Adversarial Learning with Multiscale Features and Kernel Factorization for Retinal Blood Vessel Segmentation

Domenec Puig; Farhan Akram; Hatem A. Rashwan; Md. Mostafa Kamal Sarker; Mohamed Abdel-Nasser; Nidhi Pandey; Vivek Kumar Singh

arxiv: 1907.02742 · v1 · pith:SIZ6OLGFnew · submitted 2019-07-05 · 📡 eess.IV · cs.CV

Adversarial Learning with Multiscale Features and Kernel Factorization for Retinal Blood Vessel Segmentation

Farhan Akram , Vivek Kumar Singh , Hatem A. Rashwan , Mohamed Abdel-Nasser , Md. Mostafa Kamal Sarker , Nidhi Pandey , Domenec Puig This is my paper

Pith reviewed 2026-05-25 02:17 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords retinal blood vessel segmentationadversarial learningmultiscale featureskernel factorizationfundus imagesDRIVE datasetSTARE datasetsqueeze excitation

0 comments

The pith

An adversarial network using multiscale features and kernel factorization segments retinal blood vessels more accurately than prior methods on DRIVE and STARE datasets.

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

The paper proposes an adversarial learning approach for segmenting blood vessels in eye fundus images. The generator network applies spatial pyramid pooling, kernel factorization, and squeeze excitation blocks to capture features at multiple scales while keeping computation low. The discriminator uses convolutional layers plus a squeeze excitation block to tell apart generated masks from ground truth. Images receive edge sharpening and Gaussian regularization before input, and morphological operations clean the output masks. This combination is presented as delivering better qualitative and quantitative results than existing methods on the DRIVE and STARE datasets.

Core claim

The central claim is that an adversarial framework, with a generator that combines spatial pyramid pooling for multiscale context, kernel factorization for efficiency, and squeeze excitation blocks for feature emphasis, paired with a discriminator that includes convolutional layers and squeeze excitation to enforce realism, outperforms state-of-the-art vessel segmentation techniques on the DRIVE and STARE datasets after edge sharpening pre-processing and morphological post-processing.

What carries the argument

The adversarial generator-discriminator pair that uses spatial pyramid pooling, kernel factorization, and squeeze excitation blocks to process multiscale features at reduced complexity.

If this is right

The method produces segmentation masks that are both visually closer to ground truth and higher in standard quantitative scores than prior techniques on the two evaluated datasets.
Kernel factorization enables multiscale feature handling while lowering the number of parameters and computation.
Edge sharpening and Gaussian regularization prepare inputs so the network reaches a better solution for thin vessel structures.
Morphological operations after inference remove isolated noise points without altering the main vessel map.

Where Pith is reading between the lines

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

If the gains hold on varied imaging conditions, the pipeline could support automated detection of retinal diseases such as diabetic retinopathy.
The same multiscale adversarial structure might transfer to segmentation of other thin, branching structures in medical images.
Testing the trained model on images from different cameras or patient populations would check whether the reported improvements are dataset-specific.

Load-bearing premise

The described mix of pre-processing, network modules, adversarial training, and post-processing yields a general performance gain that extends past the specific traits and small size of the DRIVE and STARE datasets.

What would settle it

Running the method on a new or larger fundus image collection and finding that its accuracy metrics no longer exceed those of current leading approaches.

Figures

Figures reproduced from arXiv: 1907.02742 by Domenec Puig, Farhan Akram, Hatem A. Rashwan, Md. Mostafa Kamal Sarker, Mohamed Abdel-Nasser, Nidhi Pandey, Vivek Kumar Singh.

**Figure 1.** Figure 1: The architecture of the proposed model. 2 Proposed Methodology In this work, we propose an adversarial learning based retinal blood vessels segmentation method, which includes generator and discriminator networks. The generator network comprises an encoder and a decoder layers. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Visualizing encoder and decoder layers. 3.2 Results and Discussion In this section, the results are computed using the proposed method and compared with the state-of-the-art eye blood vessel segmentation methods both quantitatively and qualitatively [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: The segmentation output of the proposed model. Input GT FCN UNet cGAN Proposed [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Comparing the proposed model with the state-of-the-art methods. 0.9860, 0.7634 and 0.9830, which are 0.4%, 7.18% and 0.16% less than the second best Jiang et al. method [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

In this paper, we propose an efficient blood vessel segmentation method for the eye fundus images using adversarial learning with multiscale features and kernel factorization. In the generator network of the adversarial framework, spatial pyramid pooling, kernel factorization and squeeze excitation block are employed to enhance the feature representation in spatial domain on different scales with reduced computational complexity. In turn, the discriminator network of the adversarial framework is formulated by combining convolutional layers with an additional squeeze excitation block to differentiate the generated segmentation mask from its respective ground truth. Before feeding the images to the network, we pre-processed them by using edge sharpening and Gaussian regularization to reach an optimized solution for vessel segmentation. The output of the trained model is post-processed using morphological operations to remove the small speckles of noise. The proposed method qualitatively and quantitatively outperforms state-of-the-art vessel segmentation methods using DRIVE and STARE datasets.

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

Incremental application of known modules to retinal vessel segmentation without ablations to back up the performance claims.

read the letter

This paper describes an adversarial network for retinal vessel segmentation that incorporates spatial pyramid pooling, kernel factorization, and squeeze-excitation blocks in the generator, plus a discriminator with convolutional layers and another squeeze-excitation block. Preprocessing uses edge sharpening and Gaussian regularization, and the output gets cleaned with morphological operations. The authors claim it outperforms prior methods on the DRIVE and STARE datasets. What is new is the specific assembly of these modules into one pipeline for this task. The paper does a decent job of explaining how each component fits into the generator and discriminator and why they might help with multiscale features and reduced complexity. The soft spots are more significant. There are no ablation studies, so the gains cannot be pinned on the proposed modules rather than the preprocessing or postprocessing that other approaches also employ. The datasets are small and fixed, with the usual 20/20 and 10/10 splits, which makes it easy for hyperparameter choices to produce apparent improvements without real architectural advance. The abstract mentions quantitative outperformance but supplies no actual metrics or comparisons here, which weakens the central claim. This kind of paper is useful for practitioners who need a working segmentation method for eye fundus images and can test it on their own data. It will not interest readers looking for new theoretical insights or broadly applicable frameworks. I do not think it deserves peer review in its current state. The lack of controls on the experimental results is a load-bearing problem for an empirical methods paper. If the authors supplied ablations and full quantitative tables with statistical support, it could be reconsidered as an engineering contribution.

Referee Report

3 major / 0 minor

Summary. The paper proposes an adversarial framework for retinal blood vessel segmentation in fundus images. The generator uses spatial pyramid pooling, kernel factorization, and squeeze-excitation blocks for multiscale feature representation with reduced complexity; the discriminator combines convolutional layers with a squeeze-excitation block. Images are pre-processed with edge sharpening and Gaussian regularization, and outputs are post-processed with morphological operations to remove noise. The central claim is that the method qualitatively and quantitatively outperforms prior state-of-the-art vessel segmentation approaches on the DRIVE and STARE datasets.

Significance. If the performance gains are shown to be attributable to the proposed architectural components rather than shared pre/post-processing steps, the work could offer a practical improvement for automated retinal analysis in ophthalmology. The use of adversarial training with explicit multiscale and factorization modules is a reasonable direction for this domain, but the current lack of supporting quantitative evidence and component ablations prevents any assessment of whether the result would hold or generalize.

major comments (3)

[Abstract] Abstract: the central claim of quantitative outperformance on DRIVE and STARE is asserted without any reported metrics (Dice, sensitivity, specificity, AUC), tables, error bars, or statistical comparisons to baselines. This absence makes the empirical contribution unverifiable and is load-bearing for the paper's main result.
[Abstract] Abstract and method description: no ablation studies are described (e.g., generator with vs. without adversarial loss, with vs. without kernel factorization or SE blocks). Without these, it is impossible to attribute any measured margins to the proposed modules rather than the shared edge-sharpening pre-processing or morphological post-processing.
[Abstract] Abstract: the evaluation uses only the small fixed splits (DRIVE 20/20, STARE 10/10) with no mention of cross-validation, multiple runs, or larger external test sets. This raises a generalization concern for the claim that the combination produces a reliable improvement.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We respond to each major comment below and indicate planned revisions to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of quantitative outperformance on DRIVE and STARE is asserted without any reported metrics (Dice, sensitivity, specificity, AUC), tables, error bars, or statistical comparisons to baselines. This absence makes the empirical contribution unverifiable and is load-bearing for the paper's main result.

Authors: The full manuscript contains tables with the requested metrics (Dice, sensitivity, specificity, AUC) and comparisons to baselines on both datasets. We will revise the abstract to explicitly report these quantitative results and comparisons. revision: yes
Referee: [Abstract] Abstract and method description: no ablation studies are described (e.g., generator with vs. without adversarial loss, with vs. without kernel factorization or SE blocks). Without these, it is impossible to attribute any measured margins to the proposed modules rather than the shared edge-sharpening pre-processing or morphological post-processing.

Authors: The current manuscript does not contain component ablations. This is a valid observation. In the revision we will add ablation experiments that isolate the contribution of the adversarial loss, kernel factorization, and SE blocks while holding pre- and post-processing fixed. revision: yes
Referee: [Abstract] Abstract: the evaluation uses only the small fixed splits (DRIVE 20/20, STARE 10/10) with no mention of cross-validation, multiple runs, or larger external test sets. This raises a generalization concern for the claim that the combination produces a reliable improvement.

Authors: Fixed splits are the established protocol for DRIVE and STARE to permit direct comparison with prior work. We will add an explicit discussion of this evaluation choice and its limitations in the revised manuscript. revision: partial

Circularity Check

0 steps flagged

No circularity; purely empirical performance claims on fixed datasets

full rationale

The manuscript describes a GAN-based segmentation architecture (generator with SPP, kernel factorization, SE blocks; discriminator with conv+SE) plus pre/post-processing steps, then reports Dice/accuracy/sensitivity/specificity on DRIVE and STARE. No equations, derivations, or first-principles predictions exist that could reduce to fitted quantities by construction. All load-bearing claims are end-to-end empirical results on fixed public datasets; no self-citation chain, uniqueness theorem, or ansatz smuggling is invoked to justify the architecture. The absence of ablations is a methodological limitation but does not create circularity in any derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is an empirical deep-learning application. The central claim rests on the unstated premise that the proposed architecture and training procedure generalize beyond the training distribution of the two cited datasets. No explicit free parameters, mathematical axioms, or invented entities are described in the abstract.

pith-pipeline@v0.9.0 · 5704 in / 1072 out tokens · 27944 ms · 2026-05-25T02:17:01.961839+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[1]

IEEE Trans Pattern Anal Mach Intell37(9), 1904– 1916 (2015)

He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell37(9), 1904– 1916 (2015)

work page 1904
[2]

In: Proc

Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piece-wise threshold probing of a matched ﬁlter response. In: Proc. of the AMIA Symposium. p. 931. American Medical Informatics Association (1998)

work page 1998
[3]

In: Proc

Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proc. of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

work page 2018
[4]

Neurocomputing 309, 179–191 (2018)

Hu, K., Zhang, Z., Niu, X., Zhang, C., Xiao, F., Gao, X.: Retinal vessel segmenta- tion of color fundus images using multiscale convolutional neural network with an improved cross-entropy loss function. Neurocomputing 309, 179–191 (2018)

work page 2018
[5]

In: Proc

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with condi- tional adversarial networks. In: Proc. of the IEEE conference on computer vision and pattern recognition. pp. 1125–1134 (2017)

work page 2017
[6]

Comput Med Imaging Graph 68, 1–15 (2018)

Jiang, Z., Zhang, H., Wang, Y., Ko, S.B.: Retinal blood vessel segmentation using fully convolutional network with transfer learning. Comput Med Imaging Graph 68, 1–15 (2018)

work page 2018
[7]

The British journal of radiology 87(1040), 20130832 (2014)

MacGillivray, T., Trucco, E., Cameron, J., Dhillon, B., Houston, J., Van Beek, E.: Retinal imaging as a source of biomarkers for diagnosis, characterization and prognosis of chronic illness or long-term conditions. The British journal of radiology 87(1040), 20130832 (2014)

work page 2014
[8]

In: Medical imaging 2004: image processing

Niemeijer, M., Staal, J., van Ginneken, B., Loog, M., Abramoﬀ, M.D.: Comparative study of retinal vessel segmentation methods on a new publicly available database. In: Medical imaging 2004: image processing. vol. 5370, pp. 648–657. International Society for Optics and Photonics (2004)

work page 2004
[9]

Expert Syst Appl 112, 229–242 (2018) Title Suppressed Due to Excessive Length 9

Oliveira, A., Pereira, S., Silva, C.A.: Retinal vessel segmentation based on fully convolutional neural networks. Expert Syst Appl 112, 229–242 (2018) Title Suppressed Due to Excessive Length 9

work page 2018
[10]

IEEE T INTELL TRANSP 19(1), 263–272 (2018)

Romera, E., Alvarez, J.M., Bergasa, L.M., Arroyo, R.: Erfnet: Eﬃcient residual factorized convnet for real-time semantic segmentation. IEEE T INTELL TRANSP 19(1), 263–272 (2018)

work page 2018
[11]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015
[12]

In: 2017 International Conference on Digital Image Computing: Techniques and Applications (DICTA)

Soomro, T.A., Aﬁﬁ, A.J., Gao, J., Hellwich, O., Khan, M.A., Paul, M., Zheng, L.: Boosting sensitivity of a retinal vessel segmentation algorithm with convolutional neural network. In: 2017 International Conference on Digital Image Computing: Techniques and Applications (DICTA). pp. 1–8. IEEE (2017)

work page 2017
[13]

Comput Med Imaging Graph 55, 2–12 (2017)

Vostatek, P., Claridge, E., Uusitalo, H., Hauta-Kasari, M., F¨ alt, P., Lensu, L.: Performance comparison of publicly available retinal blood vessel segmentation methods. Comput Med Imaging Graph 55, 2–12 (2017)

work page 2017

[1] [1]

IEEE Trans Pattern Anal Mach Intell37(9), 1904– 1916 (2015)

He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell37(9), 1904– 1916 (2015)

work page 1904

[2] [2]

In: Proc

Hoover, A., Kouznetsova, V., Goldbaum, M.: Locating blood vessels in retinal images by piece-wise threshold probing of a matched ﬁlter response. In: Proc. of the AMIA Symposium. p. 931. American Medical Informatics Association (1998)

work page 1998

[3] [3]

In: Proc

Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proc. of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

work page 2018

[4] [4]

Neurocomputing 309, 179–191 (2018)

Hu, K., Zhang, Z., Niu, X., Zhang, C., Xiao, F., Gao, X.: Retinal vessel segmenta- tion of color fundus images using multiscale convolutional neural network with an improved cross-entropy loss function. Neurocomputing 309, 179–191 (2018)

work page 2018

[5] [5]

In: Proc

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with condi- tional adversarial networks. In: Proc. of the IEEE conference on computer vision and pattern recognition. pp. 1125–1134 (2017)

work page 2017

[6] [6]

Comput Med Imaging Graph 68, 1–15 (2018)

Jiang, Z., Zhang, H., Wang, Y., Ko, S.B.: Retinal blood vessel segmentation using fully convolutional network with transfer learning. Comput Med Imaging Graph 68, 1–15 (2018)

work page 2018

[7] [7]

The British journal of radiology 87(1040), 20130832 (2014)

MacGillivray, T., Trucco, E., Cameron, J., Dhillon, B., Houston, J., Van Beek, E.: Retinal imaging as a source of biomarkers for diagnosis, characterization and prognosis of chronic illness or long-term conditions. The British journal of radiology 87(1040), 20130832 (2014)

work page 2014

[8] [8]

In: Medical imaging 2004: image processing

Niemeijer, M., Staal, J., van Ginneken, B., Loog, M., Abramoﬀ, M.D.: Comparative study of retinal vessel segmentation methods on a new publicly available database. In: Medical imaging 2004: image processing. vol. 5370, pp. 648–657. International Society for Optics and Photonics (2004)

work page 2004

[9] [9]

Expert Syst Appl 112, 229–242 (2018) Title Suppressed Due to Excessive Length 9

Oliveira, A., Pereira, S., Silva, C.A.: Retinal vessel segmentation based on fully convolutional neural networks. Expert Syst Appl 112, 229–242 (2018) Title Suppressed Due to Excessive Length 9

work page 2018

[10] [10]

IEEE T INTELL TRANSP 19(1), 263–272 (2018)

Romera, E., Alvarez, J.M., Bergasa, L.M., Arroyo, R.: Erfnet: Eﬃcient residual factorized convnet for real-time semantic segmentation. IEEE T INTELL TRANSP 19(1), 263–272 (2018)

work page 2018

[11] [11]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015

[12] [12]

In: 2017 International Conference on Digital Image Computing: Techniques and Applications (DICTA)

Soomro, T.A., Aﬁﬁ, A.J., Gao, J., Hellwich, O., Khan, M.A., Paul, M., Zheng, L.: Boosting sensitivity of a retinal vessel segmentation algorithm with convolutional neural network. In: 2017 International Conference on Digital Image Computing: Techniques and Applications (DICTA). pp. 1–8. IEEE (2017)

work page 2017

[13] [13]

Comput Med Imaging Graph 55, 2–12 (2017)

Vostatek, P., Claridge, E., Uusitalo, H., Hauta-Kasari, M., F¨ alt, P., Lensu, L.: Performance comparison of publicly available retinal blood vessel segmentation methods. Comput Med Imaging Graph 55, 2–12 (2017)

work page 2017