{"id":"d44808fb-70fc-45ec-ab4c-7341f5a31e10","arxiv_id":"2506.16556","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage neural network that turns sparse CT slices into a signed distance field improves vascular reconstruction over voxel-classification baselines on hepatic vessel datasets.","lead":"VesselSDF reconstructs blood vessels from sparse CT scans by first finding vessel voxels with a U-Net, then refining them into a continuous signed distance field that describes vessel surfaces. It reports better surface accuracy than standard segmentation networks on two liver vessel datasets, though volume scores are similar on one dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's JD column is defined inconsistently (caption: 'Jaccard similarity'; text: 'Jaccard Distance'; arrow: higher-better) and matches neither IoU nor 1-IoU; with no error bars or significance tests, the abstract's 'significantly outperforms' is not supported by the reported experiments.","rationale":"The paper's goal is to show that reformulating vessel segmentation as two-stage occupancy-then-SDF regression with an adaptive Gaussian regularizer yields measurably better reconstructions. The most load-bearing assumption is that Table 1, as reported, substantiates a definite comparative claim. That assumption is insecure for two concrete reasons: (1) the JD column is defined inconsistently - the caption calls it 'Jaccard similarity (JD)' (which would equal the IoU column), the Evaluation paragraph calls it 'Jaccard Distance (JD)' (which would be 1-IoU), and the arrow says higher-better while the values match neither definition; a reader cannot tell what was computed, and under a distance reading VesselSDF is worst on that column; (2) the word 'significantly' in the abstract is unsupported because no standard deviations, per-case results, significance tests, or test-sample sizes are given, and the margins are small (Dice 0.72 vs 0.69, with identical IRCADb volume metrics). I checked the reader's alternative concern - the detached-occupancy bottleneck of Eq. (3) - and while it is a genuine ceiling (Stage 2 cannot add vessels that Stage 1 missed, and the Lsur term of Eq. (9) actively discourages surfaces unsupported by the occupancy input), Table 2's ablation shows the SDF stage still adds +0.03 Dice and better CD/HD, so the bottleneck is a limitation rather than a contradiction of the measured gains. Credit is due for the component ablations, the use of two public datasets, and the anisotropic Eikonal term of Eq. (7); the unreported sigma and beta hyperparameters for Eqs. (8)-(9) further block reproduction. The verdict should remain CONDITIONAL: fix the JD definition, add paired statistics, report sigma/beta and dataset splits, and release code. My read agrees with the reader's verdict but identifies the metric/statistics inconsistency, rather than the architectural bottleneck, as the decisive weakness.","tokens_in":7202,"tokens_out":16000,"duration_ms":173600,"concrete_test":"On the Hepatic Vessels test split, recompute all five Table 1 metrics per volume with standard definitions: compute JD two ways - per-case Jaccard index (then averaged) and 1-IoU (global) - and check whether either definition reproduces the reported JD column with its stated arrow direction; then run a paired Wilcoxon signed-rank test on per-volume Dice, Chamfer distance, and Hausdorff distance for VesselSDF vs nnU-Net, reporting the test-set size n. If the JD column cannot be reproduced under either standard definition, or if the paired differences are not significant at p<0.05, the abstract's 'significantly outperforms' should be weakened to 'consistently improves surface metrics, pending statistical validation.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an empirical comparison: the abstract says VesselSDF 'significantly outperforms existing methods,' and Sec. 4.1 says 'we outperform all methods' on Hepatic Vessels. The load-bearing evidence is Table 1, and it is not internally consistent. The JD column is defined three incompatible ways: the Table 1 caption calls it 'Jaccard similarity (JD)' with an up-arrow (higher better); the Evaluation paragraph calls it 'Jaccard Distance (JD)' (lower better); and the reported values (0.48/0.45/0.38/0.31 across VesselSDF/nnU-Net/SA-UNet/3D-UNet) reconcile with 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 is what 'Jaccard Distance' should be. Under a distance reading, VesselSDF is actually the worst performer on this column of its own table. Because the metric is undefined and unreproducible, part of Table 1 cannot validate the claimed superiority. Separately, 'significantly' has no statistical content: the Dice margin is 0.03 (0.72 vs 0.69), the IRCADb volume metrics tie nnU-Net exactly (0.86/0.82/0.75), and no per-case results, standard deviations, significance tests, or test-split sizes are reported anywhere. The load-bearing premise - that the experiments as reported support a definite comparative claim - is therefore insecure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7615,"tokens_out":3725,"duration_ms":42695,"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":[{"comment":"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.","section":"Table 1 and Sec. 4 (Evaluation)"},{"comment":"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.","section":"Sec. 4.1 and Abstract"},{"comment":"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.","section":"Sec. 3.3, Eq. (3)"},{"comment":"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.","section":"Sec. 3.4, Eq. (9)"},{"comment":"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.","section":"Table 2 (Ablations)"}],"minor_comments":[{"comment":"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.","section":"Sec. 4 (Evaluation)"},{"comment":"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.","section":"Sec. 3.4, Eqs. (5)-(9)"},{"comment":"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.","section":"Sec. 3.2 and Sec. 3.3"},{"comment":"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.","section":"Sec. 4 (Implementation)"},{"comment":"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.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a promising core idea but the empirical reporting is not yet at the standard required for the claims. The metric inconsistency in Table 1 and the absence of any significance testing are the main blockers. I would encourage the editor to ask for a revision that fixes the metric definitions, adds uncertainty quantification, and acknowledges the Stage-1 dependency explicitly. The design is not fatally flawed, so rejection seems too strong at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about VesselSDF: it is a reasonable two-stage pipeline that treats vessel segmentation as occupancy followed by SDF regression, and its distance-weighted Gaussian regularizer is a genuinely new idea. But the headline claim of significant improvement is not backed by the table as printed: the JD column is defined one way in the caption, another in the text, and the numbers match neither IoU nor 1-IoU.\n\nWhat's actually novel: detaching the binary occupancy and feeding it to a second, lighter 3D U-Net that regresses an SDF, with an adaptive Gaussian smoothness term weighted by |fSDF|, plus an anisotropic Eikonal term that accounts for slice spacing. That is a sensible combination of existing components, and the ablation shows SDF refinement helps (Dice 0.69 to 0.72). On the Hepatic Vessels dataset VesselSDF improves over nnU-Net across all five metrics, and on IRCADb the surface metrics are clearly better (CD 0.60 vs 0.75, HD 3.5 vs 4.2). So there is real, modest progress here.\n\nThe soft spots are real but mostly fixable. The JD metric inconsistency is not a minor typo: the values in Table 1 and Table 2 cannot be interpreted as either Jaccard similarity or Jaccard distance. That undermines part of the evidence. There are no standard deviations, confidence intervals, or significance tests anywhere, so \"significantly outperforms\" has no statistical content. On IRCADb the volume metrics tie nnU-Net exactly (0.86/0.82/0.75), which sits awkwardly with the abstract. The paper also omits the values of sigma and beta in the regularizers, and gives no test-split details. The detached occupancy in Eq. 3 means Stage 2 cannot recover vessels missed by Stage 1, so the method's success depends on Stage 1 being nearly complete; that should be discussed. No code is released.\n\nThis is an honest engineering contribution to a narrow subfield. A reader working on vascular reconstruction from sparse CT will find the two-stage recipe and the regularizer worth trying. It deserves peer review after a revision that fixes the table, adds uncertainty estimates, and releases code. I would not cite it in its current form.","headline":"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.","tokens_in":8121,"tokens_out":2772,"would_cite":false,"duration_ms":27854,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["signed distance fields","vessel segmentation","vascular network reconstruction","sparse CT slices","adaptive Gaussian regularization","3D U-Net","hepatic vessels"],"falsifier":"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.","tokens_in":7035,"feed_emoji":"🩻","tokens_out":9180,"duration_ms":85249,"temperature":0.7,"pith_summary":"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.","feed_headline":"Distance fields beat voxel labels for vessel reconstruction","feed_subtitle":"Distance regression yields smoother, better-connected 3D vessel networks from sparse CT slices.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Motivates the continuous implicit SDF representation for vascular modeling that VesselSDF builds on.","marker":"[1]"},{"why":"Supplies the eikonal regularization term used to enforce near-unit SDF gradients.","marker":"[7]"},{"why":"The strongest baseline; VesselSDF reports better Dice and surface distances against it on Hepatic Vessels.","marker":"[9]"},{"why":"Provides the Hepatic Vessels dataset used for the main evaluation and ablations.","marker":"[2]"},{"why":"Provides the IRCADb dataset used for the second evaluation.","marker":"[19]"},{"why":"A baseline 3D U-Net architecture compared in Table 1.","marker":"[5]"},{"why":"A baseline 3D SA-UNet architecture compared in Table 1.","marker":"[8]"},{"why":"Source of the attention gates used in the Stage 1 occupancy network.","marker":"[13]"},{"why":"Foundational U-Net encoder-decoder design that the occupancy stage is built on.","marker":"[15]"}],"fun_headline_variants":["VesselSDF: SDF regression beats voxel classification","No floating segments: VesselSDF uses SDF priors","Distance fields enhance vessel reconstruction","From voxels to distance fields for better vasculature"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["VesselSDF: SDF regression beats voxel classification","No floating segments: VesselSDF uses SDF priors","Distance fields enhance vessel reconstruction","From voxels to distance fields for better vasculature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000356,"raw_usage":{"total_tokens":1916,"prompt_tokens":913,"completion_tokens":1003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":940}},"tokens_in":529,"tokens_out":1003,"duration_ms":11089,"temperature":1.0,"reasoning_tokens":940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:36:41.756374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"In: Statistical Atlases and Computational Models of the Heart (2022)","cited_arxiv_id":null,"evidence_quote":"Motivates the continuous implicit SDF representation for vascular modeling that VesselSDF builds on."},{"cited_title":"In: ICML (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the eikonal regularization term used to enforce near-unit SDF gradients."},{"cited_title":"Nature Methods (2021)","cited_arxiv_id":null,"evidence_quote":"The strongest baseline; VesselSDF reports better Dice and surface distances against it on Hepatic Vessels."},{"cited_title":"Nature communications (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the Hepatic Vessels dataset used for the main evaluation and ablations."},{"cited_title":"IRCAD Tech Report (2010)","cited_arxiv_id":null,"evidence_quote":"Provides the IRCADb dataset used for the second evaluation."},{"cited_title":"In: MICCAI (2016)","cited_arxiv_id":null,"evidence_quote":"A baseline 3D U-Net architecture compared in Table 1."},{"cited_title":"3D SA-UNet: 3D Spatial Attention UNet with 3D Atrous Spatial Pyramid Pooling for White Matter Hyperintensities Segmentation","cited_arxiv_id":"2309.08402","evidence_quote":"A baseline 3D SA-UNet architecture compared in Table 1."}],"review_version":1}