Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Deep Self-knowledge Distillation: A hierarchical supervised learning for coronary artery segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a segmentation network can improve itself by distilling, at every decoder stage, the patch-level foreground/background distributions of its own previous-epoch predictions, and that this hierarchical self-distillation

desk verdict The paper's central new loss term is broken as written; the empirical claims cannot be trusted until it is fixed. read the letter →

arxiv 2509.03173 v1 pith:SKOQGMIK submitted 2025-09-03 eess.IV cs.CVcs.LG

Mingfeng Lin This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords coronaryarterysegmentationself-knowledgedistillationdeepsupervisionX-rayangiographyknowledgemedicalimageU-Nethierarchical
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

This paper aims to show that an encoder-decoder segmentation network can be made more accurate for coronary artery segmentation by teaching itself from its own earlier training state, and by supervising every decoder depth rather than only the final output. The proposed Deep Self-knowledge Distillation adds two terms to the standard dice loss: a Deep Distribution Loss that matches patch-level foreground/background distributions between the student's and the previous epoch teacher's side outputs, and a Pixel-wise Self-knowledge Distillation Loss that blends teacher output with ground truth into a soft pixel target. On the XCAD and DCA1 angiography datasets, the method outperforms nine comparison models on Dice, accuracy, sensitivity, and IoU, and ablations show both loss components contribute. If the claim holds, it gives a training-only route to better vessel segmentation—no extra inference cost, no external teacher, and a principled use of the hierarchical knowledge already inside the network.

What carries the argument

The mechanism is hierarchical self-distillation with two cooperating constraints. Each decoder side output—the prediction-like map formed after upsampling a decoder feature map to input resolution—is divided into patches; counting foreground and background pixels per patch gives a distribution vector, which is softmax-smoothed with a temperature and matched between student and teacher by KL divergence (the Deep Distribution Loss). In parallel, the teacher's final output is linearly combined with ground truth with weight α to form a softened label, and the student's final output is supervised against it by cross entropy (the Pixel-wise Self-knowledge Distillation Loss), with α increasing line

What would settle it

Feed a real training image through the method and compute the distribution vector for a set of patches: because the formula multiplies the indicator flags across all pixels in a patch, any mixed patch containing both vessel and background yields a (0,0) vector. If most patches are mixed, the KL term is comparing almost empty vectors and cannot be the mechanism behind the reported gains; replacing the product with an actual count would settle the claim.

Watch

Extended reading notes

Core claim

The paper proposes Deep Self-knowledge Distillation as a training-only supervision scheme for encoder-decoder segmentation networks. At each decoder depth, the side output is turned into a probabilistic distribution vector by counting foreground and background pixels in patches; during training the student's vectors are pulled toward the teacher's—the model from the previous epoch—via KL divergence, forming the Deep Distribution Loss. In parallel, the teacher's final output is linearly blended with the ground truth to form a soft target that supervises the student pixel-wise, forming the Pixel-wise Self-knowledge Distillation Loss. On U-Net3+ over XCAD and DCA1, the combined objective raises

Load-bearing premise

The main premise is that the model from the previous epoch is a reliable teacher and that patch-level foreground/background counts extracted from its decoder side outputs carry hierarchical knowledge worth transferring to the current model.

Editorial extensions

If this is right

  • Adopting the two losses on a U-Net3+ baseline raises DSC by 2.60 percentage points on XCAD and 4.24 points on DCA1, with the combined loss outperforming either loss alone.
  • The gains come at zero extra inference cost, since the teacher and side-output branches are used only during training and discarded afterward.
  • The method works on any encoder-decoder segmentation network with accessible decoder side outputs, so the same hierarchical recipe can be dropped onto other U-Net-style backbones.
  • The loose distribution-level supervision at side outputs combined with tight pixel-level supervision at the final output jointly improve generalization; ablations identify τ=3, n=16, and α_T=0.5 as the best settings.
  • On both datasets the model surpasses nine comparison methods across all four reported metrics: DSC, ACC, SEN, and IOU.

Reading between the lines

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

  • The distribution-vector definition multiplies indicators inside each patch, so mixed vessel/background patches collapse to a (0,0) vector; using an actual count would preserve information in mixed patches and may strengthen the Deep Distribution Loss—an inexpensive modification to test.
  • Because the losses are defined on generic decoder side outputs and patch statistics rather than coronary-specific features, the recipe should transfer to other thin-structure segmentation tasks such as retinal vessel, airway, or road extraction.
  • The α schedule is linear by fiat; replacing it with a validation-gated schedule could make the softened target more reliable when teacher quality plateaus or dips.
  • A control experiment with full-resolution KL divergence (no patch aggregation) would isolate whether the reported robustness comes from the coarseness of the distribution supervision or from distribution matching itself.
Share X Bluesky LinkedIn Reddit HN

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 paper proposes Deep Self-knowledge Distillation (DSKD) for coronary artery segmentation in X-ray angiography. The method augments a U-Net3+ baseline with two auxiliary losses: a Deep Distribution Loss (LDDL) that compares probabilistic distribution vectors derived from decoder side outputs of the student and teacher models via KL divergence, and a Pixel-wise Self-knowledge Distillation Loss (LPSDL) that supervises the final output against a linear combination of the teacher prediction and ground truth. Experiments on the XCAD and DCA1 datasets report improved DSC, ACC, SEN, and IOU over several U-Net variants and transformer-based methods, and ablations claim both loss components contribute.

Significance. If the proposed hierarchical distillation mechanism were sound, the paper would offer a practical way to exploit multi-scale decoder features without additional inference cost, which is valuable for medical image segmentation. The manuscript includes a broad comparison with nine baselines and ablation studies on all three introduced hyperparameters. However, the central methodological novelty (LDDL) is, as written, not a functioning loss: Eq. (2) defines a degenerate and non-differentiable quantity that cannot provide a learning signal, yet Table 1 reports clear improvements from adding it. This internal contradiction, together with test-set hyperparameter selection and the absence of error bars or significance tests, means the current empirical claims are not supported. The paper therefore requires substantial corrections and re-analysis before the results can be trusted.

major comments (3)
  1. [§3.2.1, Eq. (2)] The formula defines each entry of the patch distribution vector as a product of indicator functions, not as a count as stated in the text ('count the number of 0 and 1 in each patch'). The side outputs in Eq. (1) are sigmoid activations in (0,1), so the indicators 1(·=1) and 1(·=0) are zero for essentially every pixel; each patch vector is [0,0], softmax yields a uniform vector, and the KL divergence in Eq. (4) is identically zero. Even after thresholding, products of step functions are non-differentiable and give zero gradients almost everywhere. This makes it impossible for the LDDL component to contribute to training, yet Table 1 reports that adding LDDL improves DSC. Please correct Eq. (2) (e.g., sums/counts), define a differentiable soft-bin mapping, and re-run the ablations; the current central claim is unsupported as written.
  2. [§4.3.2–§4.3.4, Tables 2–4] The values τ=3, n=16, and α_T=0.5 are selected by maximizing DSC on the DCA1 evaluation split, and then the DCA1 row of Table 1 is reported with those selected values. This is test-set selection and invalidates the reported DCA1 gains as an unbiased estimate. Use a separate validation split for tuning or nested cross-validation, and report XCAD results under the same protocol. In addition, Table 1 gives single-run metrics without standard deviations or significance tests, so the summary claim of 'outperforming' is not statistically supported.
  3. [§3.2.1 and §4.3.3] The patch geometry is not well-defined. Equation (2) uses patch size s, while the text sets n = H·W/s². Table 3 reports n=16 (4×4) as optimal; for 512×512 inputs this implies s=128, but for the 300×300 DCA1 images the relationship is not satisfied. Specify exactly how n and s are computed for each dataset, including any padding or overlapping patches, since the loss depends on this mapping. Also note that Eq. (1) produces continuous sigmoid outputs, so the assertion in §3.2.1 that '˜y^i takes values in 0 and 1' is inconsistent with the model.
minor comments (5)
  1. [Abstract] The phrase 'outperforms the dice coefficient, accuracy, sensitivity and IoU' should be reworded to 'outperforms in terms of DSC, ACC, SEN, and IoU'.
  2. [Tables 2–4] The word 'bond' appears in captions; it should be 'bold'.
  3. [Algorithm 1] L_total = Φ uses the empty-set symbol; use 0 or an empty loss container. Also, 'MTeacher= Φ' should be 'MTeacher = None' or similar.
  4. [Figure 3] The caption says 'Quantitative analysis' but the figure shows visual segmentation results; this should be 'Qualitative comparison'.
  5. [§4.3.1] The notation 'LP SKL' in the text appears to be a typo for 'L_PSDL' or 'LP SDL'.

Circularity Check

2 steps flagged · score 6.0 of 10

LDDL is definitionally inert and DCA1 hyperparameters are fitted then reported; partial circularity.

  1. fitted input called prediction [Section 4.3.2–4.3.4, Tables 2–4 and Table 1 (DCA1)]
    "The ablation study results for the temperature scaling factor τ in the DCA1 dataset are presented in Table 2. ... we found that the optimal performance of the model occurred when τ = 3. Consequently, for the experiments mentioned in Table 1, we set the temperature scaling factor τ to 3."

    The same DCA1 dataset is used both to select τ, n, and α (Tables 2–4, by reported DSC/ACC/SEN/IOU) and to report the final 'U-Net3+ w/ Ltotal' DCA1 row in Table 1. No separate validation split or held-out selection is described. Thus the DCA1 comparison is partly a fitted optimum, not an independent prediction: the hyperparameters are chosen to maximize the very metrics then presented as the method's DCA1 result.

  2. other [Eq. (1)–(4), Table 1 rows 'U-Net3+ w/ LDDL']
    "the side outputs ˙X i De is defined as: ˙X i De = σ(B(f (X i De); i − 1)) (1) ... ˜Z i[m] = [ ∏_{j=1}^{s} ∏_{k=1}^{s} 1(˜yi,m[j][k] = 1), ∏_{j=1}^{s} ∏_{k=1}^{s} 1(˜yi,m[j][k] = 0)] (2)"

    By Eq. (1), side outputs are sigmoid activations, so for finite logits no value is exactly 0 or 1; both indicator products in Eq. (2) are 0 for every pixel, making each patch vector [0,0]. The softmax in Eq. (3) then yields the same uniform distribution for student and teacher, and the KL divergence in Eq. (4) is identically zero. If instead the sigmoid outputs were thresholded to 0/1, the indicator step functions would be non-differentiable, again giving no gradient. Therefore LDDL cannot change training, and the DSC improvements attributed to 'U-Net3+ w/ LDDL' in Table 1 are not derivable from the loss as defined; the claimed effect is zero by construction.

full rationale

The paper contains no load-bearing self-citations, and the XCAD benchmark provides an external, transferable check, so the work is not wholly circular. However, two by-construction reductions are present. First, the Deep Distribution Loss is defined through hard indicator products on sigmoid side outputs: because σ outputs are not exactly 0/1, every patch vector is [0,0], the softmax is the same for teacher and student, and the KL loss is identically zero (and non-differentiable). The Table 1 gains attributed to LDDL are therefore not explained by the defined loss. Second, the DCA1 results in Table 1 use hyperparameters (τ=3, n=16, α=0.5) that were selected on DCA1 in Tables 2–4; the paper does not describe a separate validation selection, so the DCA1 comparison is at least partly a fitted optimum rather than an independent prediction. The pixel-wise self-distillation and dice-loss components remain substantive, and XCAD gives independent evidence, so the circularity is partial rather than total.

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

The central claim rests on unproven modeling assumptions about what the teacher model knows and what patch distributions represent. No independent evidence is given for these assumptions beyond the reported tables, which themselves lack error bars.

free parameters (3)
  • temperature scaling factor tau = 3
    Selected by ablation on DCA1 (Table 2); controls smoothing of patch distribution vectors.
  • patch number n = 16 (4x4 patches)
    Selected by ablation on DCA1 (Table 3); determines granularity of the distribution vector.
  • soft label weight alpha_T = 0.5
    Selected by ablation on DCA1 (Table 4); linearly scheduled as alpha_t = alpha_T * t/T.
assumptions (4)
  • domain assumption Decoder side feature maps at different depths encode useful hierarchical knowledge that can be distilled.
    Section 3.2 motivation; never independently verified.
  • domain assumption The model from the previous epoch is a sufficiently reliable teacher.
    Section 3.2.2 and Algorithm 1; no criterion for teacher quality is provided.
  • domain assumption Patch-level counts of foreground/background pixels, after softmax and temperature scaling, form a meaningful distribution for KL distillation.
    Section 3.2.1, Eqs. (2)-(3); Eq. (2) as written computes products, not counts.
  • domain assumption Ground-truth annotations in XCAD and DCA1 are accurate.
    Section 4.1; no inter-observer variability is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Self-knowledge Distillation: A hierarchical supervised learning for coronary artery segmentation." pith.science (2026). https://pith.science/paper/SKOQGMIK

@misc{pith2026250903173,
  author       = {Pith},
  title        = {Pith review of: Deep Self-knowledge Distillation: A hierarchical supervised learning for coronary artery segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SKOQGMIK}},
  note         = {Machine review of arXiv:2509.03173}
}
read the original abstract

Coronary artery disease is a leading cause of mortality, underscoring the critical importance of precise diagnosis through X-ray angiography. Manual coronary artery segmentation from these images is time-consuming and inefficient, prompting the development of automated models. However, existing methods, whether rule-based or deep learning models, struggle with issues like poor performance and limited generalizability. Moreover, current knowledge distillation methods applied in this field have not fully exploited the hierarchical knowledge of the model, leading to certain information waste and insufficient enhancement of the model's performance capabilities for segmentation tasks. To address these issues, this paper introduces Deep Self-knowledge Distillation, a novel approach for coronary artery segmentation that leverages hierarchical outputs for supervision. By combining Deep Distribution Loss and Pixel-wise Self-knowledge Distillation Loss, our method enhances the student model's segmentation performance through a hierarchical learning strategy, effectively transferring knowledge from the teacher model. Our method combines a loosely constrained probabilistic distribution vector with tightly constrained pixel-wise supervision, providing dual regularization for the segmentation model while also enhancing its generalization and robustness. Extensive experiments on XCAD and DCA1 datasets demonstrate that our approach outperforms the dice coefficient, accuracy, sensitivity and IoU compared to other models in comparative evaluations.

Figures

Figures reproduced from arXiv: 2509.03173 by the authors.

Figure 1
Figure 1. The structure of Deep Self-knowledge Distillation. yij ∈ {0, 1}. The objective of training the segmentation network N, is to generate predictions xˆ = N(x) ≈ y effectively, where xˆ denotes the model’s predicted output. In segmentation networks employing an encoder-decoder archi￾tecture (e.g., U-Net [22]), the decoder comprises a series of up￾sampling and convolutional layers. Consequently, each layer in the network… view at source ↗
Figure 2
Figure 2. The workflow of processing the side outputs to probabilistic distribution vectors. 0 and 1 in each patch, thus representing the distribution vector Z˜i for different layers. The distribution vector Z˜i ∈ R n×2 is defined as: Z˜i [m] = [Ys j=1 Ys k=1 1(˜y i,m[j][k] = 1), Ys j=1 Ys k=1 1(˜y i,m[j][k] = 0)] (2) where s denotes the patch size of the side output, Z˜i [m] denotes the mth row of the distribution vector Z˜i… view at source ↗
Figure 3
Figure 3. Quantitative analysis of our proposed method compared to other comparative models. Distribution Loss LDDL and the Pixel-wise Self-Knowledge Distil￾lation Loss LP SKL. We also examine the effects of the temperature scaling factor τ for smoothing the probability distributions, the num￾ber of patches n, and the hyper-parameter α that adjusts the linear combination ratio of the teacher model’s output to the ground truth… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 22 canonical work pages

  1. [1]

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

    work page

  2. [2]

    C. E. Cardenas, R. E. McCarroll, L. E. Court, B. A. Elgohari, H. Elha- lawani, C. D. Fuller, M. J. Kamal, M. A. Meheissen, A. S. Mohamed, A. Rao, et al. Deep learning algorithm for auto-delineation of high- risk oropharyngeal clinical target volumes with built-in dice similarity coefficient parameter optimization function. International Journal of Radiati...

    work page doi:10.1016/j.ijrobp.2018.01.114 2018

  3. [3]

    Cervantes-Sanchez, I

    F. Cervantes-Sanchez, I. Cruz-Aceves, A. Hernandez-Aguirre, M. A. Hernandez-Gonzalez, and S. E. Solorio-Meza. Automatic segmentation of coronary arteries in x-ray angiograms using multiscale analysis and artificial neural networks. Applied Sciences , 9(24):5507, 2019. doi: https://doi.org/10.3390/app9245507

    doi:10.3390/app9245507 2019

  4. [4]

    Z. Fan, J. Mo, B. Qiu, W. Li, G. Zhu, C. Li, J. Hu, Y . Rong, and X. Chen. Accurate retinal vessel segmentation via octave convolu- tion neural network. arXiv preprint arXiv:1906.12193 , 2019. doi: https://doi.org/10.48550/arXiv.1906.12193

    work page Pith review arXiv doi:10.48550/arxiv.1906.12193 1906

  5. [5]

    Gharleghi, N

    R. Gharleghi, N. Chen, A. Sowmya, and S. Beier. Towards automated coronary artery segmentation: A systematic review. Computer Methods and Programs in Biomedicine , 225:107015, 2022. doi: https://doi.org/ 10.1016/j.cmpb.2022.107015

    work page arXiv 2022

  6. [6]

    I. S. Golub, O. G. Termeie, S. Kristo, L. P. Schroeder, S. Laksh- manan, A. M. Shafter, L. Hussein, D. Verghese, J. Aldana-Bitar, V . S. Manubolu, et al. Major global coronary artery calcium guidelines. Car- diovascular Imaging, 16(1):98–117, 2023. doi: https://doi.org/10.1016/ j.jcmg.2022.06.018

    work page 2023

  7. [7]

    J. Gou, B. Yu, S. J. Maybank, and D. Tao. Knowledge distillation: A survey. International Journal of Computer Vision , 129(6):1789–1819,

    work page

  8. [8]

    T. Han, D. Ai, X. Li, J. Fan, H. Song, Y . Wang, and J. Yang. Coronary artery stenosis detection via proposal-shifted spatial-temporal trans- former in x-ray angiography. Computers in Biology and Medicine , 153:106546, 2023. doi: https://doi.org/10.1016/j.compbiomed.2023. 106546

    doi:10.1016/j.compbiomed.2023 2023

Show all 36 references
  1. [9]

    Hinton, O

    G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 , 2015. doi: https://doi.org/ 10.48550/arXiv.1503.02531

    2015 doi arXiv

  2. [10]

    Huang, L

    H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y . Iwamoto, X. Han, Y .- W. Chen, and J. Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP 2020-2020 IEEE international con- ference on acoustics, speech and signal processing (ICASSP) , pages 1055–1...

    2020

  3. [11]

    Hung, C.-Y

    P.-K. Hung, C.-Y . Liu, C.-Y . Hsu, C.-Y . Huang, W.-J. Lee, T.-D. Wang, and C.-M. Chen. Automated coronary artery segmentation and calcified/non-calcified plaque measurement. In Proceedings of the In- ternational Conference on Image Processing, Computer Vision, and Pattern Re...

    2013

  4. [12]

    D. Jha, P. H. Smedsrud, M. A. Riegler, D. Johansen, T. De Lange, P. Halvorsen, and H. D. Johansen. Resunet++: An advanced archi- tecture for medical image segmentation. In 2019 IEEE international symposium on multimedia (ISM) , pages 225–2255. IEEE, 2019. doi: https://doi.org/...

    2019

  5. [13]

    M. Ji, S. Shin, S. Hwang, G. Park, and I.-C. Moon. Refine my- self by teaching myself: Feature refinement via self-knowledge distil- lation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition , pages 10664–10673, 2021. doi: https: //doi.org/...

    2021

  6. [14]

    Jiang, Y

    W. Jiang, Y . Li, Y . Jia, Y . Feng, Z. Yi, M. Chen, and J. Wang. Ori- net: Orientation-guided neural network for automated coronary arteries segmentation. Expert Systems with Applications , 238:121905, 2024. doi: https://doi.org/10.1016/j.eswa.2023.121905

    2024

  7. [15]

    K. B. Khan, M. S. Siddique, M. Ahmad, and M. Mazzara. A hybrid unsupervised approach for retinal vessel segmentation. BioMed Re- search International, 2020(1):8365783, 2020. doi: https://doi.org/10. 1155/2020/8365783

    2020

  8. [16]

    K. Kim, B. Ji, D. Yoon, and S. Hwang. Self-knowledge distillation: A simple way for better generalization. arXiv preprint arXiv:2006.12000, 3:1, 2020. doi: https://doi.org/10.48550/arXiv.2006.12000

    2006 doi arXiv

  9. [17]

    X. Li, S. Pang, R. Zhang, J. Zhu, X. Fu, Y . Tian, and J. Gao. Attransunet: An enhanced hybrid transformer architecture for ultra- sound and histopathology image segmentation. Computers in Biol- ogy and Medicine , 152:106365, 2023. doi: https://doi.org/10.1016/j. compbiomed.20...

    2023

  10. [18]

    Y . Li, Y . Wu, J. He, W. Jiang, J. Wang, Y . Peng, Y . Jia, T. Xiong, K. Jia, Z. Yi, et al. Automatic coronary artery segmentation and di- agnosis of stenosis by deep learning based on computed tomographic coronary angiography. European Radiology, 32(9):6037–6045, 2022. doi: ...

    2022 doi

  11. [19]

    Y . Ma, Y . Hua, H. Deng, T. Song, H. Wang, Z. Xue, H. Cao, R. Ma, and H. Guan. Self-supervised vessel segmentation via adversarial learn- ing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7536–7545, 2021. doi: https://doi.org/10.1109/ ICCV...

    2021

  12. [20]

    Oktay, J

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Mis- awa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainz, et al. Atten- tion u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. doi: https://doi.org/10.48550/arXiv.1804. 03999

    2018 doi arXiv

  13. [21]

    S. Park, J. Kim, and Y . S. Heo. Semantic segmentation using pixel-wise adaptive label smoothing via self-knowledge distillation for limited la- beling data. Sensors, 22(7):2623, 2022. doi: https://doi.org/10.3390/ s22072623

    2022

  14. [22]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional net- works for biomedical image segmentation. In Medical image comput- ing and computer-assisted intervention–MICCAI 2015: 18th interna- tional conference, Munich, Germany, October 5-9, 2015, proceedings, part III 1...

    2015

  15. [23]

    N. Shen, T. Xu, S. Huang, F. Mu, and J. Li. Expert-guided knowledge distillation for semi-supervised vessel segmentation. IEEE Journal of Biomedical and Health Informatics , 2023. doi: https://doi.org/10.1109/ JBHI.2023.3312338

    2023

  16. [24]

    A. Song, L. Xu, L. Wang, B. Wang, X. Yang, B. Xu, B. Yang, and S. E. Greenwald. Automatic coronary artery segmentation of ccta images with an efficient feature-fusion-and-rectification 3d-unet. IEEE Journal of Biomedical and Health Informatics , 26(8):4044–4055, 2022. doi: htt...

    2022

  17. [25]

    K. Sun, B. Xiao, D. Liu, and J. Wang. Deep high-resolution rep- resentation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 5693–5703, 2019. doi: https://doi.org/10.1109/CVPR.2019. 00584

    2019 doi

  18. [26]

    X. J. Tang, S. J. Zhou, and Q. M. Hu. Segmentation of coronary ct angiography images based on deformable model with new edge mea- sures. Applied Mechanics and Materials , 333:888–896, 2013. doi: https://doi.org/10.4028/www.scientific.net/AMM.333-335.888

    2013 doi

  19. [27]

    Wang and Y

    Y . Wang and Y . Yao. Application of artificial intelligence methods in carotid artery segmentation: a review. IEEE Access, 2023. doi: https: //doi.org/10.1109/ACCESS.2023.3243162

    2023

  20. [28]

    Y . Wang, X. Meng, and D. Wang. Automated coronary artery anal- ysis system in 3d cta images. In 2010 3rd International Conference on Biomedical Engineering and Informatics, volume 1, pages 122–126. IEEE, 2010. doi: https://doi.org/10.1109/BMEI.2010.5639471

    2010

  21. [29]

    Yang and S

    Z. Yang and S. Farsiu. Directional connectivity-based segmentation of medical images. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11525–11535, 2023. doi: https://doi.org/10.1109/cvpr52729.2023.01109

    2023

  22. [30]

    Zhang, X

    H. Zhang, X. Ran, and W. Zhou. Self-knowledge distillation-based staged extraction and multiview collection network for rgb-d mirror seg- mentation. IEEE Signal Processing Letters , 31:1029–1033, 2024. doi: https://doi.org/10.1109/LSP.2024.3386470

    2024

  23. [31]

    Zhang, L

    N. Zhang, L. Yu, D. Zhang, W. Wu, S. Tian, X. Kang, and M. Li. Ct-net: Asymmetric compound branch transformer for medical im- age segmentation. Neural Networks , 170:298–311, 2024. doi: https: //doi.org/10.1016/j.neunet.2023.11.034

    2024 doi

  24. [32]

    Zhang, Q

    Z. Zhang, Q. Liu, and Y . Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters , 15(5):749–753, 2018. doi: https://doi.org/10.1109/LGRS.2018.2802944

    2018

  25. [33]

    Zheng, M

    Y . Zheng, M. Sun, X. Wang, T. Cao, X. Zhang, L. Xing, and Z. Fang. Self-distillation object segmentation via pyramid knowledge represen- tation and transfer. Multimedia Systems, 29(5):2615–2631, 2023. doi: https://doi.org/10.1007/s00530-023-01121-x

    2023 doi

  26. [34]

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learn- ing for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th ...

    2018

  27. [2021]

    doi: https://doi.org/10.1007/s11263-021-01453-z

    doi

  28. [2022]

    doi: https://doi.org/10.1007/978-3-031-25066-8_9

    doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.