pith. sign in

arxiv: 2606.29744 · v1 · pith:43M653MDnew · submitted 2026-06-29 · 💻 cs.CV

HTC-SGA Former: A Hybrid Transformer-CNN Network with Self-Guided Attention and a New Boundary-Weighted Adaptive Loss for Coronary DSA Vessel Segmentation

Rayan Merghani Ahmed , Marwa Omer Mohammed Omer , Mohamed Elmanna , Shijie Li , Bin Li , Shoujun Zhoua This is my paper

Pith reviewed 2026-06-30 06:39 UTC · model grok-4.3

classification 💻 cs.CV
keywords coronary vessel segmentationDSA imageshybrid Transformer-CNNself-guided attentionboundary-weighted lossmedical image segmentationlightweight model
0
0 comments X

The pith

HTC-SGA Former segments coronary DSA vessels more accurately than 14 prior methods using only 0.81 million parameters.

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

This paper proposes HTC-SGA Former, a hybrid Transformer-CNN network, to segment vessels in coronary digital subtraction angiography images where thin low-contrast structures and severe class imbalance make reliable extraction difficult. It pairs a CNN encoder for local morphology with a Transformer decoder, adding a Multi-Scale Global-Local Window Attention block for contextual modeling, a Self-Guided Feature Attention module to strengthen weak-vessel signals, and a Boundary-Weighted Adaptive Compound Loss to focus on boundaries. The work shows these components together produce better thin-vessel recovery, continuity, and boundary accuracy than existing approaches. A sympathetic reader would care because more precise vessel maps could improve computer-aided diagnosis and intervention planning for coronary artery disease while keeping the model small enough for practical use.

Core claim

HTC-SGA Former employs a CNN encoder for local vessel morphology extraction and a Transformer decoder for contextual feature modeling. A Multi-Scale Global-Local Window Attention block performs efficient global-local contextual modeling, a Self-Guided Feature Attention module enhances weak-vessel responses, and a Boundary-Weighted Adaptive Compound Loss emphasizes thin-vessel boundaries while adaptively balancing recovery and refinement. On private right and left coronary artery DSA subsets this architecture outperforms 14 state-of-the-art segmentation methods while using only 0.81M parameters; the loss also improves results when substituted into four different encoder-decoder backbones.

What carries the argument

Multi-Scale Global-Local Window Attention (MS-GLWA) block combined with Self-Guided Feature Attention (SGFA) module and Boundary-Weighted Adaptive Compound Loss (BWACL), which together supply global-local context, weak-vessel emphasis, and boundary-focused optimization inside the hybrid encoder-decoder.

If this is right

  • Thin distal branches and vessel boundaries become recoverable at higher fidelity than with prior encoder-decoder or pure Transformer models.
  • Vessel continuity improves, reducing fragmentation that currently limits downstream CAD analysis.
  • BWACL can be dropped into other segmentation networks to raise performance without changing their architecture.
  • The 0.81M-parameter footprint makes real-time or edge deployment feasible for clinical DSA workflows.
  • Complementary global-local modeling plus adaptive boundary weighting supports more reliable computer-assisted cardiovascular interventions.

Where Pith is reading between the lines

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

  • The same lightweight hybrid pattern could be tried on other low-contrast tubular structures such as retinal or hepatic vessels where class imbalance is also severe.
  • Public benchmark datasets for coronary segmentation would allow direct comparison and test whether gains persist outside the private subsets used here.
  • If BWACL generalizes across backbones, it could serve as a drop-in replacement for standard losses in any medical segmentation pipeline that must preserve fine boundaries.

Load-bearing premise

The observed performance gains are produced by the MS-GLWA, SGFA, and BWACL components rather than by dataset properties, preprocessing steps, or hyperparameter choices.

What would settle it

An ablation study on the same private datasets that replaces MS-GLWA with standard window attention, removes SGFA, and swaps BWACL for binary cross-entropy plus Dice loss, yet still matches or exceeds the reported scores, would falsify the claim that these three elements drive the improvement.

Figures

Figures reproduced from arXiv: 2606.29744 by Bin Li, Marwa Omer Mohammed Omer, Mohamed Elmanna, Rayan Merghani Ahmed, Shijie Li, Shoujun Zhoua.

Figure 1
Figure 1. Figure 1: Comparison of medical image segmentation architectures: (a) pure CNN encoder-decoder (U-Net (Ronneberger et al., 2015)); (b) hybrid CNN with Transformer bottleneck (e.g., TransUNet (Chen et al., 2021)); (c) pure Transformer encoder-decoder (e.g., SwinUNet (Cao et al., 2022)); and (d) the HTC-SGA Former combining a lightweight CNN encoder with a Transformer decoder. The lower panel compares the standard Tra… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of HTC-SGA Former, comprising a lightweight CNN encoder and a Transformer-based decoder incorporating Efficient WF, MS-GLWA, and SGFA for coronary vessel segmentation. better handle class imbalance, structural continuity, and fine-vessel delineation in difficult medical images, including coronary angiography. Overcoming these limitations is essential for clinically reliable coronary DS… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the MS-GLWA module. The global branch performs enhanced multi-scale QKV generation and window-based self-attention, while the local branch preserves fine spatial details through convolutions. F3 = E3 (Pool(F2 )) ∈ ℝ B×48×H∕4×W∕4 (3) F4 = E4 (Pool(F3 )) ∈ ℝ B×48×H∕8×W∕8 (4) where 𝐸𝑖 denotes the 𝑖-th encoding block, Pool(⋅) denotes max pooling, and 𝐵, 𝐻, and 𝑊 represent the batch size and spa… view at source ↗
Figure 4
Figure 4. Figure 4: Structure of the SGFA module. The module generates a preliminary vessel-guided attention map, extracts multi-scale contextual features through parallel dilated convolutions, and fuses them for final vessel-focused refinement. 3.3.2. Self-guided feature attention module (SGFA) Although MS-GLWA improves global-local contextual modeling, accurate coronary DSA segmentation still requires refinement of weak and… view at source ↗
Figure 5
Figure 5. Figure 5: Recall comparison of different loss functions across U-Net, Attention U-Net, U-Net++, and Attention U-Net++. (a) Right coronary artery subset. (b) Left coronary artery subset. BWACL achieves the highest recall. To further validate the vessel-recovery capability of BWACL, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative segmentation comparison on the right coronary subset using U-Net (blue border) and U-Net++ (orange border) trained with and without BWACL. Yellow rectangles highlight regions where BWACL improves thin-vessel recovery, boundary delineation, and false-positive suppression [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative segmentation comparison on the left coronary subset using Attention U-Net (blue border) and Attention U-Net++ (orange border) trained with and without BWACL. Yellow rectangles highlight regions where BWACL improves thin-vessel recovery, boundary delineation, and false-positive suppression. preserves weak distal branches, reduces fragmented predictions, and enhances boundary delineation across a… view at source ↗
Figure 8
Figure 8. Figure 8: The figure illustrates the trade-off between parameter count (M), Dice coefficient (DSC), and ASD for different segmentation models. Point size and color represent model complexity and ASD, respectively. (a) Right coronary artery subset. (b) Left coronary artery subset [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison on the right coronary artery subset between HTC-SGA Former and state-of-the-art segmentation methods. Yellow rectangles highlight challenging regions with thin distal branches, vessel discontinuities, boundary ambiguity, and false-positive responses. Qualitative comparison with state-of-the-art methods To further evaluate HTC-SGA Former qualitatively, Figs. 9 and 10 present visualiza… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison on the left coronary artery subset between HTC-SGA Former and state-of-the-art segmentation methods. Yellow rectangles highlight challenging regions with thin distal branches, complex bifurcations, boundary localization, and false-positive responses. Compared with competing methods, HTC-SGA Former produces segmentation results closer to the ground truth across both coronary subsets.… view at source ↗
read the original abstract

Accurate coronary Digital Subtraction Angiography (DSA) vessel segmentation is essential for computer-aided diagnosis and treatment planning of coronary artery disease (CAD). However, thin low-contrast vessels, background interference, and severe vessel-background class imbalance make reliable segmentation of weak distal branches and vessel boundaries challenging. Existing methods struggle to balance global contextual reasoning with preservation of weak vessels, vessel continuity, and fine boundaries. To address these limitations, we propose HTC-SGA Former, a lightweight hybrid Transformer-CNN framework for coronary DSA vessel segmentation. It employs a CNN encoder for local vessel morphology extraction and a Transformer decoder for contextual feature modeling. A Multi-Scale Global-Local Window Attention (MS-GLWA) block performs efficient global-local contextual modeling, while a Self-Guided Feature Attention (SGFA) module enhances weak-vessel responses. In addition, a Boundary-Weighted Adaptive Compound Loss (BWACL) emphasizes thin-vessel boundaries and adaptively balances vessel recovery and boundary refinement. Experiments on private right and left coronary artery DSA subsets show that HTC-SGA Former outperforms 14 state-of-the-art segmentation methods while maintaining a compact architecture with only 0.81M parameters. BWACL also improves performance over binary cross-entropy and Dice losses across four encoder-decoder architectures, demonstrating strong cross-backbone applicability. HTC-SGA Former improves thin-vessel recovery, vessel continuity, and boundary localization through complementary global-local contextual modeling, vessel-focused refinement, and adaptive optimization, supporting reliable and computationally efficient coronary vessel analysis for future computer-assisted cardiovascular interventions.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes HTC-SGA Former, a lightweight hybrid Transformer-CNN for coronary DSA vessel segmentation. It uses a CNN encoder and Transformer decoder augmented by a Multi-Scale Global-Local Window Attention (MS-GLWA) block, a Self-Guided Feature Attention (SGFA) module, and a Boundary-Weighted Adaptive Compound Loss (BWACL). The central empirical claims are that the model outperforms 14 state-of-the-art segmentation methods on private right- and left-coronary DSA subsets while using only 0.81 M parameters, and that BWACL improves results over binary cross-entropy and Dice losses across four encoder-decoder backbones.

Significance. If the reported gains prove reproducible and attributable to the proposed modules rather than dataset-specific tuning, the compact architecture and boundary-aware loss could be useful for clinical vessel analysis. The absence of any public-dataset validation or statistical rigor, however, prevents assessment of generalizability, so the practical significance remains limited.

major comments (2)
  1. [Experiments] Experiments section: all quantitative comparisons (Tables reporting Dice, sensitivity, etc., versus 14 SOTA methods) and all ablation results for MS-GLWA, SGFA, and BWACL are confined to two private right/left coronary DSA subsets. No acquisition parameters, annotation protocol, train/val/test splits, or preprocessing details are supplied that would permit reproduction, and no experiments on any public coronary or vessel dataset appear. This directly blocks verification of the outperformance and component-attribution claims that constitute the paper's central contribution.
  2. [Results] Results and ablation subsections: the abstract asserts that HTC-SGA Former and BWACL outperform baselines, yet no statistical significance tests, standard deviations, or error bars are reported, nor is any protocol described that guarantees identical hyper-parameters, data augmentation, and optimization settings across all compared methods and backbones. Without these, the attribution of gains to the proposed modules cannot be isolated from confounding factors.
minor comments (1)
  1. [Abstract] Abstract: the 0.81 M parameter count is stated without any corresponding parameter counts for the 14 reference methods, making the compactness claim difficult to evaluate.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for improved reproducibility and statistical rigor. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: all quantitative comparisons (Tables reporting Dice, sensitivity, etc., versus 14 SOTA methods) and all ablation results for MS-GLWA, SGFA, and BWACL are confined to two private right/left coronary DSA subsets. No acquisition parameters, annotation protocol, train/val/test splits, or preprocessing details are supplied that would permit reproduction, and no experiments on any public coronary or vessel dataset appear. This directly blocks verification of the outperformance and component-attribution claims that constitute the paper's central contribution.

    Authors: We agree that additional experimental details are required. In the revised manuscript we will add a dedicated subsection with full acquisition parameters (e.g., imaging system, contrast protocol, resolution), annotation protocol (number of experts, inter-observer agreement), exact train/val/test split sizes and selection criteria, and all preprocessing steps. Because the data are private clinical DSA images governed by patient privacy regulations, public release is not possible; we will state this limitation explicitly. We will also explore inclusion of at least one public vessel segmentation dataset for supplementary validation. revision: partial

  2. Referee: [Results] Results and ablation subsections: the abstract asserts that HTC-SGA Former and BWACL outperform baselines, yet no statistical significance tests, standard deviations, or error bars are reported, nor is any protocol described that guarantees identical hyper-parameters, data augmentation, and optimization settings across all compared methods and backbones. Without these, the attribution of gains to the proposed modules cannot be isolated from confounding factors.

    Authors: We concur that statistical controls are essential. The revised manuscript will report standard deviations from repeated experiments (different random seeds), include paired statistical significance tests (e.g., Wilcoxon signed-rank) between HTC-SGA Former and each baseline, and add error bars to all quantitative tables and figures. We will also document the hyper-parameter search procedure and confirm that every compared method and backbone was trained and evaluated under identical data-augmentation and optimization protocols. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical architecture proposal with no derivation chain

full rationale

The paper introduces HTC-SGA Former (CNN encoder + Transformer decoder with MS-GLWA and SGFA modules) plus BWACL loss, then reports experimental metrics on two private DSA datasets. No equations derive predictions from inputs, no fitted parameters are relabeled as predictions, and no self-citation chain supports a uniqueness claim. All performance assertions rest on direct comparisons to 14 baselines and cross-backbone ablations; the derivation is therefore self-contained and contains no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on three newly introduced modules whose independent value is asserted via experiments on private data; no free parameters are explicitly fitted beyond standard training, and no new physical or mathematical entities are postulated.

axioms (1)
  • domain assumption Supervised learning on pixel-wise annotated DSA images is a valid way to train vessel segmentation models.
    Implicit foundation of all supervised segmentation papers; invoked by the experimental setup described in the abstract.
invented entities (3)
  • MS-GLWA block no independent evidence
    purpose: Efficient global-local contextual modeling
    New attention block proposed to address limitations of existing methods.
  • SGFA module no independent evidence
    purpose: Enhance weak-vessel responses
    New attention module introduced for thin low-contrast vessels.
  • BWACL loss no independent evidence
    purpose: Emphasize thin-vessel boundaries and adaptively balance classes
    New compound loss function proposed to improve boundary localization and vessel recovery.

pith-pipeline@v0.9.1-grok · 5845 in / 1472 out tokens · 31020 ms · 2026-06-30T06:39:51.154651+00:00 · methodology

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

37 extracted references · 37 canonical work pages · 2 internal anchors

  1. [1]

    Biomedical Signal Processing and Control 123, 110539

    Msa-unet3+: Multi-scale attention unet3+ with new supervised prototypical contrastive loss for coronary dsa image segmentation. Biomedical Signal Processing and Control 123, 110539. doi:10.1016/j.bspc.2026.110539. Algarni, M., Al-Rezqi, A., Saeed, F., Alsaeedi, A., Ghabban, F.,

    work page doi:10.1016/j.bspc.2026.110539 2026

  2. [2]

    PeerJ Computer Science 8, e993

    Multi-constraints based deep learning model for automated segmentation and diagnosis of coronary artery disease in x-ray angiographic images. PeerJ Computer Science 8, e993. doi:10.7717/peerj-cs.993. Cao,H.,Wang,Y.,Chen,J.,Jiang,D.,Zhang,X.,Tian,Q.,Wang,M.,2022. Swin-unet:Unet-likepuretransformerformedicalimagesegmentation, in: European conference on comp...

    work page doi:10.7717/peerj-cs.993 2022

  3. [3]

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.,

    doi:10.1038/s41598-024-57198-5. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.,

    work page doi:10.1038/s41598-024-57198-5

  4. [4]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

    Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 doi:10.48550/arXiv.2102.04306. Rayan Merghani Ahmed et al.:Preprint submitted to ElsevierPage 18 of 20 HTC-SGA Former Deng, H., Li, Y., Liu, X., Cheng, K., Fang, T., Min, X.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.04306

  5. [5]

    Medical Physics 52, 3135–3150

    Multi-scale dual attention embedded u-shaped network for accurate segmentation of coronary vessels in digital subtraction angiography. Medical Physics 52, 3135–3150. doi:10.1002/mp.17618. Deng, H., Liu, X., Fang, T., Li, Y., Min, X.,

    work page doi:10.1002/mp.17618

  6. [6]

    Frąk, W., Wojtasińska, A., Lisińska, W., Młynarska, E., Franczyk, B., Rysz, J.,

    doi:10.1186/s40537-024-00904-x. Frąk, W., Wojtasińska, A., Lisińska, W., Młynarska, E., Franczyk, B., Rysz, J.,

    work page doi:10.1186/s40537-024-00904-x

  7. [7]

    Fu,B.,Peng,Y.,He,J.,Tian,C.,Sun,X.,Wang,R.,2024a

    doi:10.3390/ biomedicines10081938. Fu,B.,Peng,Y.,He,J.,Tian,C.,Sun,X.,Wang,R.,2024a. Hmsu-net:Ahybridmulti-scaleu-netbasedonacnnandtransformerformedicalimage segmentation. Computers in Biology and Medicine 170, 108013. doi:10.1016/j.compbiomed.2024.108013. Fu, Z., Fu, Z., Lu, C., Yan, J., Fei, J., Han, H., 2024b. Robust implementation of foreground extrac...

    work page doi:10.1016/j.compbiomed.2024.108013 2024

  8. [8]

    Measurement 222, 113553

    A multi-scale global attention network for blood vessel segmentation from fundus images. Measurement 222, 113553. doi:10.1016/j.measurement.2023.113553. Gao,Y.,Wang,Y.,Ai,D.,Shang,F.,Song,H.,Fan,J.,Xiao,D.,Wang,Y.,Yang,J.,2026. Iterativejointlearningintegratingtemporalandgeometric information for vessel segmentation in x-ray coronary angiography. Medical ...

    work page doi:10.1016/j.measurement.2023.113553 2023

  9. [9]

    IEEE Transactions on Medical Imaging 42, 2763–2775

    H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medical Imaging 42, 2763–2775. doi:10.1109/tmi.2023.3264513. He,K.,Zhang,X.,Ren,S.,Sun,J.,2016. Deepresiduallearningforimagerecognition,in:ProceedingsoftheIEEEconferenceoncomputervision and pattern recognition, pp. 770–778. doi:10.1109/cvpr.2016.90....

    work page doi:10.1109/tmi.2023.3264513 2023

  10. [10]

    Medical Physics 52, e17970

    Deep learning model for coronary artery segmentation and quantitative stenosis detection in angiographic images. Medical Physics 52, e17970. doi:10.1002/mp.17970. Huang, X., Deng, Z., Li, D., Yuan, X., Fu, Y.,

    work page doi:10.1002/mp.17970

  11. [11]

    IEEE transactions on medical imaging 42, 1484–1494

    Missformer: An effective transformer for 2d medical image segmentation. IEEE transactions on medical imaging 42, 1484–1494. doi:10.1109/TMI.2022.3230943. Huang, X., Gong, H., Zhang, J.,

    work page doi:10.1109/tmi.2022.3230943 2022

  12. [12]

    IEEE Journal of Biomedical and Health Informatics 28, 4048–4061

    Hst-mrf: heterogeneous swin transformer with multi-receptive field for medical image segmentation. IEEE Journal of Biomedical and Health Informatics 28, 4048–4061. doi:10.1109/jbhi.2024.3397047. Li,C.,Tan,Y.,Chen,W.,Luo,X.,Gao,Y.,Jia,X.,Wang,Z.,2020. Attentionunet++:Anestedattention-awareu-netforliverctimagesegmentation, in: 2020 IEEE international confer...

    work page doi:10.1109/jbhi.2024.3397047 2024

  13. [13]

    Computer Methods and Programs in Biomedicine 233, 107493

    Mestrans: Multi-scale embedding spatial transformer for medical image segmentation. Computer Methods and Programs in Biomedicine 233, 107493. doi:10.1016/j.cmpb.2023.107493. Liu,Z.,Lin,Y.,Cao,Y.,Hu,H.,Wei,Y.,Zhang,Z.,Lin,S.,Guo,B.,2021. Swintransformer:Hierarchicalvisiontransformerusingshiftedwindows, in: Proceedings of the IEEE/CVF international conferen...

    work page doi:10.1016/j.cmpb.2023.107493 2023

  14. [14]

    Biomedical Signal Processing and Control 101, 107175

    Segmentation of coronary arteries from x-ray angiographic images using density based spatial clustering of applications with noise (dbscan). Biomedical Signal Processing and Control 101, 107175. doi:10.1016/j.bspc.2024.107175. Milletari, F., Navab, N., Ahmadi, S.A.,

    work page doi:10.1016/j.bspc.2024.107175 2024

  15. [15]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation, in: 2016 fourth international conference on 3D vision (3DV), Ieee. pp. 565–571. doi:10.1109/3dv.2016.79. Moccia, S., De Momi, E., El Hadji, S., Mattos, L.S.,

    work page doi:10.1109/3dv.2016.79 2016

  16. [16]

    Computer methods and programs in biomedicine 158, 71–91

    Blood vessel segmentation algorithms—review of methods, datasets and evaluation metrics. Computer methods and programs in biomedicine 158, 71–91. doi:10.1016/j.cmpb.2018.02.001. Molenaar, M.A., Hebbo, E., Selder, J.L., Shekiladze, N., Sandesara, P.B., Nicholson, W.J., Asselbergs, F.W., Ahmad, S., Gold, D.A., Sakr, S.M., et al.,

    work page doi:10.1016/j.cmpb.2018.02.001 2018

  17. [17]

    JACC: Advances 4, 102360

    Deep learning–based segmentation of coronary arteries and stenosis detection in x-ray coronary angiography. JACC: Advances 4, 102360. doi:10.1016/j.jacadv.2025.102360. NobreMenezes,M.,Silva,J.L.,Silva,B.,Rodrigues,T.,Guerreiro,C.,Guedes,J.P.,Santos,M.O.,Oliveira,A.L.,Pinto,F.J.,2023. Coronaryx-ray angiography segmentation using artificial intelligence: a ...

    work page doi:10.1016/j.jacadv.2025.102360 2025

  18. [18]

    Attention U-Net: Learning Where to Look for the Pancreas

    Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 doi:10.48550/arXiv.1804.03999. Pagliaro, B.R., Cannata, F., Stefanini, G.G., Bolognese, L.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1804.03999

  19. [19]

    Heart failure reviews 25, 53–65

    Myocardial ischemia and coronary disease in heart failure. Heart failure reviews 25, 53–65. doi:10.1007/s10741-019-09831-z. Pan, L.S., Li, C.W., Su, S.F., Tay, S.Y., Tran, Q.V., Chan, W.P.,

    work page doi:10.1007/s10741-019-09831-z

  20. [20]

    Scientific reports 11, 14493

    Coronary artery segmentation under class imbalance using a u-net based architecture on computed tomography angiography images. Scientific reports 11, 14493. doi:10.1038/s41598-021-93889-z. Park, J., Kweon, J., Kim, Y.I., Back, I., Chae, J., Roh, J.H., Kang, D.Y., Lee, P.H., Ahn, J.M., Kang, S.J., et al.,

    work page doi:10.1038/s41598-021-93889-z

  21. [21]

    Medical physics 50, 7822–7839

    Selective ensemble methods for deep learning segmentation of major vessels in invasive coronary angiography. Medical physics 50, 7822–7839. doi:10.1002/mp.16554. Peng,J.,Wang,P.,Pedersoli,M.,Desrosiers,C.,2024. Boundary-awareinformationmaximizationforself-supervisedmedicalimagesegmentation. Medical Image Analysis 94, 103150. doi:10.1016/j.media.2024.10315...

    work page doi:10.1002/mp.16554 2024

  22. [22]

    Computers in Biology and Medicine 156, 106493

    Semi-supervised segmentation of coronary dsa using mixed networks and multi-strategies. Computers in Biology and Medicine 156, 106493. doi:10.1016/j.compbiomed.2022.106493. Qiu, M., Zhang, C., Song, Z.,

    work page doi:10.1016/j.compbiomed.2022.106493 2022

  23. [23]

    Medical physics 49, 1739–1753

    Dynamic boundary-insensitive loss for magnetic resonance medical image segmentation. Medical physics 49, 1739–1753. doi:10.1002/mp.15386. Ronneberger, O., Fischer, P., Brox, T.,

    work page doi:10.1002/mp.15386

  24. [24]

    U-net: Convolutional networks for biomedical image segmentation, in: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Springer. pp. 234–241. doi:10.1007/978-3-319-24574-4_28. Samuel, P.M., Veeramalai, T.,

    work page doi:10.1007/978-3-319-24574-4_28 2015

  25. [25]

    Computer methods and programs in biomedicine 198, 105769

    Vssc net: vessel specific skip chain convolutional network for blood vessel segmentation. Computer methods and programs in biomedicine 198, 105769. doi:10.1016/j.cmpb.2020.105769. Shen, Y., Chen, Z., Tong, J., Jiang, N., Ning, Y.,

    work page doi:10.1016/j.cmpb.2020.105769 2020

  26. [26]

    The International Journal of Cardiovascular Imaging 39, 1571–1579

    Dbcu-net: deep learning approach for segmentation of coronary angiography images. The International Journal of Cardiovascular Imaging 39, 1571–1579. doi:10.1007/s10554-023-02849-3. Tang, F., Ding, J., Quan, Q., Wang, L., Ning, C., Zhou, S.K.,

    work page doi:10.1007/s10554-023-02849-3

  27. [27]

    Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion, in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), IEEE. pp. 1–5. doi:10.1109/isbi56570. 2024.10635609. Tang, F., Wang, L., Ning, C., Xian, M., Ding, J., 2023a. Cmu-net: a strong convmixer-based medical ultrasound image segmentation network, in...

    work page doi:10.1109/isbi56570 2024

  28. [28]

    Physics in Medicine & Biology 69, 025012

    Coronary vessel segmentation in coronary angiography with a multi- scale u-shaped transformer incorporating boundary aggregation and topology preservation. Physics in Medicine & Biology 69, 025012. doi:10.1088/1361-6560/ad0b63. Wang, H., Qi, Y., Liu, W., Guo, K., Lv, W., Liang, Z.,

    work page doi:10.1088/1361-6560/ad0b63

  29. [29]

    IEEE Journal of Biomedical and Health Informatics doi:10.1109/jbhi.2025.3601025

    Dpgnet: A boundary-aware medical image segmentation framework via uncertainty perception. IEEE Journal of Biomedical and Health Informatics doi:10.1109/jbhi.2025.3601025. Wang, R., Chen, S., Ji, C., Fan, J., Li, Y.,

    work page doi:10.1109/jbhi.2025.3601025 2025

  30. [30]

    Medical image analysis 78, 102395

    Boundary-aware context neural network for medical image segmentation. Medical image analysis 78, 102395. doi:10.1016/j.media.2022.102395. Woo, S., Park, J., Lee, J.Y., Kweon, I.S.,

    work page doi:10.1016/j.media.2022.102395 2022

  31. [31]

    Cbam: Convolutional block attention module, in: Proceedings of the European conference on computer vision (ECCV), pp. 3–19. doi:10.1007/978-3-030-01234-2_1. Wu, H., Min, W., Gai, D., Huang, Z., Geng, Y., Wang, Q., Chen, R.,

    work page doi:10.1007/978-3-030-01234-2_1

  32. [32]

    Computers in Biology and Medicine 178, 108671

    Hd-former: A hierarchical dependency transformer for medical image segmentation. Computers in Biology and Medicine 178, 108671. doi:10.1016/j.compbiomed.2024.108671. Wu, W., Dai, T., Chen, Z., Huang, X., Xiao, J., Ma, F., Ouyang, R.,

    work page doi:10.1016/j.compbiomed.2024.108671 2024

  33. [33]

    Engineering Applications of Artificial Intelligence 139, 109626

    Adaptive patch contrast for weakly supervised semantic segmentation. Engineering Applications of Artificial Intelligence 139, 109626. doi:10.1016/j.engappai.2024.109626. Xia, C., Rook, M., Pelgrim, G.J., Sidorenkov, G., Wisselink, H.J., Van Bolhuis, J.N., van Ooijen, P.M., Guo, J., Oudkerk, M., Groen, H., et al.,

    work page doi:10.1016/j.engappai.2024.109626 2024

  34. [34]

    xia et al

    Early imaging biomarkers of lung cancer, copd and coronary artery disease in the general population: rationale and design of the imalife (imaging in lifelines) study: C. xia et al. European journal of epidemiology 35, 75–86. doi:10.1007/s10654-019-00519-0. Xie, X., Zhang, W., Pan, X., Xie, L., Shao, F., Zhao, W., An, J.,

    work page doi:10.1007/s10654-019-00519-0

  35. [35]

    Biomedical Signal Processing and Control 81, 104437

    Canet: Context aware network with dual-stream pyramid for medical image segmentation. Biomedical Signal Processing and Control 81, 104437. doi:10.1016/j.bspc.2022.104437. Yang, S., Kweon, J., Roh, J.H., Lee, J.H., Kang, H., Park, L.J., Kim, D.J., Yang, H., Hur, J., Kang, D.Y., et al.,

    work page doi:10.1016/j.bspc.2022.104437 2022

  36. [36]

    Scientific reports 9, 16897

    Deep learning segmentation of major vessels in x-ray coronary angiography. Scientific reports 9, 16897. doi:10.1038/s41598-019-53254-7. Zeng,Y.,Liu,X.,Xiao,N.,Li,Y.,Jiang,Y.,Feng,J.,Guo,S.,2019. Automaticdiagnosisbasedonspatialinformationfusionfeatureforintracranial aneurysm. IEEE transactions on medical imaging 39, 1448–1458. doi:10.1109/tmi.2019.2951439...

    work page doi:10.1038/s41598-019-53254-7 2019

  37. [37]

    arXiv (Cornell University) URL:https://arxiv.org/abs/1912.05074

    Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. arXiv (Cornell University) URL:https://arxiv.org/abs/1912.05074. Rayan Merghani Ahmed et al.:Preprint submitted to ElsevierPage 20 of 20

    work page arXiv 1912