Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Improved Vessel Segmentation with Symmetric Rotation-Equivariant U-Net

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that replacing standard convolutions with symmetric rotation-equivariant kernels makes a U-Net segment retinal vessels accurately on rotated inputs while using 0.12M parameters, about 1.7% of a 6.97M-parameter…

desk verdict A plausible plug-in for rotation-robust U-Net segmentation, but the headline performance margins are within the noise floor and need a proper evaluation before they carry weight. read the letter →

arxiv 2501.14592 v1 pith:GH6Z7ZRW submitted 2025-01-24 eess.IV cs.CVcs.LG

classification eess.IVcs.CVcs.LG
keywords retinalvesselsegmentationrotationequivariancesymmetricconvolutionkernelU-Netparameterefficiencyfundusimagingsemanticconvolutionalneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a U-Net whose ordinary convolution kernels are replaced by symmetric rotation-equivariant kernels can segment retinal vessels accurately even when test images are rotated, while using far fewer trainable parameters than standard or state-of-the-art networks. On the DRIVE dataset, the proposed SRE U-Net reports the best accuracy, specificity, IoU, Dice, and AUC among the compared methods, and it outperforms a 6.97M-parameter state-of-the-art segmentation network using only 1.7% of that parameter count. The core idea is that rotational equivariance can be built directly into a standard segmentation architecture through a cheap kernel constraint, rather than through large group-convolution or steerable-filter stacks. If the paper is right, rotation-robust segmentation becomes practical in data- and compute-limited clinical settings, where image orientation is often inconsistent or unknown.

What carries the argument

The load-bearing object is the symmetric rotation-equivariant convolution kernel (SRE-Conv kernel). It partitions a $k \times k$ kernel into $b = \lfloor k/2 \rfloor + 2$ concentric bands; a binary index matrix $I$ assigns each kernel entry to one band, and the kernel is reconstructed as $\psi(\Theta_i \cdot I)$, where $\Theta_i$ holds only $b$ trainable values per channel. This construction forces the kernel to be centrally symmetric, which makes the convolution commute with 90-degree rotations and reflections and, the paper argues, generalizes to finer rotations through the learned band structure. Two architectural choices complete the mechanism: every convolution in the encoder and decoder is replaced with SRE-Conv, and decoder upsampling uses linear interpolation instead of trainable strided up-convolution, because strided up-convolution would break equivariance under rotation.

What would settle it

Train a comparison U-Net with the identical architecture and kernel-size schedule as SRE U-Net but standard asymmetric convolutions, matched for parameter count, and evaluate both on the DRIVE test set rotated in 1-degree steps. If the standard network matches SRE U-Net's Dice and AUC under rotation, the symmetric kernel is not what carries the reported robustness.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that applying SRE-Conv, a convolution kernel constrained to be centrally symmetric and encoded by a small number of band parameters, throughout a U-Net yields a segmentation model that is rotation- and reflection-equivariant in its feature maps and stays accurate when inputs are rotated by small angles (1 to 5 degrees) as well as by 90 degrees. In the reported experiments, the SRE U-Net (0.12M parameters) attains the highest averaged accuracy (0.9600), specificity (0.9870), IoU (0.6795), Dice (0.8087), and AUC (0.9767) among U-Net, Group U-Net, ES U-Net, FR U-Net, and Frangi baselines, with FR U-Net slightly ahead on sensitivity. The authors take this as evidence that symmetric kernels give the network both parameter efficiency and rotational robustness, and that the SRE kernel can serve as a drop-in replacement for the convolution layers in segmentation backbones.

Load-bearing premise

The central claim rests on the assumption that the improved accuracy and rotation stability come from the symmetric kernel's equivariance, rather than from the smaller model, the two-downsampling architecture, the interpolation upsampling, or from configuring kernel sizes on the test set.

Editorial extensions

If this is right

  • Retinal vessel segmentation can become rotation-robust without the memory overhead of group convolution: SRE U-Net uses 0.12M parameters versus 1.92M for Group U-Net and 0.72M for ES U-Net in the comparison.
  • Because the SRE kernel is designed as a drop-in replacement for standard convolutions, the same equivariance benefit could transfer to other U-Net-based segmentation tasks where object orientation is arbitrary.
  • A smaller, cheaper segmentation model makes rotation-equivariant learning usable in settings with limited training data or GPU memory.
  • The reported stability under small rotations implies fewer inconsistent predictions when clinical fundus images are acquired at slightly different angles.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The exact equivariance guarantee holds for grid-preserving rotations such as 90 degrees and reflections; the observed robustness to continuous 1-to-5-degree rotations is an empirical property that likely depends on the band structure, interpolation, and data augmentation together, not on a proven symmetry for arbitrary angles.
  • A cleaner test of the kernel's contribution would hold every architectural choice fixed and vary only kernel symmetry; the paper changes downsampling depth and upsampling method at the same time, so part of the gain may come from those changes rather than the symmetric kernel alone.
  • The same kernel could be evaluated on histopathology or other orientation-agnostic imaging where vessels and tissue structures appear at arbitrary angles; a public multi-dataset benchmark would show whether the parameter-efficiency advantage persists outside DRIVE.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes SRE U-Net, a U-Net variant in which all convolution kernels are replaced by symmetric rotation-equivariant (SRE) kernels with a small number of trainable parameters, two downsampling stages, and linear-interpolation upsampling. The authors evaluate the method on the DRIVE retinal vessel segmentation dataset, using both original test images and images rotated in 1-degree increments within +/-5 degrees, and compare against standard U-Net, Group U-Net, ES U-Net, FR U-Net, and the Frangi filter. They report that SRE U-Net achieves the best accuracy, specificity, IoU, Dice, and AUC while using only 0.12M parameters, and they claim that it surpasses the state-of-the-art FR U-Net while using 1.7% of its trainable parameters. The paper also includes an ablation over kernel sizes and qualitative difference maps.

Significance. If the empirical claims were established, the paper would provide a useful plug-in convolution kernel that improves rotation robustness and parameter efficiency for segmentation: reducing from 0.48M parameters for the U-Net baseline to 0.12M and from 6.97M for FR U-Net to 0.12M is striking, and the code is made publicly available. The use of a public benchmark, multiple metrics, and a rotation-based test protocol are strengths. However, the central quantitative claims are not currently supported: the kernel configuration is selected on the same test set used for the final evaluation, no error bars or significance tests are reported, and the headline differences are smaller than the likely noise floor. The 90-degree equivariance is also a design consequence of the kernel construction rather than an independently discovered property. The direction is plausible, but the reported margins need substantially stronger evidence.

major comments (3)
  1. [Sec. 3.4, Table 2] The kernel configuration k=[9,7,5] is selected by maximizing Dice on the same rotated test set used for the final evaluation: Table 2 reports Dice averaged over +/-5 degrees of 0.8064 for [9,7,5] versus 0.7960 for [7,7,7], and the selected configuration is labelled 'Proposed'. Because the test set contains only 20 images and no error bars are given, this test-set-based model selection makes the final headline numbers optimistically biased. The configuration choice should be made on a held-out validation set, or the authors should report nested cross-validation; without this, the reported margins cannot be taken at face value.
  2. [Sec. 3.2, Table 1] No error bars, repeated runs, or statistical significance tests are reported. The decisive comparisons are very small: at 0 degrees, SRE U-Net Dice is 0.8087 versus 0.8079 for FR U-Net and 0.7998 for U-Net; at +/-5 degrees, the values are 0.8046 versus 0.8027 and 0.7947, respectively. With n=20 test images and per-image Dice variability typical of this dataset, the standard error of the mean is on the order of 0.01-0.02, so the claimed 'significantly surpasses' and 'best across most metrics' are not established. The authors should provide per-image metric distributions, multiple training seeds, or statistical significance tests for the differences.
  3. [Sec. 2.1, Eq. (1); Sec. 3.2] The strict rotation-equivariance claim is broader than the construction supports. The SRE kernel is invariant under grid-preserving rotations and reflections, such as 90-degree rotations, by construction of the bands, but the paper's central evidence includes continuous small-angle rotations (1 to 5 degrees) where exact equivariance does not follow from Eq. (1). No analytical statement is given for small angles, and the experiments do not isolate the kernel's contribution from the architectural changes (two downsampling stages, linear-interpolation upsampling) or from the kernel-size selection. Please either narrow the equivariance claim to the discrete symmetry group or add an analytical or carefully controlled empirical argument for small-angle robustness.
minor comments (5)
  1. [Title, Sec. 2] The title contains a typo: 'EQUIV ARIANT' should be 'EQUIVARIANT'.
  2. [Sec. 2.2] Please clarify whether the standard U-Net baseline and the other equivariant baselines also use two downsampling stages and linear-interpolation upsampling; the text says 'same architecture' but the original U-Net uses transposed convolutions, so the comparison is otherwise not architecture-matched.
  3. [Fig. 3] The reported MSE values need a precise definition (mean over pixels, normalized, or image-level) and the qualitative result is shown for one image; a quantitative aggregate over the full test set would be more informative.
  4. [Sec. 2.1, Eq. (1)] The band-index construction would be clearer if the dimensions of the binary index matrix I and the reshaping operator psi were defined explicitly in the text; Fig. 1 partially provides this, but the notation in Eq. (1) is otherwise ambiguous.
  5. [Sec. 4] The claim that SRE-Conv can be 'easily integrated into other existing segmentation frameworks' is speculative and should be framed as future work rather than an established result.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline Dice advantage is a test-set-selected maximum rather than an independent prediction

  1. fitted input called prediction [Sec. 3.1 (Experimental Settings); Sec. 3.2 (Table 1); Sec. 3.4 (Table 2)]
    "We implement our SRE U-net using SRE-Conv with sizes k=[9,7,5] at each layer (ablation studies in Sec. 3.4 justify our model architecture). ... Table 2. SRE-Conv Ablation Studies. Model performance using different kernel size configurations ... Dice are averaged across rotation angles in the test set. ... [9, 7, 5](Proposed) 0.12M 0.8064"

    The same ±5° rotated DRIVE test set is used both to select the kernel configuration and to report the final evaluation. Table 2's Dice values are averaged over the test set, and the [9,7,5] row is the highest of the four configurations; the paper labels it 'Proposed'. Table 1 then reports this selected configuration as SRE U-Net's performance (Dice 0.8046 at ±5°, 0.8087 at 0°) and bases the headline conclusion on it: 'surpasses the performance of the SoTA segmentation method [20] ... using only 1.7% of the trainable parameters.' Thus the reported advantage is not an independent prediction of the method; it is the result of optimizing the model selector on the same evaluation data.

full rationale

The DRIVE segmentation experiments themselves are externally measured: test images, ground-truth labels, and baseline methods are independent of the paper's construction, so the core benchmarking is not circular by construction. The SRE-Conv kernel is taken from an overlapping-author preprint ([18], Du and Onofrey), but Eq. (1) and the binary index matrix make the central-symmetry construction explicit, so the exact 90°/reflection equivariance is a direct design property rather than a self-citation-dependent result. However, the paper's central empirical claim—that SRE U-Net 'significantly surpasses' baselines and the SoTA method with 1.7% of parameters—is partially circular because the [9,7,5] kernel configuration was selected by maximizing Dice on the same rotated test set used for the final comparison. The reported gap (e.g., 0.8046 vs 0.8027 for FR U-Net at ±5°) is also within a plausible noise floor for 20 test images, but that is a statistical-correctness concern rather than a circularity concern. The load-bearing circular step is the test-set-based model selection: the headline numbers are selected maxima, not independent predictions.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's own contribution is architectural application, so its correct behavior depends on the SRE kernel definition, the band-matrix construction, and the labeling of the test set; there are no newly invented physical entities.

free parameters (3)
  • Kernel size configuration per layer = [9,7,5]
    Chosen after evaluating four configurations (Table 2) on the same DRIVE test set used for the headline results, so it is a test-set-selected architectural hyperparameter.
  • Number of downsampling stages = 2
    Hand-selected in Sec 3.1 to compress the model; not ablated, yet the paper credits the architecture's parameter efficiency partly to this choice.
  • Training epochs = 6000
    Hand-set without a validation-based stopping criterion; on a 20-image training set this risks overfitting and affects the comparison.
assumptions (4)
  • domain assumption The SRE-Conv kernel defined by Eq. 1 with the binary band matrix I is rotation- and reflection-equivariant.
    The paper inherits this result from the authors' prior work (ref 18) and does not prove it here; the entire method rests on this property.
  • domain assumption All deep-learning baselines were trained under the same hyperparameters, augmentation, and architecture with only the equivariant layers changed, so differences in Table 1 are attributable to the kernel.
    Stated in Sec 3.1 but no training curves, config files, or seed analysis are provided to verify comparable training.
  • domain assumption The DRIVE test set of 20 images and the binary vessel labels are reliable enough to rank methods on the reported metrics.
    Used without independent annotation review or confidence intervals; a 20-image test set is small for detecting the small metric differences in Table 1.
  • ad hoc to paper Nearest-neighbor resampling of rotated test images does not introduce a bias favoring any particular method.
    The rotation protocol (Sec 3.1) is chosen to avoid interpolation artifacts, but no experiment compares resampling schemes or verifies that the equivariant method is not preferentially helped.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improved Vessel Segmentation with Symmetric Rotation-Equivariant U-Net." pith.science (2026). https://pith.science/paper/GH6Z7ZRW

@misc{pith2026250114592,
  author       = {Pith},
  title        = {Pith review of: Improved Vessel Segmentation with Symmetric Rotation-Equivariant U-Net},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GH6Z7ZRW}},
  note         = {Machine review of arXiv:2501.14592}
}
read the original abstract

Automated segmentation plays a pivotal role in medical image analysis and computer-assisted interventions. Despite the promising performance of existing methods based on convolutional neural networks (CNNs), they neglect useful equivariant properties for images, such as rotational and reflection equivariance. This limitation can decrease performance and lead to inconsistent predictions, especially in applications like vessel segmentation where explicit orientation is absent. While existing equivariant learning approaches attempt to mitigate these issues, they substantially increase learning cost, model size, or both. To overcome these challenges, we propose a novel application of an efficient symmetric rotation-equivariant (SRE) convolutional (SRE-Conv) kernel implementation to the U-Net architecture, to learn rotation and reflection-equivariant features, while also reducing the model size dramatically. We validate the effectiveness of our method through improved segmentation performance on retina vessel fundus imaging. Our proposed SRE U-Net not only significantly surpasses standard U-Net in handling rotated images, but also outperforms existing equivariant learning methods and does so with a reduced number of trainable parameters and smaller memory cost. The code is available at https://github.com/OnofreyLab/sre_conv_segm_isbi2025.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Equivariant Imaging Biomarkers for Robust Unsupervised Segmentation of Histopathology

    eess.IV 2025-05 conditional novelty 3.0 of 10

    A rotation-equivariant convolutional network produces more rotation-consistent unsupervised segmentations of prostate histopathology images than standard or group-equivariant baselines.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Improved Vessel Segmentation with Symmetric Rotation-Equivariant U-Net

    INTRODUCTION Segmentation is a fundamental task in medical imaging anal- ysis that involves identifying and delineating regions of inter- est, such as organs, lesions, and tissues. Accurate segmenta- tion is essential for many clinical applications, including dis- ease diagnosis, treatment planning, and monitoring of disease © 2025 IEEE. Personal use of t...

  2. [2]

    Symmetric Rotation-Equivariant (SRE) Convolution CNNs are equivariant with respect to translation

    METHODS 2.1. Symmetric Rotation-Equivariant (SRE) Convolution CNNs are equivariant with respect to translation. This means that translating the input to a convolutional layer will result in translating the output by the same amount. To achieve rota- tional equivariance, we apply SRE-Conv kernels [18], which parameterize the kernels to be centrally symmetr...

  3. [3]

    Experimental Setup Dataset: We evaluate using the public retina vessel DRIVE dataset [19], which consists of 40 2D RGB fundus images with paired binary vessel segmentation labels

    EXPERIMENTS AND RESULTS 3.1. Experimental Setup Dataset: We evaluate using the public retina vessel DRIVE dataset [19], which consists of 40 2D RGB fundus images with paired binary vessel segmentation labels. We partition the dataset into equal halves for training and testing. Baselines Comparison: We choose the standard U-Net [1] with the same architectu...

  4. [4]

    We integrate this kernel into the standard U-Net framework to learn rotation- and reflection-equivariant segmentation features, while also reducing the model size

    DISCUSSION AND CONCLUSION In this work, we present a novel application of an efficient symmetric rotation-equivariant convolution (SRE-Conv) ker- nel to the task of semantic segmentation. We integrate this kernel into the standard U-Net framework to learn rotation- and reflection-equivariant segmentation features, while also reducing the model size. The e...

  5. [5]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI. Springer, 2015, pp. 234– 241

  6. [6]

    Learning calibrated medi- cal image segmentation via multi-rater agreement mod- eling,

    Wei Ji, Yefeng Zheng, et al., “Learning calibrated medi- cal image segmentation via multi-rater agreement mod- eling,” in IEEE CVPR, 2021, pp. 12341–12351

  7. [7]

    nnu- net: a self-configuring method for deep learning-based biomedical image segmentation,

    Fabian Isensee, Klaus H Maier-Hein, et al., “nnu- net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

  8. [8]

    Atlas-based semantic segmentation of prostate zones,

    Jiazhen Zhang, John A Onofrey, et al., “Atlas-based semantic segmentation of prostate zones,” in MICCAI. Springer, 2022, pp. 570–579

Show all 27 references
  1. [9]

    Exploiting cyclic symmetry in convo- lutional neural networks,

    Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu, “Exploiting cyclic symmetry in convo- lutional neural networks,” in International conference on machine learning. PMLR, 2016, pp. 1889–1898

  2. [10]

    Rotation equivari- ant and invariant neural networks for microscopy image analysis,

    Benjamin Chidester, Jian Ma, et al., “Rotation equivari- ant and invariant neural networks for microscopy image analysis,” Bioinformatics, vol. 35, no. 14, pp. i530–i537, 2019

  3. [11]

    Group equivariant con- volutional networks,

    Taco Cohen and Max Welling, “Group equivariant con- volutional networks,” in International conference on machine learning. PMLR, 2016, pp. 2990–2999

  4. [12]

    Sample efficient semantic segmentation using rotation equivariant con- volutional networks,

    Jasper Linmans, Max Welling, et al., “Sample efficient semantic segmentation using rotation equivariant con- volutional networks,” arXiv preprint arXiv:1807.00583, 2018

  5. [13]

    Improved semantic segmentation for histopathology using rotation equivari- ant convolutional networks,

    Jim Winkens, Max Welling, et al., “Improved semantic segmentation for histopathology using rotation equivari- ant convolutional networks,” in MIDL, 2018

  6. [14]

    Beyond cnns: ex- ploiting further inherent symmetries in medical image segmentation,

    Shuchao Pang, Zhenmei Yu, et al., “Beyond cnns: ex- ploiting further inherent symmetries in medical image segmentation,” IEEE transactions on cybernetics, 2022

  7. [15]

    Rotation-equivariant semantic instance segmentation on biomedical images,

    Karl Bengtsson Bernander, Ingela Nystr ¨om, et al., “Rotation-equivariant semantic instance segmentation on biomedical images,” in Annual conference on med- ical image understanding and analysis . Springer, 2022, pp. 283–297

  8. [16]

    Harmonic networks: Deep translation and rotation equivariance,

    Daniel E Worrall, Gabriel J Brostow, et al., “Harmonic networks: Deep translation and rotation equivariance,” in IEEE CVPR, 2017, pp. 5028–5037

  9. [17]

    Learning steerable filters for rotation equivariant cnns,

    Maurice Weiler, Fred A Hamprecht, and Martin Storath, “Learning steerable filters for rotation equivariant cnns,” in IEEE CVPR, 2018, pp. 849–858

  10. [18]

    General e (2)- equivariant steerable cnns,

    Maurice Weiler and Gabriele Cesa, “General e (2)- equivariant steerable cnns,” Advances in neural infor- mation processing systems, vol. 32, 2019

  11. [19]

    A program to build E(N)-equivariant steerable CNNs,

    Gabriele Cesa, Leon Lang, and Maurice Weiler, “A program to build E(N)-equivariant steerable CNNs,” in ICLR, 2022

  12. [20]

    Use of sym- metric kernels for convolutional neural networks,

    Viacheslav Dudar and Vladimir Semenov, “Use of sym- metric kernels for convolutional neural networks,” in ICDSIAI. Springer, 2019, pp. 3–10

  13. [21]

    Rotated ring, radial and depth wise separable radial convolutions,

    Wolfgang Fuhl and Enkelejda Kasneci, “Rotated ring, radial and depth wise separable radial convolutions,” in IJCNN. IEEE, 2021, pp. 1–8

  14. [22]

    Sre-conv: Symmetric rotation equivariant convolution for biomedical image classification,

    Yuexi Du, John A Onofrey, et al., “Sre-conv: Symmetric rotation equivariant convolution for biomedical image classification,” arXiv preprint arXiv:2501.09753, 2025

  15. [23]

    Ridge-based vessel segmentation in color images of the retina,

    Joes Staal, Bram Van Ginneken, et al., “Ridge-based vessel segmentation in color images of the retina,”IEEE transactions on medical imaging , vol. 23, no. 4, pp. 501–509, 2004

  16. [24]

    Full-resolution net- work and dual-threshold iteration for retinal vessel and coronary angiograph segmentation,

    Wentao Liu, Feng Gao, et al., “Full-resolution net- work and dual-threshold iteration for retinal vessel and coronary angiograph segmentation,” IEEE journal of biomedical and health informatics , vol. 26, no. 9, pp. 4623–4634, 2022

  17. [25]

    Multiscale vessel enhancement filtering,

    Alejandro F Frangi, Max A Viergever, et al., “Multiscale vessel enhancement filtering,” in MICCAI. Springer, 1998, pp. 130–137

  18. [26]

    Attention gated networks: Learning to leverage salient regions in medi- cal images,

    Jo Schlemper, Daniel Rueckert, et al., “Attention gated networks: Learning to leverage salient regions in medi- cal images,” Medical image analysis , vol. 53, pp. 197– 207, 2019

  19. [27]

    Unet++: Re- designing skip connections to exploit multiscale features in image segmentation,

    Zongwei Zhou, Jianming Liang, et al., “Unet++: Re- designing skip connections to exploit multiscale features in image segmentation,” IEEE transactions on medical imaging, vol. 39, no. 6, pp. 1856–1867, 2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.