pith. sign in

arxiv: 2605.01382 · v1 · submitted 2026-05-02 · 💻 cs.CV · cs.AI

Sparse Representation Learning for Vessels

Chinmay Prabhakar , Bastian Wittmann , Paul B\"uschl , Hongwei Bran Li , Bjoern Menze , Suprosanna Shit This is my paper

Pith reviewed 2026-05-09 15:18 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords sparse representation learning3D vascular networksvariational autoencodersparse convolutionsmedical image compressionvessel classificationgenerative vascular modeling
0
0 comments X

The pith

A sparse variational autoencoder compresses entire organ vascular networks 512-fold while preserving clinical features in its latent space.

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

The paper introduces VAEsselSparse, an encoder-decoder model for creating compact representations of full organ-level vascular networks from high-resolution 3D scans. It exploits the natural sparsity of blood vessels and tubular structures with sparse convolutions and attention to reach 8x8x8 spatial compression. This yields better reconstruction than dense models and produces a latent space that supports direct classification of vessel abnormalities and generation of realistic vasculature. A sympathetic reader would care because prior methods have been limited to small sub-regions or simplified trees, making whole-organ analysis at clinical resolution computationally infeasible.

Core claim

VAEsselSparse is an efficient encoder-decoder model to obtain a meaningful yet compact representation of the entire organ-level vascular network at sub-millimeter resolution. It leverages the inherent sparsity of 3D vascular structures via sparse convolutions and attention mechanisms, achieving substantial spatial compression rates of 8 x 8 x 8. The model demonstrates superior reconstruction performance compared to dense counterparts and previous methods. The resulting latent space retains clinically relevant discriminative features readily usable for classification tasks such as aneurysm/stenosis or subvariants of the circle of Willis. Moreover, the compact latent space serves as an有效代表 for

What carries the argument

VAEsselSparse variational autoencoder that applies sparse convolutions and attention mechanisms to exploit the natural sparsity of 3D vascular data for high-ratio compression and retention of discriminative features.

If this is right

  • Full organ vascular networks become analyzable at sub-millimeter resolution without restricting to small sub-regions.
  • The latent space directly supports classification of clinical conditions such as aneurysms, stenosis, and circle-of-Willis variants.
  • Compact representations enable generative models to learn vessel-specific priors and synthesize realistic vasculature.
  • Reconstruction fidelity exceeds that of standard dense variational autoencoders on vascular imaging tasks.

Where Pith is reading between the lines

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

  • The method could lower memory and compute barriers enough to support real-time vascular analysis during clinical procedures.
  • Sparse operations might transfer to other sparse tubular structures in medical imaging, such as airways or neural tracts.
  • Generative synthesis from the latent space could augment training data for rare vascular conditions.
  • The compact representation opens a path to joint modeling of vascular networks across multiple organs or imaging modalities.

Load-bearing premise

Sparsity patterns in real clinical vascular data remain sufficiently regular and stable across patients and organs for sparse convolutions and attention to deliver both high compression and faithful reconstruction without losing critical signals.

What would settle it

A controlled test on diverse multi-organ, multi-patient vascular datasets that shows reconstruction error rising sharply or classification accuracy falling below dense-model baselines would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.01382 by Bastian Wittmann, Bjoern Menze, Chinmay Prabhakar, Hongwei Bran Li, Paul B\"uschl, Suprosanna Shit.

Figure 1
Figure 1. Figure 1: Overview. VAEsselSparse represents an encoder-decoder architecture comprised of sparse convolution and sparse attention layers and is, therefore, tailored to operate on sparse input masks. Our learned latent space is suitable for a wide variety of tasks, inclding generative modeling and classification. In contrast, our work operates directly on the entire organ-level high-resolution segmentation map and do… view at source ↗
Figure 2
Figure 2. Figure 2: Representative samples from the reconstruction experiments on COSTA (top row), AIB (middle row), and PARSE (bottom row) test sets. ATM dataset, as VesselGPT is only applicable to tree-like structures. For the classification task, we compare against the ResNet and ViT baselines operating in the voxel space and against the PCA+RF model on the dense VAE features. Implementation Details: VAEsselSparse’s encode… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of 15 PCA components of VAEsselSparse’s latent space on the INSTED dataset. The analysis reveals emerging clusters based on the classes healthy/stenosis/aneurysm (right), whereas the dense VAE fails to capture any (left). Metrics: For the reconstruction task, we report Dice, clDice [21], mean absolute Betti 0 error (|∆β0|), and Betti 1 error (|∆β1|). For the classification task, we repo… view at source ↗
Figure 4
Figure 4. Figure 4: Unconditional generated samples from the denoising U-Net-based flow￾matching model trained on VAEsselSparse’s latent space of the ATM dataset. both the INSTED and TopCoW datasets, which were unseen to VAEsselSparse dur￾ing training, with a simple MLP over strong ResNet and ViT baselines. Even with PCA+RF, VAEsselSparse achieves competitive performance, whereas the dense VAE model underperforms. We observe … view at source ↗
read the original abstract

Analyzing human vasculature and vessel-like, tubular structures, such as airways, is crucial for disease diagnosis and treatment. Current methods often rely on small sub-regions or simplified tree-like structures, rendering analysis of entire organ-level networks at clinical resolution computationally challenging. To this end, we propose VAEsselSparse, an efficient encoder-decoder model to obtain a meaningful yet compact representation of the entire organ-level vascular network at sub-millimeter resolution. VAEsselSparse leverages the inherent sparsity of 3D vascular structures via sparse convolutions and attention mechanisms, achieving substantial spatial compression rates of 8 x 8 x 8. We demonstrate superior reconstruction performance compared to dense counterparts and previous methods. Importantly, the resulting latent space retains clinically relevant discriminative features readily usable for classification tasks, such as aneurysm/stenosis or subvariants of the circle of Willis. Moreover, the compact latent space of VAEsselSparse serves as an effective representation for learning vessel-specific priors through generative models, enabling the synthesis of realistic vasculature.

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 / 2 minor

Summary. The manuscript proposes VAEsselSparse, an encoder-decoder model that exploits the sparsity of 3D vascular structures using sparse convolutions and attention mechanisms. It claims to achieve an 8×8×8 spatial compression rate for full organ-level networks at sub-millimeter resolution, with superior reconstruction performance relative to dense counterparts and prior methods, while preserving clinically relevant discriminative features in the latent space for downstream tasks such as aneurysm/stenosis classification and circle-of-Willis subvariant identification, and enabling generative modeling of realistic vasculature.

Significance. If the empirical claims are substantiated with rigorous quantitative validation, the work could enable scalable analysis of complete high-resolution vascular networks that are currently intractable due to memory and compute constraints. Retaining clinical utility in a highly compressed latent representation would support both diagnostic pipelines and data augmentation via generative models, addressing a practical bottleneck in medical image analysis.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): the claims of 'superior reconstruction performance' and 'retained clinically relevant discriminative features' are asserted without reference to specific quantitative metrics (e.g., PSNR/SSIM values, Dice scores, or classification AUCs), baselines (dense VAE, prior sparse methods), or ablation tables. This makes it impossible to judge whether the reported gains are robust or arise from post-hoc hyperparameter choices.
  2. [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the central premise that sparse convolutions plus attention deliver faithful 8× compression while preserving clinical signals rests on the assumption that vascular sparsity patterns are sufficiently regular and stable across patients, organs, and pathologies. No quantitative characterization of sparsity statistics, no cross-organ or cross-pathology experiments, and no failure-mode analysis on data with altered vessel density are provided; if sparsity deviates, the operators risk either under-coverage or fallback to dense behavior, directly undermining both compression and downstream utility.
minor comments (2)
  1. [§3.2] Notation for the sparse attention module should be defined explicitly (e.g., how the attention mask is derived from the sparse feature map) to avoid ambiguity when reproducing the architecture.
  2. [Figures 4–6] Figure captions for reconstruction and latent-space visualizations should include the exact compression factor, dataset split, and clinical labels used in each panel.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the claims of 'superior reconstruction performance' and 'retained clinically relevant discriminative features' are asserted without reference to specific quantitative metrics (e.g., PSNR/SSIM values, Dice scores, or classification AUCs), baselines (dense VAE, prior sparse methods), or ablation tables. This makes it impossible to judge whether the reported gains are robust or arise from post-hoc hyperparameter choices.

    Authors: We agree that the claims would be more verifiable with explicit numerical results. In the revised manuscript we will add three tables to §4: Table 1 reporting reconstruction metrics (PSNR, SSIM, volumetric Dice) for VAEsselSparse versus dense VAE and prior methods; Table 2 reporting classification AUCs for aneurysm/stenosis detection and accuracy for circle-of-Willis subvariant identification; and Table 3 presenting ablation results on the individual contributions of sparse convolutions and attention. These tables will directly support the stated performance advantages. revision: yes

  2. Referee: [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the central premise that sparse convolutions plus attention deliver faithful 8× compression while preserving clinical signals rests on the assumption that vascular sparsity patterns are sufficiently regular and stable across patients, organs, and pathologies. No quantitative characterization of sparsity statistics, no cross-organ or cross-pathology experiments, and no failure-mode analysis on data with altered vessel density are provided; if sparsity deviates, the operators risk either under-coverage or fallback to dense behavior, directly undermining both compression and downstream utility.

    Authors: We will strengthen §3 by adding quantitative sparsity statistics, including the distribution and average non-zero voxel occupancy (typically 1–5 %) across the full dataset. In §4.3 we will expand the ablation section with failure-mode experiments on subsets containing higher local vessel density due to aneurysms or stenoses, confirming that reconstruction and downstream performance remain stable. Cross-organ validation (e.g., pulmonary or hepatic vessels) is outside the scope of the present cerebral-vasculature study; we will explicitly note this limitation and identify it as future work. revision: partial

standing simulated objections not resolved
  • Cross-organ and multi-pathology experiments beyond cerebral vasculature, which would require new datasets not used in the current work.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents VAEsselSparse as an empirical encoder-decoder architecture that applies sparse convolutions and attention to exploit vascular sparsity for 8x8x8 compression. No equations, predictions, or derivations are shown that reduce reconstruction performance, latent-space discriminability, or generative utility to fitted parameters or self-definitional inputs. Claims of superiority over dense baselines and retention of clinical features (aneurysm/stenosis classification, circle-of-Willis variants) are framed as experimental outcomes rather than forced by construction. The sparsity-regularity precondition is an explicit modeling assumption, not a circular definition or self-citation load-bearing step. No uniqueness theorems, ansatzes smuggled via prior self-work, or renamings of known results appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract contains no explicit mathematical axioms, free parameters, or newly postulated entities; the method rests on standard assumptions of variational autoencoders and the empirical observation that vascular data is sparse.

pith-pipeline@v0.9.0 · 5488 in / 1139 out tokens · 31359 ms · 2026-05-09T15:18:37.462800+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

30 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    Insted: Intracranial aneurysm and intracranial artery stenosis detection and seg- mentation challenge.https://www.codabench.org/competitions/2139/(2024), accessed: 2026-02-26

    work page 2024

  2. [2]

    In: Proceedings of the MIC- CAI

    Alhonnoro, T., Pollari, M., Lilja, M., Flanagan, R., Kainz, B., Muehl, J., Mayrhauser, U., Portugaller, H., Stiegler, P., Tscheliessnigg, K.: Vessel segmenta- tion for ablation treatment planning and simulation. In: Proceedings of the MIC- CAI. pp. 45–52. Springer (2010)

    work page 2010

  3. [3]

    Vector represen- tations of vessel trees.arXiv preprint arXiv:2506.11163, 2025

    Batten, J., Schaap, M., Sinclair, M., Bai, Y., Glocker, B.: Vector representations of vessel trees. arXiv preprint arXiv:2506.11163 (2025)

    work page arXiv 2025

  4. [4]

    In: Proceedings of the MICCAI

    Chen, S., Zhang, G., Lai, J., Shen, B., Zhang, S., Dong, C., Chen, X., Li, Y.: Hier- archical part-based generative model for realistic 3d blood vessel. In: Proceedings of the MICCAI. pp. 257–267. Springer (2025)

    work page 2025

  5. [5]

    In: International Conference on Pattern Recognition

    Cheng, H., Zheng, L., Yan, Z., Zhang, H., Meng, B., Xu, X.: Fusion of machine learninganddeepneuralnetworksforpulmonaryarteriesandveinssegmentationin lung cancer surgery planning. In: International Conference on Pattern Recognition. pp. 422–438. Springer (2024)

    work page 2024

  6. [6]

    Nature Com- munications16(1), 2262 (2025)

    Chu, Y., Luo, G., Zhou, L., Cao, S., Ma, G., Meng, X., Zhou, J., Yang, C., Xie, D., Mu, D., et al.: Deep learning-driven pulmonary artery and vein segmentation reveals demography-associated vasculature anatomical differences. Nature Com- munications16(1), 2262 (2025)

    work page 2025

  7. [7]

    In: Proceedings of the MICCAI

    Feldman, P., Fainstein, M., Siless, V., Delrieux, C., Iarussi, E.: Vesselvae: Recur- sive variational autoencoders for 3d blood vessel synthesis. In: Proceedings of the MICCAI. pp. 67–76. Springer (2023)

    work page 2023

  8. [8]

    In: Proceedings of the MICCAI

    Feldman, P., Sinnona, M., Delrieux, C., Siless, V., Iarussi, E.: Vesselgpt: Autore- gressive modeling of vascular geometry. In: Proceedings of the MICCAI. pp. 662–

    work page

  9. [9]

    In: 2025 WACV

    Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V., Tang, Y., Simon, B., Belue, M., Harmon, S., Turkbey, B., et al.: Maisi: Medical ai for synthetic imaging. In: 2025 WACV. pp. 4430–4441. IEEE (2025)

    work page 2025

  10. [10]

    In: Medical Imaging with Deep Learning (2025)

    Kuipers, T.P., Konduri, P.R., Bekkers, E.J., Marquering, H.: Self-supervised syn- thetic cerebral vessel tree generation using semantic signed distance fields. In: Medical Imaging with Deep Learning (2025)

    work page 2025

  11. [11]

    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)

    work page 2024

  12. [12]

    The Lancet Digital Health4(4), e256–e265 (2022)

    Lin, A., Manral, N., McElhinney, P., Killekar, A., Matsumoto, H., Kwiecinski, J., Pieszko, K., Razipour, A., Grodecki, K., Park, C., et al.: Deep learning-enabled coronary ct angiography for plaque and stenosis quantification and cardiac risk prediction: an international multicentre study. The Lancet Digital Health4(4), e256–e265 (2022)

    work page 2022

  13. [13]

    Flow Matching for Generative Modeling

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022) 10 Prabhakar et al

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  14. [14]

    Efficient automatic segmentation for multi-level pulmonary arteries: The parse challenge

    Luo, G., Wang, K., Liu, J., Li, S., Liang, X., Li, X., Gan, S., Wang, W., Dong, S., Wang, W., et al.: Efficient automatic segmentation for multi-level pulmonary arteries: The parse challenge. arXiv preprint arXiv:2304.03708 (2023)

    work page arXiv 2023

  15. [15]

    IEEE transactions on med- ical imaging43(12), 4442–4456 (2024)

    Mou, L., Lin, J., Zhao, Y., Liu, Y., Ma, S., Zhang, J., Lv, W., Zhou, T., Liu, J., Frangi, A.F., et al.: Costa: A multi-center tof-mra dataset and a style self- consistency network for cerebrovascular segmentation. IEEE transactions on med- ical imaging43(12), 4442–4456 (2024)

    work page 2024

  16. [16]

    Medical Image Analysis97, 103253 (2024)

    Nan, Y., Xing, X., Wang, S., Tang, Z., Felder, F.N., Zhang, S., Ledda, R.E., Ding, X., Yu, R., Liu, W., et al.: Hunting imaging biomarkers in pulmonary fibrosis: benchmarks of the aiib23 challenge. Medical Image Analysis97, 103253 (2024)

    work page 2024

  17. [17]

    In: Proceedings of the MICCAI

    Prabhakar, C., Shit, S., Amiranashvili, T., Li, H.B., Menze, B.: Semantically con- sistent discrete diffusion for 3d biological graph modeling. In: Proceedings of the MICCAI. pp. 594–604. Springer (2025)

    work page 2025

  18. [18]

    In: Proceed- ings of the MICCAI

    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: Proceed- ings of the MICCAI. pp. 3–13. Springer (2024)

    work page 2024

  19. [19]

    In: Proceedings of the CVPR

    Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large- scale 3d generative modeling using sparse voxel hierarchies. In: Proceedings of the CVPR. pp. 4209–4219 (2024)

    work page 2024

  20. [20]

    Nature communications11(1), 6090 (2020)

    Shi, Z., Miao, C., Schoepf, U.J., Savage, R.H., Dargis, D.M., Pan, C., Chai, X., Li, X.L., Xia, S., Zhang, X., et al.: A clinically applicable deep-learning model for detecting intracranial aneurysm in computed tomography angiography images. Nature communications11(1), 6090 (2020)

    work page 2020

  21. [21]

    In: Proceedings of the CVPR

    Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the CVPR. pp. 16560–16569 (2021)

    work page 2021

  22. [22]

    arXiv preprint arXiv:2311.01138 (2023)

    Støverud, K.H., Bouget, D., Pedersen, A., Leira, H.O., Langø, T., Hofstad, E.F.: Aeropath: An airway segmentation benchmark dataset with challenging pathology. arXiv preprint arXiv:2311.01138 (2023)

    work page arXiv 2023

  23. [23]

    In: Proceedings of the MICRO (2023)

    Tang, H., Yang, S., Liu, Z., Hong, K., Yu, Z., Li, X., Dai, G., Wang, Y., Han, S.: Torchsparse++: Efficient training and inference framework for sparse convolution on gpus. In: Proceedings of the MICRO (2023)

    work page 2023

  24. [24]

    arXiv preprint arXiv:2502.14753 (2025)

    Varma, M., Kumar, A., Van der Sluijs, R., Ostmeier, S., Blankemeier, L., Cham- bon, P., Bluethgen, C., Prince, J., Langlotz, C., Chaudhari, A.: Medvae: Efficient automated interpretation of medical images with large-scale generalizable autoen- coders. arXiv preprint arXiv:2502.14753 (2025)

    work page arXiv 2025

  25. [25]

    In: Proceedings of the CVPR

    Wittmann, B., Wattenberg, Y., Amiranashvili, T., Shit, S., Menze, B.: vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation. In: Proceedings of the CVPR. pp. 20874–20884 (2025)

    work page 2025

  26. [26]

    arXiv preprint arXiv:1804.04381 (2018)

    Wolterink, J.M., Leiner, T., Isgum, I.: Blood vessel geometry synthesis using gen- erative adversarial networks. arXiv preprint arXiv:1804.04381 (2018)

    work page arXiv 2018

  27. [27]

    Direct3d-s2: Gigascale 3d generation made easy with spatial sparse attention.arXiv preprint arXiv:2505.17412, 2025

    Wu, S., Lin, Y., Zhang, F., Zeng, Y., Yang, Y., Bao, Y., Qian, J., Zhu, S., Cao, X., Torr, P., et al.: Direct3d-s2: Gigascale 3d generation made easy with spatial sparse attention. arXiv preprint arXiv:2505.17412 (2025)

    work page arXiv 2025

  28. [28]

    In: Proceedings of the CVPR

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: Proceedings of the CVPR. pp. 21469–21480 (2025)

    work page 2025

  29. [29]

    ArXiv pp

    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 VAEsselSparse 11 topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv pp. arXiv–2312 (2025)

    work page 2025

  30. [30]

    Medical image analysis90, 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 analysis90, 102957 (2023)

    work page 2023