Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper uses a stochastic label-consistent projector during discrete diffusion sampling to enforce semantic edge-label constraints, yielding 100% valid circle-of-Willis and airway graphs and better downstream vessel labeling.

desk verdict Stochastic label-consistent projection is a real extension, but the headline downstream win is confounded by training-set size and the semantic validity metric is true by construction. read the letter →

arxiv 2507.04856 v1 pith:66KAPZLO submitted 2025-07-07 cs.CV

Chinmay Prabhakar , Suprosanna Shit , Tamaz Amiranashvili , Hongwei Bran Li , Bjoern Menze This is my paper
classification cs.CV
keywords discretediffusion3DbiologicalgraphssemanticconsistencygraphgenerationcircleofWillisairwaytreesvessellabelinglinkprediction
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 tries to make discrete diffusion models generate 3D biological graphs—blood-vessel networks and airway trees—that are not just statistically similar to real ones but anatomically valid. Its central claim is that a stochastic "label-consistent projection" step during sampling can enforce semantic edge-label constraints (for example, forbidding a basal artery from being directly connected to an anterior cerebral artery) without sacrificing the distribution of graph statistics. On two datasets, circle of Willis and lung airways, the method produces 100% semantically valid generated graphs, whereas the baselines produce 70% and 10% valid samples respectively. The paper also claims the synthetic circle-of-Willis graphs train a downstream vessel labeler to higher balanced accuracy (95.45) than training on real data (93.77), and that the same model works out of the box as a link predictor on airway trees.

What carries the argument

The central object is the semantic-consistency projector, a sampling-time module that checks candidate edges against an edge-label incompatibility matrix $\Omega \in \{0,1\}^{c\times c}$, where $\Omega[c_i,c_j]=1$ means edge types $c_i$ and $c_j$ cannot be adjacent around the same node. The projector works with an edge-deletion noising transition $Q_t = \alpha_t^e I + (1-\alpha_t^e)\mathbf{1} e_E^\top$, which treats "no edge" as an absorbing state so reverse sampling only adds edges. When the edge denoiser proposes an edge whose label is incompatible, the projector resamples that edge label up to $k$ times from the posterior $p_\phi(E^{t-1}_{ij}\mid E^t_{ij})$ and accepts the edge if any resampled label is compatible; otherwise it deletes the edge. This stochastic repair is what lets the method enforce semantics without the oversparsity caused by hard deletion.

What would settle it

Take a dataset with an independent expert-written set of forbidden vessel or airway label adjacencies that was not used to build the constraint matrix, generate a large batch of graphs, and check whether any generated graph contains a forbidden adjacency. One such graph would show the validity guarantee is only as good as the hand-written constraint list.

Watch

Extended reading notes

Core claim

The paper claims that anatomical validity in generated 3D biological graphs can be enforced as a byproduct of sampling rather than by post-processing. It achieves this by showing that semantic edge-label consistency is invariant under an edge-deletion noising process, so a stochastic projector can be applied at every reverse-diffusion step. The outcome on the circle-of-Willis and airway datasets is 100% semantically valid generated graphs with distribution statistics closer to real data than the uniform-noising baseline; on the circle of Willis, synthetic graphs used as training data push vessel-labeling balanced accuracy to 95.45 versus 93.77 for real data; and on airway trees the same model predicts missing links at 84.97 balanced accuracy while keeping all predictions valid. The paper also states its guarantee is limited to edge-deletion invariants, so properties such as a single connected component are not enforced.

Load-bearing premise

The method assumes the hand-built list of forbidden label pairs completely captures anatomical plausibility and that checking each node's local neighbors is enough, so a missing or wrongly included pair would make "valid" graphs anatomically wrong or artificially sparse.

Editorial extensions

If this is right

  • All generated samples satisfy the semantic-consistency criterion defined by $\Omega$ (100% on both datasets), so downstream users can skip rejection sampling.
  • Synthetic circle-of-Willis graphs train a vessel labeler that outperforms one trained on real data (95.45 vs 93.77 balanced accuracy, avg F1 0.967 vs 0.946).
  • The same model can be used for link prediction on airway graphs with roughly 30% missing edges, reaching 84.97 balanced accuracy while keeping 100% semantic validity.
  • Edge-deletion noising is better suited to sparse biological graphs than uniform-label noising, because early noising steps do not inject implausible long-range edges.

Reading between the lines

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

  • Editorial extension: because the projector is applied at sampling time and does not depend on a specific architecture, the same stochastic-repair idea could be ported to other discrete diffusion models with pairwise compatibility constraints, such as molecular or material graphs.
  • Editorial extension: the validity guarantee is only as strong as the hand-defined $\Omega$; learning this matrix from anatomical data or clinical ontologies would generalize the method to constraints the authors did not encode.
  • Editorial extension: the paper's stated limitation that node coordinates are frozen during edge denoising suggests a testable extension where coordinates and edges are denoised jointly or alternately, which could improve anatomical fidelity further.
  • Editorial extension: the reported low intervention rate (about 2% of generated edges) suggests the projector acts as a targeted safety valve rather than a wholesale override, which may explain why distribution statistics stay close to real data.
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

4 major / 4 minor

Summary. The paper proposes a discrete diffusion model for generating 3D biological graphs, specifically circle of Willis vessel graphs and airway trees. The method has two stages: a point-cloud DDPM generates node coordinates, and an edge denoiser with an edge-deletion (absorbing-state) noising process generates edge labels. During sampling, a novel 'label-consistent projector' checks newly sampled edges against a manually defined edge-label adjacency incompatibility matrix Ω (Eq. 1) and, when a violation occurs, resamples the edge label up to k times before rejecting the edge. The authors report 100% semantic validity on both datasets, improved distribution statistics relative to baselines, and downstream gains in vessel labeling and link prediction.

Significance. If validated, the paper would make a useful contribution: it adapts edge-deletion discrete diffusion to sparse biological graphs, and it extends the hard-projection idea of Madeira et al. [9] to semantic edge-label constraints via stochastic resampling. The code is promised publicly, and the downstream applications (vessel labeling, link prediction) are practically relevant. However, the current evidence is weakened by the fact that the headline semantic-validity metric is the same predicate the projector enforces, and by a downstream comparison that confounds training-set size with data source. The central method is plausible and likely fixable, but several load-bearing evaluation points need to be reworked before the paper's claims are supported.

major comments (4)
  1. [Section 3, Eq. (1); Section 4, Tables 1 and 3] The semantic validity metric S.V. is defined by exactly the predicate P_Ω(E) in Eq. (1), and the projector is designed to enforce P_Ω at every denoising step. Consequently, the reported 100% S.V. scores in Tables 1 and 3 are guaranteed by construction and do not provide independent empirical evidence that the generated graphs are anatomically valid. The paper should either (a) report an independent anatomical validity measure, e.g., expert review or a held-out constraint set, or (b) explicitly frame the 100% S.V. as a correctness check of the projector rather than an empirical result.
  2. [Section 4, 'Downstream Tasks', Table 2] The claim that synthetic samples outperform real data for vessel labeling is confounded by training-set size: the synthetic condition trains on 1000 generated graphs while the real-data baseline trains on a 50-graph train split. A 20-fold difference in training volume could explain the 1.7-point balanced-accuracy gap even if the generated graphs are no more realistic than the real ones. The paper needs a matched comparison, e.g., training the real-data baseline on the same number of graphs (via augmentation or repeated sampling) or training all methods on equal-size synthetic corpora, and it should report error bars over multiple seeds.
  3. [Section 3, 'Semantically Consistent Projection'] The text states that the projector 'allows us to sample from the posterior distribution constrained by our semantic plausibility,' but this is not established. The k-resampling scheme is a heuristic: it samples a label from the model's posterior, checks P_Ω, and accepts the first valid label, which does not generally equal conditioning on the validity event. The theoretical result from [9] is about hard projection for invariant properties, not about stochastic resampling. The authors should either provide a correctness argument for the resampling distribution or soften the claim to say that the projector biases samples toward valid configurations in an empirically tested way.
  4. [Section 4, 'Semantic Validity (S.V.)' and Eq. (1)] The Ω matrix is manually constructed from the TopCoW annotation protocol, and the same protocol is used to define edge labels and the validity criterion. There is no evidence that Ω is complete (no omitted forbidden adjacencies) or not over-restrictive. If Ω is incomplete, the projector could certify graphs that are anatomically implausible; if it is over-restrictive, it could force unrealistic sparsity. The paper should validate Ω against independent anatomical knowledge or at least report sensitivity to its construction, e.g., by comparing S.V. under an expert-reviewed subset of Ω.
minor comments (4)
  1. [Abstract and Introduction] The abstract contains a typo, 'link predictior'; the Introduction also repeats this in the contributions list. Please correct.
  2. [Section 4, 'Datasets'] The sentence describing CROWN node counts says 'between 13 and 27 in CoW (average 8)', which is numerically inconsistent; likely the average should be near 18 or the range is a typo. Please clarify.
  3. [Section 4, 'Metrics'] There is an apparent typo 'A TM' for 'ATM' in the semantic validity paragraph. Please fix.
  4. [Section 3, Eq. (1)] Equation (1) is notationally dense and hard to parse; the tensor product over edges at a node would benefit from an explicit example or a rephrased definition, e.g., in terms of the number of pairs of incident edges whose labels are marked incompatible.

Circularity Check

2 steps flagged · score 4.0 of 10

The 100% semantic-validity result is entailed by the projector, and the validity standard is inherited from the authors' own TopCoW protocol; downstream labeling/link-prediction results are independent evidence.

  1. self definitional [Section 3 'Semantically Consistent Projection' and Section 4 'Semantic Validity (S.V.)']
    "The projector begins with Et−1 ← Et and adds edges in Et−1 that satisfy the semantic-consistency criteria. Specifically, for each edge eij ∈ ˆEt−1 \ Et our projector adds Et−1 ← Et−1∪{eij} if PΩ(Et−1∪{eij}) = True. If not, we sample k new edge labels from eij ∼ pϕ(Et−1|Et). ... The projector accepts the edge if any of the k resampled edge labels satisfy the semantic consistency; otherwise, it rejects the edge."

    The semantic-validity metric is exactly the predicate PΩ(E)=True that the projector uses as its acceptance test. Every edge that would make PΩ false is either relabeled until consistent or rejected before the final graph is emitted. Therefore the 'S.V.% 100' entries for Ours and Ours(k=0) in Table 1 are logically forced by the sampling procedure, not empirically discovered. The non-tautological content of Table 1 is restricted to the KL distribution statistics, which are measured against the real data independently of the projector's own constraint check.

  2. self citation load bearing [Section 4, 'Semantic Validity (S.V.)' and reference [24] (TopCoW)]
    "Semantic Validity (S.V.): We use edge-label based criterion for semantic validity. CoW: The CoW graph has no structural constraints, and the TopCoW annotation protocol [24] determines its validity, please refer [24] for details. For CoW, we have 13x13 Ω matrix for 13 edge-labels."

    The authority for anatomical validity is delegated entirely to the TopCoW annotation protocol [24], whose author list overlaps with the present paper (Li, Menze, and likely other co-authors). The Ω matrix that defines 'valid' in Eq. (1) is exactly the constraint set enforced by the projector, so the reported semantic validity measures agreement with the authors' own protocol rather than with an independent anatomical standard. This self-citation chain is load-bearing for the 'anatomically plausible graphs' claim, though softened by TopCoW being a public multi-author challenge dataset.

full rationale

The paper's core methodological claim, that stochastic projection enforces semantic consistency, is true by construction: the projector only emits graphs satisfying PΩ, and the S.V. metric is PΩ, making 100% S.V. a definitional outcome rather than a prediction. The Ω standard itself is inherited from the authors' own TopCoW protocol, a mild self-citation that further weakens the independence of the validity claim. However, the paper does contain substantial independent evaluation: KL distribution statistics are compared with real data, and the downstream vessel labeling and airway link-prediction tasks use external ground-truth labels and held-out TopCoW/ATM samples. Table 2's 'synthetic beats real' comparison is statistically confounded (1000 synthetic training graphs vs. 50 real graphs, no error bars), but a confound is not a circularity; the downstream result is not equivalent to the projector's input. On balance, the semantic-validity headline is circular to a moderate degree, while the utility claims have independent content, supporting a score of 4 rather than a higher score.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces a new algorithmic component (stochastic label-consistent projector) but no new physical or modeled entities. The main loaded assumptions are the correctness/completeness of the Omega matrix and the invariance of the semantic constraint under edge-deletion noising. Free parameters include the resampling count k, the Omega matrix itself, and the link-prediction denoising step count.

free parameters (3)
  • k (resampling attempts) = 4
    Number of label resamples per edge during projection. Chosen by hand, affects the density vs. validity trade-off, and no ablation or sensitivity analysis is reported.
  • Omega constraint matrix = 13x13 for CoW, 4x4 for ATM
    Hand-specified forbidden-adjacency matrix defining semantic validity. It is not fitted to data but is a manual modeling choice that determines the validity definition and therefore the 100% S.V. result.
  • Link prediction denoising steps = 100
    The diffusion model is trained with 500 noising steps but link prediction uses 100 steps, chosen without analysis of the effect on performance.
assumptions (5)
  • standard math Standard DDPM and discrete diffusion Markov-chain assumptions hold for coordinate and edge noising.
    Invoked in Section 3 for the coordinate DDPM and the edge transition matrix Q^t.
  • standard math Edge-deletion noising preserves the P_Omega constraint because it only removes edges and never changes labels.
    Stated in Section 3: the deletion process does not change relationships among neighboring edges, motivating the label-aware projector.
  • domain assumption The manually defined Omega matrix correctly and completely characterizes anatomical plausibility for both datasets.
    The projector enforces exactly Omega in Eq. 1; if Omega is incomplete or wrong, the generated graphs can be spuriously valid. Omega is derived from TopCoW protocol for CoW and a radius-based hierarchy for ATM, with no independent validation.
  • domain assumption Training graphs satisfy the Omega constraint, so the denoiser learns to propose consistent edges.
    The projector corrects violations, but if training data contained many violations the model could over-rely on correction. The paper does not report the S.V. of the training graphs.
  • domain assumption The graph transformer edge denoiser, trained on fixed node coordinates, assigns meaningful posterior probabilities even when coordinates come from a separately trained point cloud generator.
    The two-stage generation fixes X^0 during edge denoising, as acknowledged in the Limitations section, so the edge posterior is conditioned on potentially imperfect coordinates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling." pith.science (2026). https://pith.science/paper/66KAPZLO

@misc{pith2026250704856,
  author       = {Pith},
  title        = {Pith review of: Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66KAPZLO}},
  note         = {Machine review of arXiv:2507.04856}
}
read the original abstract

3D spatial graphs play a crucial role in biological and clinical research by modeling anatomical networks such as blood vessels,neurons, and airways. However, generating 3D biological graphs while maintaining anatomical validity remains challenging, a key limitation of existing diffusion-based methods. In this work, we propose a novel 3D biological graph generation method that adheres to structural and semantic plausibility conditions. We achieve this by using a novel projection operator during sampling that stochastically fixes inconsistencies. Further, we adopt a superior edge-deletion-based noising procedure suitable for sparse biological graphs. Our method demonstrates superior performance on two real-world datasets, human circle of Willis and lung airways, compared to previous approaches. Importantly, we demonstrate that the generated samples significantly enhance downstream graph labeling performance. Furthermore, we show that our generative model is a reasonable out-of-the-box link predictior.

Figures

Figures reproduced from arXiv: 2507.04856 by the authors.

Figure 1
Figure 1. Our method first generates node coordinates using a point cloud denoiser. Subsequently, the edge denoiser is trained with an edge-deletion noise model. During inference, the edge denoiser suggests a set of new edges. Our novel semantic consistent projector discards an edge if none of k resampled edge labels meet the validity criteria. to generate airway trees. Recently, data-driven methods have been proposed [7,14,2… view at source ↗
Figure 2
Figure 2. Qualitative examples of graphs generated by [14], Construct, and our method compared to ground truth samples. The constraint violations are highlighted in red.[14] produces samples with label and structural violations (cycles in airway trees). Construct can produce samples with label violations. Our method tries to adhere to both label and structural constraints, potentially fixing the issues (shown in blue). a safe… view at source ↗
Figure 3
Figure 3. Downstream applications of the diffusion model. Left: The vessel labeler re￾ceives circle of Willis graphs without any label and assigns multi-class labels based on TopCoW annotation protocol [24]. Right: The diffusion model predicts missing links on the airway tree samples. We show the ground truth samples for reference. Acknowledgments. This work has been supported by the Helmut Horten Foun￾dation. S. S. is suppor… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [9]

    Advances in Neural Information Processing Systems 37, 137218–137262 (2025)

    Madeira, M., Vignac, C., Thanou, D., Frossard, P.: Generative modelling of struc- turally constrained graphs. Advances in Neural Information Processing Systems 37, 137218–137262 (2025)

    work page 2025

  2. [1]

    Advances in Neural Information Processing Systems 34, 17981–17993 (2021)

    Austin, J., Johnson, D.D., Ho, J., Tarlow, D., Van Den Berg, R.: Structured de- noising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981–17993 (2021)

    work page 2021

  3. [2]

    BMC bioinformatics22(1), 1–28 (2021)

    Drees, D., Scherzinger, A., Hägerling, R., Kiefer, F., Jiang, X.: Scalable robust graph and feature extraction for arbitrary vessel networks in large volumetric datasets. BMC bioinformatics22(1), 1–28 (2021)

    work page 2021

  4. [3]

    In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention

    Feldman, P., Fainstein, M., Siless, V., Delrieux, C., Iarussi, E.: Vesselvae: Recur- sive variational autoencoders for 3d blood vessel synthesis. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 67–76. Springer (2023)

    2023

  5. [4]

    Advances in neural information processing systems36, 12489–12517 (2023)

    Gruver, N., Stanton, S., Frey, N., Rudner, T.G., Hotzel, I., Lafrance-Vanasse, J., Rajpal, A., Cho, K., Wilson, A.G.: Protein design with guided discrete diffusion. Advances in neural information processing systems36, 12489–12517 (2023)

    work page 2023

  6. [5]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  7. [6]

    Journal of the Royal Society Interface 19(191), 20220087 (2022) 10 Prabhakar and Shit et al

    Jessen, E., Steinbach, M.C., Debbaut, C., Schillinger, D.: Rigorous mathemati- cal optimization of synthetic hepatic vascular trees. Journal of the Royal Society Interface 19(191), 20220087 (2022) 10 Prabhakar and Shit et al

    work page 2022

  8. [7]

    In: Medical Imaging with Deep Learning (2024)

    Kuipers, T.P., Konduri, P.R., Marquering, H., Bekkers, E.J.: Generating cerebral vessel trees of acute ischemic stroke patients using conditional set-diffusion. In: Medical Imaging with Deep Learning (2024)

    2024

Show all 26 references
  1. [8]

    arXiv preprint arXiv:2310.16834 (2023)

    Lou, A., Meng, C., Ermon, S.: Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834 (2023)

    2023 arXiv

  2. [10]

    Clinical Biomechanics66, 60–65 (2019)

    Van de Moortele, T., Goerke, U., Wendt, C.H., Coletti, F.: Airway morphology and inspiratory flow features in the early stages of chronic obstructive pulmonary disease. Clinical Biomechanics66, 60–65 (2019)

    2019

  3. [11]

    PloS one 15(4), e0230259 (2020)

    Nousias,S.,Zacharaki,E.I.,Moustakas,K.:Avatree:Anopen-sourcecomputational modelling framework modelling anatomically valid airway tree conformations. PloS one 15(4), e0230259 (2020)

    2020

  4. [12]

    In: Thirty-Fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)

    Paetzold, J.C., McGinnis, J., Shit, S., Ezhov, I., Büschl, P., Prabhakar, C., Sekuboyina, A., Todorov, M., Kaissis, G., Ertürk, A., et al.: Whole brain ves- sel graphs: A dataset and benchmark for graph learning and neuroscience. In: Thirty-Fifth Conference on Neural Informati...

    2021

  5. [13]

    Journal of microscopy199(2), 149–161 (2000)

    Pothuaud, L., Porion, P., Lespessailles, E., Benhamou, C., Levitz, P.: A new method for three-dimensional skeleton graph analysis of porous media: applica- tion to trabecular bone microarchitecture. Journal of microscopy199(2), 149–161 (2000)

    2000

  6. [14]

    In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

    Prabhakar, C., Shit, S., Musio, F., Yang, K., Amiranashvili, T., Paetzold, J.C., Li, H.B., Menze, B.: 3d vessel graph generation using denoising diffusion. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. pp. 3–13. Springer (2024)

    2024

  7. [15]

    Pulmonary circulation6(1), 70–81 (2016)

    Rahaghi, F.N., Ross, J.C., Agarwal, M., González, G., Come, C., Diaz, A., Vegas- Sánchez-Ferrero, G., Hunsaker, A., Estépar, R.S.J., Waxman, A., et al.: Pulmonary vascular morphology as an imaging biomarker in chronic thromboembolic pul- monary hypertension. Pulmonary circulat...

    2016

  8. [16]

    In: Eurographics 2021 - Short Papers

    Rauch, N., Harders, M.: Interactive Synthesis of 3D Geometries of Blood Ves- sels. In: Eurographics 2021 - Short Papers. The Eurographics Association (2021). https://doi.org/10.2312/egs.20211012

    2021 doi

  9. [17]

    Journal of Cerebral Blood Flow & Metabolism29(8), 1429–1443 (2009)

    Reichold,J.,Stampanoni,M.,Keller,A.L.,Buck,A.,Jenny,P.,Weber,B.:Vascular graph model to simulate the cerebral blood flow in realistic vascular networks. Journal of Cerebral Blood Flow & Metabolism29(8), 1429–1443 (2009)

    2009

  10. [18]

    Mathematics 8(12), 2236 (2020)

    Savinkov, R., Grebennikov, D., Puchkova, D., Chereshnev, V., Sazonov, I., Bocharov, G.: Graph theory for modeling and analysis of the human lymphatic system. Mathematics 8(12), 2236 (2020)

    2020

  11. [19]

    In: Forty-first International Conference on Machine Learning (2024)

    Sharma, K., Kumar, S., Trivedi, R.: Diffuse, sample, project: plug-and-play con- trollable graph generation. In: Forty-first International Conference on Machine Learning (2024)

    2024

  12. [20]

    arXiv preprint arXiv:2209.14734 (2022)

    Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V., Frossard, P.: Digress: Discrete denoising diffusion for graph generation. arXiv preprint arXiv:2209.14734 (2022)

    2022 arXiv

  13. [21]

    arXiv preprint arXiv:2302.09048 (2023)

    Vignac, C., Osman, N., Toni, L., Frossard, P.: Midi: Mixed graph and 3d denoising diffusion for molecule generation. arXiv preprint arXiv:2302.09048 (2023)

    2023 arXiv

  14. [22]

    In: Medical Imaging with Deep Learning (2024) Semantically Consistent Discrete Diffusion 11

    Vos, I., Ruigrok, Y., Kuijf, H.: Results of the crown challenge on automated as- sessment of circle of willis morphology. In: Medical Imaging with Deep Learning (2024) Semantically Consistent Discrete Diffusion 11

    2024

  15. [23]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wittmann, B., Wattenberg, Y., Amiranashvili, T., Shit, S., Menze, B.: vesselfm: A foundation model for universal 3d blood vessel segmentation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20874–20884 (2025)

    2025

  16. [24]

    arXiv preprint arXiv:2312.17670 (2024)

    Yang, K., Musio, F., Ma, Y., Juchler, N., Paetzold, J.C., Al-Maskari, R., Höher, L., Li, H.B., Hamamci, I.E., Sekuboyina, A., et al.: Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. arXiv preprint ...

    2024 arXiv

  17. [25]

    arXiv preprint arXiv:2401.09500 (2024)

    Yang, N., Zeng, K., Lu, H., Wu, Y., Yuan, Z., Chen, D., Jiang, S., Wu, J., Wang, Y., Yan, J.: Morphgrower: A synchronized layer-by-layer growing approach for plausible neuronal morphology generation. arXiv preprint arXiv:2401.09500 (2024)

    2024 arXiv

  18. [26]

    Medical image analysis 90, 102957 (2023)

    Zhang, M., Wu, Y., Zhang, H., Qin, Y., Zheng, H., Tang, W., Arnold, C., Pei, C., Yu, P., Nan, Y., et al.: Multi-site, multi-domain airway tree modeling. Medical image analysis 90, 102957 (2023)

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.