Pith. sign in

REVIEW 4 major objections 5 minor 43 references

SFD-Mamba2Net claims that joining a Hessian-based structural prior, a dual-axis bidirectional Mamba2 bottleneck, and wavelet high-frequency refinement yields the best coronary-artery segmentation and stenosis detection among eight methods o

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

SFD-Mamba2Net combines Hessian vesselness priors, a dual-stream Mamba2 module, and wavelet high-frequency enhancement to segment coronary arteries and detect stenoses in angiography, reporting improved Dice and stenosis detection but not uniformly across all metrics.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A reasonable incremental Mamba2-vessel-segmentation paper whose headline claim is contradicted by its own Table 2, and whose dataset description has an unresolved count mismatch; worth engaging but needs real revision. the 4 major comments →

arxiv 2509.08934 v2 pith:D35LJLHZ submitted 2025-09-10 cs.CV

SFD-Mamba2Net: Structure-Guided Frequency-Enhanced Dual-Stream Mamba2 Network for Coronary Artery Segmentation

Nan Mu , Ruiqi Song , Zhihui Xu , Jingfeng Jiang , Chen Zhao This is my paper
classification cs.CV
keywords coronary artery segmentationinvasive coronary angiographyMamba2state space dualityfrequency-domain enhancementwavelet decompositionstenosis detectionvesselness filter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper is trying to establish that one end-to-end network can handle both coronary artery segmentation and stenosis detection in invasive coronary angiography, despite low contrast, noise, and fine multi-scale vessels. The proposed SFD-Mamba2Net combines three mechanisms: a parameter-free curvature-aware structural prior from multi-scale Hessian vesselness at the input, a dual-axis bidirectional Mamba2 bottleneck built on state-space duality, and a wavelet-based decoder that progressively sharpens high-frequency detail while preserving low-frequency topology. On the authors' dataset, which is reported inconsistently (Section 4.1: 187 LCA and 127 RCA images; Table 1: 403 LCA and 213 RCA images), the network is claimed to outperform seven state-of-the-art models on eight segmentation metrics and to achieve the highest stenosis-detection recall and precision (TPR=0.60, PPV=0.64). A sympathetic reader would care because if those results generalize, automated ICA analysis becomes a realistic decision-support tool for grading coronary artery disease.

Core claim

The central claim is that SFD-Mamba2Net, by explicitly injecting geometric vesselness priors, bidirectional state-space modeling along orthogonal axes, and progressive wavelet high-frequency refinement, produces more complete and boundary-accurate coronary vessel segmentations than existing U-Net, Transformer, state-space, and frequency-based models. The paper reports the best Dice (88.10%), F1 (89.22%), IoU (80.81%), Accuracy (98.80%), and Specificity (99.50%) among eight compared methods, and in a centerline-based stenosis-detection pipeline it reports the highest true positive rate (0.60) and positive predictive value (0.64), with lower absolute and relative severity errors than the other

What carries the argument

The architecture rests on three named modules. CASE (Curvature-Aware Structural Enhancement) is a parameter-free multi-scale Hessian vesselness filter applied to the input to amplify slender tubular structures and suppress background. AA-DS Mamba2 (Axial-Alternating Dual-Stream Mamba2) is a bottleneck module that flattens features along width and height in two alternating paths, each containing forward and backward Mamba2 sequences under the State Space Duality (SSD) framework, to capture long-range spatial dependencies with linear complexity. PHFP (Progressive High-Frequency Perception) is a decoder module that recursively applies Haar wavelet decomposition, enhances high-frequency sub-band

Load-bearing premise

The central claim rests on the unstated assumption that the reported metrics come from a fair, patient-independent train/test split on a correctly described dataset; the paper gives no split procedure and its abstract counts (187 LCA + 127 RCA) disagree with Table 1 (403 LCA + 213 RCA), so leakage could inflate all results.

What would settle it

Run the same eight methods under a strictly patient-external split (all frames of a given patient placed entirely in training or entirely in testing) on the exact dataset reported in Table 1, and check whether SFD-Mamba2Net still leads VM-UNet by the reported margins; a drop of Dice or stenosis metrics to within one standard deviation, or evidence that training and test frames share the same patients, would falsify the generalization claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported results are correct, automated ICA analysis can localize and grade coronary stenoses with fewer misses and false positives than the seven comparison models, including for moderate and severe lesions.
  • The ablation results indicate each module contributes and the full combination is best, so the proposed recipe of geometric priors, state-space long-range modeling, and wavelet frequency refinement is a workable template for low-contrast tubular-structure segmentation.
  • Because the CASE module is parameter-free, it can be attached to other backbones during inference without adding learnable parameters, providing structural guidance at no training cost.
  • The stratified stenosis results imply the pipeline can distinguish minimal, mild, moderate, and severe narrowing, supporting its stated clinical value for risk stratification and interventional planning.

Where Pith is reading between the lines

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

  • A strict patient-independent split is the decisive next test: the paper never describes the train/test split or whether all frames from the same patient were kept together, so if nearby frames from the same angiography run appear in both training and testing, the reported margins could shrink under external validation.
  • The wavelet high-frequency refinement idea is likely transferable to other thin-structure segmentation problems, such as retinal vessels, airways, or road networks, since PHFP is not coronary-specific.
  • The paper's own Section 5.4 acknowledges that 2D ICA is a projection of 3D arteries; this sets a geometric upper bound on stenosis grading accuracy, suggesting that multi-view fusion or 3D reconstruction should be the next priority rather than further single-image accuracy gains.
  • The dataset-count inconsistency (314 vs 616 images) should be resolved and the metrics recomputed on the clarified set before accepting the headline comparison.
  • The parameter-free vesselness prior could also serve as a self-supervision signal, not just a concatenated input, in settings where manual annotations are scarce.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SFD-Mamba2Net, an encoder-decoder network for coronary artery segmentation in invasive coronary angiography (ICA), integrating a Hessian-based Curvature-Aware Structure Enhancement (CASE/MASE) module, an Axial-Alternating Dual-Stream Mamba2 (AA-DS Mamba2) bottleneck, and a Progressive High-Frequency Perception (PHFP) decoder module based on multi-level wavelet decomposition. The authors evaluate segmentation on a private ICA dataset against seven published models using eight metrics, and also report stenosis detection metrics obtained by feeding segmentation outputs into a centerline-based stenosis-detection algorithm. The abstract and conclusion claim consistent superiority across all eight segmentation metrics and best stenosis-detection performance.

Significance. If the reported results are reliable, the architecture offers a useful combination of structural priors, bidirectional state-space modeling, and frequency-domain refinement for vascular imaging, and the public code release is a practical strength. The paper is an empirical architecture comparison rather than a theoretical contribution, so its value rests on the validity and clarity of the experimental evaluation. The stenosis-detection evaluation reuses the authors' prior algorithm with several hand-set thresholds, but because the same pipeline is applied to every segmentation model, the relative comparisons are not circular. However, the headline claim of consistent superiority is contradicted by the paper's own quantitative results, and the private dataset and missing split details limit the generalizability claims.

major comments (4)
  1. [Abstract and §4.2, Table 2] The abstract states that SFD-Mamba2Net 'consistently outperformed state-of-the-art methods across eight segmentation metrics.' This is not supported by Table 2. SFD-Mamba2Net ranks first only on Dice, Acc, F1, IoU, and Spec; it is second on Sens (87.48 vs. DSCNet 91.84), third on HD95 (7.57 vs. VM-UNet 5.64), and second on ASSD (1.25 vs. ConDSeg 1.17). Section 5.1 concedes these exceptions, but the abstract and conclusion still make the unqualified claim. This is a load-bearing inconsistency that must be fixed by either rewriting the claim to 'best on 5 of 8 metrics' or providing evidence that the differences on the three remaining metrics are not meaningful.
  2. [§4.1 and Table 1] The dataset description is internally inconsistent. Section 4.1 says 'The dataset comprises 187 left coronary artery (LCA) images and 127 right coronary artery (RCA) images,' while Table 1 lists totals of 403 LCA and 213 RCA images. The paper never specifies the train/test split procedure, whether it is patient-independent, or how many images/patients are in each split. Section 4.4 mentions '132 test images' without explaining how this number follows from the stated totals. If multiple frames from the same patient appear in both training and test sets, the reported performance ranking could reflect memorization rather than generalization. This concerns the central comparative claim and must be resolved with exact split details and patient-level separation.
  3. [§4.3, Table 3] The ablation results do not show that the full model is uniformly best. In Table 3, the U+A+P variant achieves HD95 = 7.10 ± 1.41, while the full SFD-Mamba2Net has HD95 = 7.57 ± 0.35; the full model is also not best on ASSD relative to U+M+A (1.25 vs. 1.51, so it is best) but the HD95 regression is unexplained. The text in §4.3 says the full model 'achieves the best performance across Dice, Acc, F1, IoU, and Specificity,' which is accurate, but the qualitative statement that the full model integrates the modules 'to yield an accurate segmentation map' glosses over the HD95 anomaly. The authors should either explain the HD95 behavior or temper the claim that the full model is the optimal combination.
  4. [§4.4, Algorithm 3 and Table 4] The stenosis detection evaluation depends on several threshold parameters (detection radius r = 10, D_thresh = 1.8 mm, L_thresh = 20, b_thresh = 0.1, plus second-derivative heuristics). No sensitivity analysis is provided, and the same pipeline is used for all models, so the relative ranking is fair. However, the absolute TPR/PPV numbers (0.60/0.64) are claimed as clinically relevant while the detection ratio is only 44.74% and 1,212 false negatives remain in 132 test images. The clinical-applicability conclusion in §5.2 is stronger than the evidence supports; the authors should moderate it or add a failure analysis.
minor comments (5)
  1. [Title and throughout] The title has a typo: 'Strcture-Guided' should be 'Structure-Guided.' Also, the module is called CASE in the abstract and Section 3.2, but Figure 2 and Section 4.3 call it MASE; please unify the nomenclature.
  2. [§4.1, Eq. (15)] The IoU formula in Eq. (15) is written as TP / (TN + FP + FN), which is not the standard IoU definition (TP / (TP + FP + FN)). If this is a typo, correct it; if not, the metric is misdefined and the reported IoU values are suspect.
  3. [Table 2 header] The column header says 'Swim-Unet' but the model and references use 'Swin-Unet.' Please correct the typo.
  4. [§4.3, Table 3] The ablation variant names are inconsistent with the module names: MASE in the table versus CASE in the text. Also, the standard deviations in Table 3 (e.g., HD95 = 7.10 ± 1.41 for U+A+P) overlap substantially with those of the full model (7.57 ± 0.35), so the discussion of 'further optimization' should acknowledge this overlap.
  5. [§5.1, Figure 11 reference] The text says 'Visualizations of difference maps (Figure 11)' but Figure 11 shows PHFP Grad-CAM heatmaps; the difference maps are in Figure 17. Please correct the cross-reference.

Circularity Check

0 steps flagged

No circularity found; empirical architecture comparison against external baselines, with a minor non-load-bearing self-citation in stenosis evaluation.

full rationale

The paper is an empirical architecture study and contains no theoretical derivation chain that could collapse into its inputs. The segmentation model is assembled from existing building blocks (conv-BN-LeakyReLU encoder/decoder, Hessian-based vesselness, Mamba2 SSD, Haar wavelet transforms) and is compared directly with seven external baselines (U-Net, U-Net++, Swin-Unet, DSCNet, VM-UNet, Perspective-UNet, ConDSeg) using the same metrics and protocol (Tables 2-3). The stenosis evaluation in Section 4.4 does reuse the authors' own prior detection algorithm [39], and Chen Zhao and Zhihui Xu are co-authors of both works, so a self-citation is present. However, it is not load-bearing in a circular sense: the algorithm is a fixed, externally published pipeline with hand-set thresholds (r=10 pixels, L_thresh=20, D_thresh=1.8 mm, b_thresh=0.1), and it is applied identically to every model's predicted vessel contours. Relative rankings in Table 4 are therefore not forced by construction, and no parameter is fitted to the proposed model and then renamed as a prediction. The paper's internal contradictions (e.g., the abstract claims superiority on all eight segmentation metrics while Table 2 shows DSCNet, VM-UNet, and ConDSeg lead on Sensitivity, HD95, and ASSD; the dataset counts in the abstract vs. Table 1 differ) are correctness and reporting concerns, not circularity. The Limitations section (5.4) also explicitly acknowledges dataset scale and static-view limitations, which further indicates the authors' claims are empirical rather than derivational. Because the central comparison is self-contained against external benchmarks, the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central empirical claim rests on standard deep learning assumptions: representative training data, correct annotations, and a fair evaluation protocol. The most fragile assumptions are the validity of the Hessian vesselness prior for low-contrast ICA images and the reliability of a privately held dataset with an unspecified split.

free parameters (5)
  • Vesselness hyperparameters beta and c in Eq. (7) = not specified
    These control the sensitivity of the CASE vesselness response; without stated values the module is not exactly reproducible.
  • Gaussian scale set sigma in CASE module = 1, 2, 3
    The multiscale range is chosen by hand with no ablation justifying the choice.
  • Stenosis detection radius r = 10 pixels
    Set empirically in Algorithm 3; affects matching between predicted and ground-truth stenotic segments.
  • Stenosis detection thresholds = D_thresh=1.8 mm, L_thresh=20 px, b_thresh=0.1
    Clinical heuristics in the detection pipeline; fixed for all models but influence absolute TPR/PPV values.
  • PHFP wavelet decomposition levels n = not specified
    The number of progressive wavelet levels in Eq. (8)-(9) is not stated, affecting the module's capacity and detail recovery.
axioms (5)
  • domain assumption Hessian eigenvalue analysis correctly distinguishes tubular vessels from background in low-contrast ICA images.
    The CASE module relies on Frangi-style vesselness, assuming bright, elongated vessels on darker background; this may fail for overlapping, faint, or tortuous vessels.
  • standard math Haar wavelet transforms are linearly additive, so high-frequency subbands can be summed before inverse transformation.
    This is a standard wavelet property used in Eq. (9), though the network's learned depthwise convolutions on subbands may introduce aliasing not captured by the linear model.
  • domain assumption Mamba2's SSD chunked parallel scan captures long-range spatial dependencies when sequences are flattened along width and height.
    The core assumption of AA-DS Mamba2; the paper shows Grad-CAM heatmaps but provides no theoretical or empirical proof that bidirectional axial scanning is superior to other scan orders for vascular topology.
  • domain assumption Ground-truth annotations by interventional cardiologists are consistent and correct for both segmentation and stenosis severity.
    The paper itself notes in Section 5.3 that manual ICA interpretation has high inter-observer variability, yet the ground truth is treated as noise-free.
  • domain assumption The stenosis detection algorithm from [39] is a valid and unbiased clinical metric when applied to predicted contours.
    The stenosis evaluation depends entirely on this external algorithm; its own failure modes, such as centerline extraction and reference diameter estimation, are inherited by every model tested.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SFD-Mamba2Net: Structure-Guided Frequency-Enhanced Dual-Stream Mamba2 Network for Coronary Artery Segmentation." pith.science (2026). https://pith.science/paper/D35LJLHZ

@misc{pith2026250908934,
  author       = {Pith},
  title        = {Pith review of: SFD-Mamba2Net: Structure-Guided Frequency-Enhanced Dual-Stream Mamba2 Network for Coronary Artery Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D35LJLHZ}},
  note         = {Machine review of arXiv:2509.08934}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Background: Coronary Artery Disease (CAD) is one of the leading causes of death worldwide. Invasive Coronary Angiography (ICA), regarded as the gold standard for CAD diagnosis, necessitates precise vessel segmentation and stenosis detection. However, ICA images are typically characterized by low contrast, high noise levels, and complex, fine-grained vascular structures, which pose significant challenges to the clinical adoption of existing segmentation and detection methods. Objective: This study aims to improve the accuracy of coronary artery segmentation and stenosis detection in ICA images by integrating multi-scale structural priors, state-space-based long-range dependency modeling, and frequency-domain detail enhancement strategies. Methods: We propose SFD-Mamba2Net, an end-to-end framework tailored for ICA-based vascular segmentation and stenosis detection. In the encoder, a Curvature-Aware Structural Enhancement (CASE) module is embedded to leverage multi-scale responses for highlighting slender tubular vascular structures, suppressing background interference, and directing attention toward vascular regions. In the decoder, we introduce a Progressive High-Frequency Perception (PHFP) module that employs multi-level wavelet decomposition to progressively refine high-frequency details while integrating low-frequency global structures. Results and Conclusions: SFD-Mamba2Net consistently outperformed state-of-the-art methods across eight segmentation metrics, and achieved the highest true positive rate and positive predictive value in stenosis detection.

Figures

Figures reproduced from arXiv: 2509.08934 by Chen Zhao, Jingfeng Jiang, Nan Mu, Ruiqi Song, Zhihui Xu.

Figure 2
Figure 2. Figure 2: Overview of the proposed SFD-Mamba2Net framework. The architecture adopts an asymmetric encoder￾decoder design, integrating the MASE, AA-DS Mamba2, and PHFP modules to jointly strengthen vascular structure representation, spatial dependency modeling, and boundary reconstruction, ultimately improving segmentation accuracy and enhancing the robustness of stenosis detection. 3.1. Encoder-Decoder Structure As … view at source ↗
Figure 3
Figure 3. Figure 3: Schematic Diagram of the CASE Module. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the vesselness response mechanism. (a) Local zoom-in of vascular regions illustrating the correspondence between Hessian eigenvalues (𝜆2 , 𝜆2 ) and linear structures; (b) Variation trends of 𝑅𝑏 and 𝑆 values across different locations in ICA images [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization demonstrating the effectiveness of the CASE module. (a) Overlay of the input image and ground￾truth labels; (b) and (c) Grad-CAM heatmaps [31] generated by standard convolution and the CASE module, respectively. Red regions indicate strong activations most critical to the decision, yellow and green indicate moderate importance, and blue denotes minimal contribution. It is worth emphasizing th… view at source ↗
Figure 6
Figure 6. Figure 6: Schematic illustration of the proposed AA-DS Mamba2 architecture. As shown in [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Schematic diagram of the proposed DS Mamba2 module. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Structural diagram of the Mamba2 module. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual analysis of the effectiveness of the AA [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Schematic diagram of the PHFP module. Specifically, the PHFP module constructs its filters based on the Haar wavelet, which consists of one low-pass filter ( 𝑓𝐿𝐿 = [1/√2 1/√2] ) and three high-pass filters ( 𝑓𝐿𝐻 = [1/√2 − 1/√2], 𝑓𝐻𝐿 = [ 1/√2 1/√2 −1/√2 −1/√2 ], 𝑓𝐻𝐻 = [ 1/√2 −1/√2 −1/√2 1/√2 ]). For each decoder input feature [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of the effectiveness of PHFP. (a) Overlap of the input image and ground truth; (b) and (c) Grad [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visual comparison of different models’ effectiveness on ICA images. (a) Overlay of the input image and [PITH_FULL_IMAGE:figures/full_fig_p033_13.png] view at source ↗
Figure 18
Figure 18. Figure 18: Stenosis detection results. (a) Original image; (b) Ground [PITH_FULL_IMAGE:figures/full_fig_p045_18.png] view at source ↗

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

43 extracted references · 11 linked inside Pith

  1. [1]

    Okrainec, D.K

    K. Okrainec, D.K. Banerjee, M.J. Eisenberg, Coronary artery disease in the developing world, American heart journal, 148 (2004) 7-15

    2004

  2. [2]

    Z. Li, Y. Zhang, G. Liu, H. Shao, W. Li, X. Tang, A robust coronary artery identification and centerline extraction method in angiographies, Biomedical Signal Processing and Control, 16 (2015) 1-8

    2015

  3. [3]

    Kim, A new computerized measurement approach of carotid artery stenosis on tomographic image sequence, Academic radiology, (2010)

    D.Y. Kim, A new computerized measurement approach of carotid artery stenosis on tomographic image sequence, Academic radiology, (2010)

    2010

  4. [4]

    Otsu, A threshold selection method from gray -level histograms, Automatica, 11 (1975) 23 - 27

    N. Otsu, A threshold selection method from gray -level histograms, Automatica, 11 (1975) 23 - 27

    1975

  5. [5]

    Soille, L.M

    P. Soille, L.M. Vincent, Determining watersheds in digital pictures via flooding simulations, Visual Communications and Image Processing'90: Fifth in a Series, SPIE, 1990, pp. 240-250

    1990

  6. [6]

    Ronneberger, P

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

    2015

  7. [7]

    Bahdanau, K

    D. Bahdanau, K. Cho, Y. Bengio, Neural machine translation by jointly learning to align and translate, arXiv preprint arXiv:1409.0473, (2014)

    Pith/arXiv arXiv 2014

  8. [8]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems, 30 (2017)

    2017

  9. [9]

    J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A.L. Yuille, Y. Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306, (2021)

    Pith/arXiv arXiv 2021

  10. [10]

    A. Gu, K. Goel, C. Ré, Efficiently modeling long sequences with structured state spaces, arXiv preprint arXiv:2111.00396, (2021)

    Pith/arXiv arXiv 2021

  11. [11]

    T. Dao, A. Gu, Transformers are ssms: Generalized models and efficient algorithms through structured state space duality, arXiv preprint arXiv:2405.21060, (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    A. Gu, T. Dao, Mamba: Linear -time sequence modeling with selective state spaces, arXiv preprint arXiv:2312.00752, (2023)

    Pith/arXiv arXiv 2023

  13. [13]

    J. Ruan, J. Li, S. Xiang, Vm -unet: Vision mamba unet for medical image segmentation, arXiv preprint arXiv:2402.02491, (2024)

    Pith/arXiv arXiv 2024

  14. [14]

    Soares, J.J

    J.V. Soares, J.J. Leandro, R.M. Cesar, H.F. Jelinek, M.J. Cree, Retinal vessel segmentation using the 2-D Gabor wavelet and supervised classification, IEEE Transactions on medical Imaging, 25 (2006) 1214-1222

    2006

  15. [15]

    Huang, X.-H

    D.-X. Huang, X.-H. Zhou, X.-L. Xie, S.-Q. Liu, S.-Y. Wang, Z.-Q. Feng, M.-J. Gui, H. Li, T.- Y. Xiang, B. -X. Yao, Spironet: Spatial -frequency learning and topological channel interaction network for vessel segmentation, arXiv preprint arXiv:2406.19749, (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    W. Wang, J. Wang, C. Chen, J. Jiao, L. Sun, Y. Cai, S. Song, J. Li, Fremae: Fourier transform meets masked autoencoders for medical image segmentation, arXiv preprint arXiv:2304.10864, 4 (2023)

    Pith/arXiv arXiv 2023

  17. [17]

    Frangi, W.J

    A.F. Frangi, W.J. Niessen, K.L. Vincken, M.A. Viergever, Multiscale vessel enhancement filtering, International conference on medical image computing and computer -assisted intervention, Springer, 1998, pp. 130-137

    1998

  18. [18]

    LeCun, L

    Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient -based learning applied to document recognition, Proceedings of the IEEE, 86 (2002) 2278-2324

    2002

  19. [19]

    H. Wang, S. Xie, L. Lin, Y. Iwamoto, X.-H. Han, Y.-W. Chen, R. Tong, Mixed transformer u- net for medical image segmentation, ICASSP 2022 -2022 IEEE international conference on acoustics, speech and signal processing (ICASSP), IEEE, 2022, pp. 2390-2394

    2022

  20. [20]

    J. Hu, S. Chen, Z. Pan, S. Zeng, W. Yang, Perspective+ unet: Enhancing segmentation with bi- path fusion and efficient non-local attention for superior receptive fields, International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 2024, pp. 499-509

    2024

  21. [21]

    M. Lei, H. Wu, X. Lv, X. Wang, Condseg: A general medical image segmentation framework via contrast -driven feature enhancement, Proceedings of the AAAI Conference on Artificial Intelligence, 2025, pp. 4571-4579

    2025

  22. [22]

    Wang, J.-Q

    Z. Wang, J.-Q. Zheng, Y. Zhang, G. Cui, L. Li, Mamba-unet: Unet-like pure visual mamba for medical image segmentation, arXiv preprint arXiv:2402.05079, (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    L. Kong, J. Dong, J. Ge, M. Li, J. Pan, Efficient frequency domain-based transformers for high- quality image deblurring, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 5886-5895

    2023

  24. [24]

    N. Mu, Z. Lyu, X. Zhang, R. McBane, A.S. Pandey, J. Jiang, Exploring a frequency -domain attention-guided cascade U-Net: Towards spatially tunable segmentation of vasculature, Computers in biology and medicine, 167 (2023) 107648

    2023

  25. [25]

    R. Azad, A. Bozorgpour, M. Asadi -Aghbolaghi, D. Merhof, S. Escalera, Deep frequency re - calibration u -net for medical image segmentation, Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3274-3283

    2021

  26. [26]

    L. Chen, L. Gu, L. Li, C. Yan, Y. Fu, Frequency Dynamic Convolution for Dense Image Prediction, Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30178-30188

    2025

  27. [27]

    J. -H. Nam, N.S. Syazwany, S.J. Kim, S. -C. Lee, Modality -agnostic domain generalizable medical image segmentation by multi -frequency in multi -scale attention, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 11480-11491

    2024

  28. [28]

    Ioffe, C

    S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift, International conference on machine learning, pmlr, 2015, pp. 448 -456

    2015

  29. [29]

    Maas, A.Y

    A.L. Maas, A.Y. Hannun, A.Y. Ng, Rectifier nonlinearities improve neural network acoustic models, Proc. icml, Atlanta, GA, 2013, pp. 3

    2013

  30. [30]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: a simple way to prevent neural networks from overfitting, The journal of machine learning research, 15 (2014) 1929-1958

    2014

  31. [31]

    Selvaraju, M

    R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad -cam: Visual explanations from deep networks via gradient -based localization, Proceedings of the IEEE international conference on computer vision, 2017, pp. 618-626

    2017

  32. [32]

    Ramachandran, B

    P. Ramachandran, B. Zoph, Q.V. Le, Searching for activation functions, arXiv preprint arXiv:1710.05941, (2017)

    Pith/arXiv arXiv 2017

  33. [33]

    Zhang, R

    B. Zhang, R. Sennrich, Root mean square layer normalization, Advances in neural information processing systems, 32 (2019)

    2019

  34. [34]

    Mallat, A theory for multiresolution signal decomposition: the wavelet representation, IEEE transactions on pattern analysis and machine intelligence, 11 (2002) 674-693

    S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet representation, IEEE transactions on pattern analysis and machine intelligence, 11 (2002) 674-693

    2002

  35. [35]

    Kingma, J

    D.P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980, (2014)

    Pith/arXiv arXiv 2014

  36. [36]

    Zhou, M.M

    Z. Zhou, M.M. Rahman Siddiquee, N. Tajbakhsh, J. Liang, Unet++: A nested u-net architecture for medical image segmentation, International workshop on deep learning in medical image analysis, Springer, 2018, pp. 3-11

    2018

  37. [37]

    H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin -unet: Unet-like pure transformer for medical image segmentation, European conference on computer vision, Springer, 2022, pp. 205-218

    2022

  38. [38]

    Y. Qi, Y. He, X. Qi, Y. Zhang, G. Yang, Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation, Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 6070-6079

    2023

  39. [39]

    C. Zhao, A. Vij, S. Malhotra, J. Tang, H. Tang, D. Pienta, Z. Xu, W. Zhou, Automatic extraction and stenosis evaluation of coronary arteries in invasive coronary angiograms, Computers in biology and medicine, 136 (2021) 104667

    2021

  40. [40]

    Arbab -Zadeh, J

    A. Arbab -Zadeh, J. Hoe, Quantification of coronary arterial stenoses by multidetector CT angiography in comparison with conventional angiography: methods, caveats, and implications, JACC: Cardiovascular Imaging, 4 (2011) 191-202

    2011

  41. [41]

    Zir, S.W

    L.M. Zir, S.W. Miller, R.E. Dinsmore, J. Gilbert, J. Harthorne, Interobserver variability in coronary angiography, Circulation, 53 (1976) 627-632

    1976

  42. [42]

    Galbraith, M.L

    J.E. Galbraith, M.L. Murphy, N. de Soyza, Coronary angiogram interpretation: interobserver variability, Jama, 240 (1978) 2053-2056

    1978

  43. [43]

    Galassi, M

    F. Galassi, M. Alkhalil, R. Lee, P. Martindale, R.K. Kharbanda, K.M. Channon, V. Grau, R.P. Choudhury, 3D reconstruction of coronary arteries from 2D angiographic projections using non - uniform rational basis splines (NURBS) for accurate modelling of c oronary stenoses, PloS one, 13 (2018) e0190650

    2018

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.