel segmentation algorithms review of methods, datasets and evaluation metrics. CMPB (2018)

work page 2018

[12] [12]

In: MICCAI (2015)

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. In: MICCAI (2015)

work page 2015

[13] [13]

TPAMI (2017)

Shelhamer, E., Long, J., Darrell, T.: Fully convolutional networks for semantic segmentation. TPAMI (2017)

work page 2017

[14] [14]

In: IEEE ISBI (2014)

Sivaswamy, J., Krishnadas, S.R., et al.: Drishti-gs: Retinal image dataset for optic nerve head(onh) segmentation. In: IEEE ISBI (2014)

work page 2014

[15] [15]

IEEE TMI (2004)

Staal, J., Abr` amoﬀ, M.D., et al.: Ridge-based vessel segmentation in color images of the retina. IEEE TMI (2004)

work page 2004

[16] [16]

IEEE TMI (2003)

Tsai, A., Yezzi, A., et al.: A shape-based approach to the segmentation of medical imagery using level sets. IEEE TMI (2003)

work page 2003

[17] [17]

IEEE TMI (2019)

Wang, S., Yu, L., et al.: Patch-based output space adversarial learning for joint optic disc and cup segmentation. IEEE TMI (2019)

work page 2019

[18] [18]

arXiv:1904.09146 (2019)

Wang, W., Lai, Q., et al.: Salient object detection in the deep learning era: An in-depth survey. arXiv:1904.09146 (2019)

work page arXiv 1904

[19] [19]

IEEE PAMI (2019)

Wang, W., Shen, J., Ling, H.: A deep network solution for attention and aesthetics aware photo cropping. IEEE PAMI (2019)

work page 2019