Pith. sign in

REVIEW 5 major objections 5 minor 27 references

VesselSDF: Distance Field Priors for Vascular Network Reconstruction

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that vessel segmentation from sparse CT slices improves when posed as continuous signed-distance regression with an adaptive Gaussian regularizer, yielding smoother and better-connected reconstructions than binary voxel…

desk verdict Sensible two-stage SDF refinement with a new regularizer, but the table's JD metric is inconsistent and the 'significantly outperforms' claim lacks statistical support. read the letter →

arxiv 2506.16556 v1 pith:VHUFDGH5 submitted 2025-06-19 eess.IV cs.CV

Salvatore Esposito , Daniel Rebain , Arno Onken , Changjian Li , Oisin Mac Aodha This is my paper · ORCID
classification eess.IVcs.CV
keywords signeddistancefieldsvesselsegmentationvascularnetworkreconstructionsparseCTslicesadaptiveGaussianregularization3DU-Nethepaticvessels
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

VesselSDF claims that segmenting blood vessels from sparse CT slices is better posed as regression of a continuous signed distance field (an SDF, a function giving every point its signed distance to the nearest vessel surface) than as binary voxel classification, because the continuous representation encodes smooth tubular geometry and connectivity. The paper builds a two-stage pipeline: a 3D attention U-Net predicts binary occupancy, then a second, lighter U-Net converts that occupancy into an SDF, trained with SDF and occupancy supervision plus eikonal and adaptive Gaussian regularization. On the Hepatic Vessels dataset, the method reports Dice 0.72, Chamfer distance 0.68, and Hausdorff distance 4.1, improving on nnU-Net's 0.69, 0.82, and 4.9; on IRCADb it matches nnU-Net on volume metrics while improving surface metrics. If the paper is right, clinicians and downstream analysis get smoother, better-connected vessel reconstructions from the sparse slice spacing common in dose-reduced CT.

What carries the argument

The carrying mechanism is the SDF refiner $f_{SDF}(x; \theta_r) = f_r(\mathrm{detach}(f_o(x; \theta_o)); \theta_r)$ from Eq. (3), together with the adaptive Gaussian regularizer $L_{gauss} = \mathbb{E}_{x \in \Omega} |f_{SDF}(x)| \, \| f_{SDF}(x) - G_\sigma(f_{SDF}(x)) \|_2^2$. The SDF is a continuous function that gives each spatial point its signed distance to the nearest vessel surface, with the zero-level set defining the reconstructed surface. The two-stage design separates detection (binary occupancy) from geometry (distance regression), and gradient detachment prevents SDF-specific constraints from interfering with the segmentation task. The distance weight in the Gaussian term lets the network smooth noise far from vessel surfaces while preserving thin vessels, and the surface term $\exp(-\beta |f_{SDF}(x)|)$ suppresses floating fragments.

What would settle it

Take a held-out CT volume, run Stage 1, artificially remove one thin vessel branch from its occupancy output (simulating a missed detection), and check whether Stage 2's SDF still contains that branch; under Eq. (3) it cannot, so any recovery would indicate leakage, and any absence shows the refiner cannot correct Stage 1 false negatives.

Watch

Extended reading notes

Core claim

The paper's central claim is that reformulating vessel reconstruction as continuous SDF regression, with a detached two-stage refinement and an adaptive Gaussian regularizer, produces geometrically more accurate reconstructions than binary voxel classification. The SDF refiner in Eq. (3) receives only the detached occupancy from Stage 1 and regresses signed distances, so the vessel surface is the zero-level set of a smooth function rather than a set of independent voxel decisions. The adaptive Gaussian regularizer weights smoothing by the absolute SDF value, denoising far-field regions while leaving near-surface geometry intact, and the eikonal term accounts for anisotropic voxel spacing through $\gamma = \Delta z / \Delta x$. Ablations show that removing the SDF refinement lowers Dice from 0.72 to 0.69 and worsens Hausdorff distance from 4.1 to 4.4, while removing the Gaussian regularizer keeps Dice equal but degrades surface metrics.

Load-bearing premise

The load-bearing premise is that Stage 1 detects essentially all vessel voxels, because the SDF refiner receives only the detached binary occupancy and cannot restore branches that Stage 1 missed.

Editorial extensions

If this is right

  • The continuous SDF output removes the jagged, voxelized surface artifacts that binary classification produces on thin vessels.
  • The distance-weighted regularizer suppresses floating surface fragments without over-smoothing the vessel boundaries that matter for clinical measurements.
  • On IRCADb the volume metrics are indistinguishable from nnU-Net, but surface metrics improve, so the representation's benefit shows up in geometry rather than overlap.
  • The two-stage design makes the SDF refinement stable to train while letting the second stage operate on a clean detection target.

Reading between the lines

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

  • Since Stage 2 sees only detached binary occupancy, the pipeline's accuracy is capped by Stage 1 recall; a version that also feeds raw intensities or a consistency loss could recover missed branches, but that extension is not in the paper.
  • The distance-weighted smoothing is not specific to blood vessels; the same regularizer could plausibly improve reconstruction of other tubular anatomies from sparse slices, such as airways, ducts, or neural tracts, though the paper does not test this.
  • The reported connectivity gains are inferred from smoother surfaces rather than measured directly; counting connected components or centerline topology violations would turn 'preserves connectivity' into a quantitative claim.
  • The anisotropic eikonal term depends on slice spacing $\gamma$, so evaluating the model at spacings outside the training range would reveal how far the geometric prior generalizes.
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

5 major / 5 minor

Summary. The paper proposes VesselSDF, a two-stage framework for reconstructing vascular networks from sparse CT slices. Stage 1 predicts a binary occupancy map with an attention-gated 3D U-Net; Stage 2 refines the detached occupancy into a signed distance field (SDF) using a second 3D U-Net, supervised by SDF and occupancy losses plus Eikonal, distance-weighted Gaussian, and surface regularization terms. The method is evaluated on the Hepatic Vessels (MSD Task 08) and IRCADb datasets against nnU-Net, 3D SA-UNet, and 3D-UNet, reporting volume metrics (Dice, IoU, Jaccard) and surface metrics (Chamfer distance, Hausdorff distance). The central claims are that VesselSDF significantly outperforms existing methods and preserves vessel geometry and connectivity.

Significance. If the reported results are reliable, the two-stage occupancy-to-SDF refinement with adaptive geometric regularization is a plausible and potentially useful direction for thin-structure reconstruction in medical imaging. The paper contributes a clear architecture, a well-specified loss combination, and an ablation study. However, the empirical evidence as presented is not yet sufficient to support the main comparative claims: one metric is defined inconsistently, no statistical uncertainty is reported, and the ablations show only marginal gains for the flagship regularizer. The design also contains an inherent limitation (Eq. (3)) that the authors do not acknowledge. The ideas are worthy of further development, but the paper needs substantial revision before the claims can be accepted.

major comments (5)
  1. [Table 1 and Sec. 4 (Evaluation)] The JD column is defined inconsistently: the Table 1 caption calls it 'Jaccard similarity (JD)' with an up-arrow (higher better), while the Evaluation paragraph calls it 'Jaccard Distance (JD)' (lower better). The reported values (0.48/0.45/0.38/0.31 for VesselSDF/nnU-Net/SA-UNet/3D-UNet on Hepatic Vessels) match neither the adjacent IoU column (0.59/0.56/0.51/0.44, since Jaccard similarity equals IoU) nor 1-IoU (0.41/0.44/0.49/0.56, which would be the Jaccard distance). This metric is therefore undefined and unreproducible, and the claimed superiority on this column cannot be assessed. The authors must clarify the definition, correct the values, and ensure consistency between caption, text, and arrows.
  2. [Sec. 4.1 and Abstract] The abstract's unqualified claim that VesselSDF 'significantly outperforms existing methods' is not supported by the reported experiments. No standard deviations, confidence intervals, significance tests, or per-case results are provided anywhere. On IRCADb, the Dice, IoU, and JD values are exactly identical to nnU-Net (0.86, 0.82, 0.75), and the text in Sec. 4.1 itself concedes 'comparable performance according to volume-based metrics on IRCADb.' The statistical claim must be either substantiated with proper testing or removed and replaced with a qualified statement that separates Hepatic Vessels surface-metric gains from IRCADb volume-metric ties.
  3. [Sec. 3.3, Eq. (3)] Eq. (3) defines the SDF refiner as receiving only the detached binary occupancy from Stage 1, with no access to the original CT intensities. Any vessel missed or mislocalized by Stage 1 therefore cannot be corrected by Stage 2. This directly undermines the introduction's claim that the 'efficient SDF refinement strategy' ensures 'robust reconstruction even in challenging cases with significant inter-slice gaps' (Sec. 1). The paper should acknowledge this limitation and, ideally, provide an experiment quantifying how much Stage 2 can or cannot recover from Stage 1 false negatives.
  4. [Sec. 3.4, Eq. (9)] The surface regularization term Lsur = E[exp(-β|fSDF(x)|)] is largest when fSDF(x) = 0, i.e., exactly on the true vessel surface. As written, minimizing this term pushes the SDF away from zero everywhere, penalizing genuine surfaces as much as spurious floating fragments. The text says it suppresses 'floating' components when there is 'no strong evidence of an actual surface,' but the equation contains no evidence-dependent weighting. The authors should either correct the formulation or explain how the balance with supervised terms prevents this term from distorting true vessel boundaries.
  5. [Table 2 (Ablations)] The ablation of the Gaussian loss shows no change in Dice, IoU, or JD (0.72/0.59/0.48 for both VesselSDF and VesselSDF w/o Gaussian Loss), and only small changes in CD (0.68 vs 0.70) and HD (4.1 vs 4.3). This provides weak quantitative support for the paper's central claim that the adaptive Gaussian regularizer 'eliminates floating artifacts' and 'preserves fine vessel geometry.' No direct artifact metric (e.g., number of connected components or floating fragments) is reported. The authors should either add such a metric or temper the claim about the Gaussian regularizer's contribution.
minor comments (5)
  1. [Sec. 4 (Evaluation)] The text says 'Jaccard Distance (JD) quantifies topological similarity,' but the table caption says 'Jaccard similarity (JD).' Please pick one name and use it consistently throughout, including the arrow direction.
  2. [Sec. 3.4, Eqs. (5)-(9)] The expectation notation E_{x in Ω} is never made precise: are points sampled uniformly, near the surface, or with some importance weighting? The paper should describe the sampling strategy, as it affects the loss balance and the interpretation of the regularizers.
  3. [Sec. 3.2 and Sec. 3.3] The sentence 'Each level operates at a different spatial resolution through successive downsampling and upsampling operations...' appears twice in Sec. 3.3. Please remove the duplication.
  4. [Sec. 4 (Implementation)] The paper states that training SDFs are computed from binary ground truth but does not specify the distance transform method, truncation distance, or how anisotropic voxel spacing is handled in the supervision. This is needed for reproducibility.
  5. [Fig. 2] The qualitative figure would benefit from clearer labels, a scale bar, and an explicit visual indication of the claimed floating artifacts or connectivity improvements, so that the qualitative claims can be independently assessed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: VesselSDF is an empirical method paper; its claims rest on held-out comparison and ablations, not on a prediction that reduces to a fitted input or on a load-bearing self-citation.

full rationale

The paper's central claim is that its two-stage SDF refinement outperforms binary voxel segmentation baselines on hepatic vessel datasets. This claim is supported by quantitative comparisons in Table 1 on held-out test data and by ablations in Table 2. No mathematical derivation is presented that would create a self-definitional loop: the SDF supervision (Eq. 5) uses ground-truth SDFs computed from binary ground truth, and the occupancy loss (Eq. 6) is standard supervised cross-entropy. The regularizers (Eikonal, Gaussian, surface) are hand-specified geometric loss terms whose contributions are evaluated by removing them; they are not fitted to the test metric and then reported as predictions. The two-stage architecture's limitation that Stage 2 receives detached occupancy is an architectural choice, not a circular step. The paper does not rely on a self-citation to establish its method's validity; references to prior SDF work are contextual. While Table 1's JD column is internally inconsistent and the word 'significantly' lacks statistical support, these are correctness or reporting concerns, not circularity. The method is tested against external baselines and datasets, so no component of the claimed result is forced by construction or by an input-equivalent definition.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The central claim is empirical and rests on standard deep learning training assumptions plus hand-chosen hyperparameters. No new physical or mathematical entities are introduced.

free parameters (3)
  • Loss weights lambda_s, lambda_o, lambda_g, lambda_e, lambda_r = 0.1, 0.01, 0.1, 0.01, 0.1
    Set by hand in Section 3.4. No sensitivity analysis or tuning procedure is reported.
  • Gaussian blur width sigma = not stated (only sigma > 0)
    Appears in Eq. (8) but no value or selection procedure is reported.
  • Surface penalty beta = not stated (only beta > 0)
    Appears in Eq. (9) but no value or selection procedure is reported.
assumptions (2)
  • domain assumption Ground-truth binary vessel masks can be converted into a well-defined signed distance field used as supervision.
    Section 4 says 'We compute training SDFs from the binary ground truth' but gives no conversion algorithm and does not address errors or discretization in the masks.
  • domain assumption Vessel geometry is well represented by smooth signed distance functions with near-unit gradients.
    The Eikonal term in Eq. (7) assumes a true distance field, but thin vessels and centerlines have gradient discontinuities that may violate this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VesselSDF: Distance Field Priors for Vascular Network Reconstruction." pith.science (2026). https://pith.science/paper/VHUFDGH5

@misc{pith2026250616556,
  author       = {Pith},
  title        = {Pith review of: VesselSDF: Distance Field Priors for Vascular Network Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHUFDGH5}},
  note         = {Machine review of arXiv:2506.16556}
}
read the original abstract

Accurate segmentation of vascular networks from sparse CT scan slices remains a significant challenge in medical imaging, particularly due to the thin, branching nature of vessels and the inherent sparsity between imaging planes. Existing deep learning approaches, based on binary voxel classification, often struggle with structural continuity and geometric fidelity. To address this challenge, we present VesselSDF, a novel framework that leverages signed distance fields (SDFs) for robust vessel reconstruction. Our method reformulates vessel segmentation as a continuous SDF regression problem, where each point in the volume is represented by its signed distance to the nearest vessel surface. This continuous representation inherently captures the smooth, tubular geometry of blood vessels and their branching patterns. We obtain accurate vessel reconstructions while eliminating common SDF artifacts such as floating segments, thanks to our adaptive Gaussian regularizer which ensures smoothness in regions far from vessel surfaces while producing precise geometry near the surface boundaries. Our experimental results demonstrate that VesselSDF significantly outperforms existing methods and preserves vessel geometry and connectivity, enabling more reliable vascular analysis in clinical settings.

Figures

Figures reproduced from arXiv: 2506.16556 by the authors.

Figure 1
Figure 1. Overview of VesselSDF- our two-stage approach for vessel segmentation and reconstruction from CT scans. In the first stage, a 3D U-Net predicts a binary occupancy map. The second stage refines this occupancy into a signed distance field (SDF) using an additional 3D U-Net, guided by geometric regularization terms. The output 3D SDF, converted into a mesh, results in high-quality reconstructed vessels. 3 Method 3.1 Pr… view at source ↗
Figure 2
Figure 2. Qualitative 3D reconstruction results on the Hepatic Vessels dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    In: Statistical Atlases and Computational Models of the Heart (2022)

    Alblas, D., Brune, C., Yeung, K.K., Wolterink, J.M.: Going off-grid: Continuous implicit neural representations for 3d vascular modeling. In: Statistical Atlases and Computational Models of the Heart (2022)

    work page 2022

  2. [2]

    Nature communications (2022)

    Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nature communications (2022)

    work page 2022

  3. [3]

    In: DAGM (2023)

    Bogensperger, L., Narnhofer, D., Ilic, F., Pock, T.: Score-based generative models for medical image segmentation using signed distance functions. In: DAGM (2023)

    work page 2023

  4. [4]

    In: MICCAI (2021)

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. In: MICCAI (2021)

    work page 2021

  5. [5]

    In: MICCAI (2016)

    Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u- net: learning dense volumetric segmentation from sparse annotation. In: MICCAI (2016)

    work page 2016

  6. [6]

    In: MICCAI (2023)

    Dima, A., Zimmer, V., Menten, M., et al.: 3d arterial segmentation via single 2d projections and depth supervision in contrast-enhanced ct images. In: MICCAI (2023)

    work page 2023

  7. [7]

    In: ICML (2020)

    Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric reg- ularization for learning shapes. In: ICML (2020)

    work page 2020

  8. [8]

    3D SA-UNet: 3D Spatial Attention UNet with 3D Atrous Spatial Pyramid Pooling for White Matter Hyperintensities Segmentation

    Guo, C.: 3d sa-unet: 3d spatial attention unet with 3d aspp for white matter hyperintensities segmentation. arXiv:2309.08402 (2023)

    work page Pith review arXiv 2023

Show all 27 references
  1. [9]

    Nature Methods (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods (2021)

    2021

  2. [10]

    In: Intelligent medicine (2023)

    Li, J., et al.: A hybrid network integrating convolution and transformer for thy- moma segmentation. In: Intelligent medicine (2023)

    2023

  3. [11]

    ICML (2021)

    Ma, B., Han, Z., Liu, Y.S., Zwicker, M.: Neural-pull: Learning signed distance function from point clouds by learning to pull space onto surface. ICML (2021)

    2021

  4. [12]

    Maas, K.W., Pezzotti, N., Vermeer, A.J., Ruijters, D., Vilanova, A.: Nerf for 3d re- constructionfromx-rayangiography:Possibilitiesandlimitations.In:Eurographics Workshop on Visual Computing for Biology and Medicine (2023)

    2023

  5. [13]

    arXiv preprint arXiv:1804.03999 (2018)

    Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

    2018 arXiv

  6. [14]

    CVPR (2019)

    Park,J.J.,Florence,P.,Straub,J.,Newcombe,R.,Lovegrove,S.:Deepsdf:Learning continuous signed distance functions for shape representation. CVPR (2019)

    2019

  7. [15]

    In: MICCAI (2015)

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI (2015)

    2015

  8. [16]

    Transactions on Medical Imaging (2024)

    Shaker,A.M.,Maaz,M.,Rasheed,H.,Khan,S.,Yang,M.H.,Khan,F.S.:Unetr++: Delving into efficient and accurate 3d medical image segmentation. Transactions on Medical Imaging (2024)

    2024

  9. [17]

    In: MICCAI (2024)

    Shi, P., Hu, J., Yang, Y., Gao, Z., Liu, W., Ma, T.: Centerline boundary dice loss for vascular segmentation. In: MICCAI (2024)

    2024

  10. [18]

    In: CVPR (2021)

    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: CVPR (2021)

    2021

  11. [19]

    IRCAD Tech Report (2010)

    Soler, L., Hostettler, A., Agnus, V., Charnoz, A., Fasquel, J., Moreau, J., Oss- wald, A., Bouhadjar, M., Marescaux, J.: 3d image reconstruction for comparison VesselSDF: Distance Field Priors for Vascular Network Reconstruction 11 of algorithm database: A patient specific ana...

    2010

  12. [20]

    Fron- tiers in Neuroscience (2020)

    Tetteh, G., Efremov, V., Forkert, N., et al.: Deepvesselnet: Vessel segmentation, centerline prediction, and bifurcation detection in 3d angiographic volumes. Fron- tiers in Neuroscience (2020)

    2020

  13. [21]

    arXiv:2411.17386 (2024)

    Wittmann, B., Wattenberg, Y., Amiranashvili, T., Shit, S., Menze, B.: vesselfm: A foundation model for universal 3d blood vessel segmentation. arXiv:2411.17386 (2024)

    2024 arXiv

  14. [22]

    Medical Image Analysis (2025)

    Xu, J., Dong, A., Yang, Y., Jin, S., Zeng, J., Xu, Z., Jiang, W., Zhang, L., Dong, J., Wang, B.: Vsnet: Vessel structure-aware network for hepatic and portal vein segmentation. Medical Image Analysis (2025)

    2025

  15. [23]

    Research Square (2024)

    Yagis, E., Aslani, S., Jain, Y., Zhou, Y., Rahmani, S., Brunet, J., Bellier, A., Werlein, C., Ackermann, M., Jonigk, D., et al.: Deep learning for 3d vascular seg- mentation in phase contrast tomography. Research Square (2024)

    2024

  16. [24]

    In: MICCAI (2024)

    Zhang, Z.: Efficient slice anomaly detection network for 3d brain mri volume. In: MICCAI (2024)

    2024

  17. [25]

    Multimedia Systems (2019)

    Zhao, F., Chen, Y., Hou, Y., He, X.: Segmentation of blood vessels using rule-based and machine-learning-based methods: A review. Multimedia Systems (2019)

    2019

  18. [26]

    Transactions on Image Processing (2023)

    Zhou, H.Y., Guo, J., Zhang, Y., Han, X., Yu, L., Wang, L., Yu, Y.: nnformer: Volumetric medical image segmentation via a 3d transformer. Transactions on Image Processing (2023)

    2023

  19. [27]

    Transactions on Medical Imaging (2020)

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. Transactions on Medical Imaging (2020)

    2020

Pith tools

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