{"id":"1cefe3bf-83cc-4817-92ef-a0e7d3c95805","arxiv_id":"2506.05820","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deformable centerline graph representation, refined by a cascaded transformer-based deformation process, improves 3D vessel segmentation and centerline extraction.","lead":"This paper proposes a new way to represent blood vessels in 3D scans, as connected centerline points that are gradually reshaped to match the real vessels. It reports better segmentation and centerline quality than standard mask-based methods on four medical datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deformation never adds/removes branches: final topology is fixed by the coarse-segmentation skeleton, so missed or spurious branches cannot be corrected, undercutting topology/noise claims.","rationale":"The reader's weakest_assumption already identifies this, and I concur. The deformable representation's topology is frozen at initialization; deformation only changes geometry. This is the most load-bearing concern because it directly affects the central claim's mechanism: the paper argues that the continuous representation fixes fractures and filters noise, but in reality it can only refine whatever topology the coarse-segmentation skeleton already has. The unfair centerline comparison is a separate experimental-design issue that affects the strength of the quantitative evidence, but it does not challenge the internal logic as directly. If the topology is inherited, then the representation's claimed advantages are largely inherited from the backbone's coarse segmentation, and the reported Betti/clDice gains may reflect better coarse segmentation (trained with additional losses) rather than the deformable process. The concrete branch-matching test would establish whether the deformation ever contributes new branches; if it does not, the paper should be read as a coarse-segmentation-skeleton refinement method, and the scope of the central claim should be narrowed accordingly. The verdict remains CONDITIONAL, matching the reader's assessment, because the method may still be practically useful for well-visible vessels and the segmentation improvements are real even if the mechanism is mischaracterized.","tokens_in":15837,"tokens_out":6340,"duration_ms":66971,"concrete_test":"On HaN-Seg or ImageCAS, compute the skeleton of the coarse segmentation before deformation and the final centerline for each test image. Match branches within a 3-voxel tolerance. Report (a) the fraction of final branches matched to a coarse-skeleton branch and (b) the fraction of coarse-skeleton branches persisting to the output. Also replace the learned deformation stages with Laplacian smoothing of the skeleton and compare β0/β1 errors. If (a)≈1 and the smoothed-skeleton baseline matches DeformCL's topological errors, the deformation adds no topology, confirming the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The initial template is the coarse-segmentation skeleton (Sec 3.2). Deformation (Eq. 4) only moves points; unpooling (Eq. 5) only inserts midpoints on existing edges. Neither adds nor removes edges or branches. Thus any vessel branch absent from the coarse segmentation is absent forever, and spurious branches persist. Supplementary Fig. 13's fragmented predictions on occluded vessels confirm this. This undercuts 'noise robustness' (cannot prune false branches) and 'natural connectivity' (cannot recover missed branches). Consequently, the topological gains in Table 1 are inherited from the coarse skeleton, not produced by the deformable representation. The central claim that the continuous representation yields better topology is not established; at best, the full pipeline refines the geometry of a fixed-topology skeleton.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeformCL, a continuous representation of 3D vessels as graphs of deformable centerline points. Starting from a coarse segmentation produced by the backbone, the method skeletonizes the coarse mask, builds a minimum-spanning-tree graph, selects k control points, and interpolates an initial centerline template. A cascade of Transformer feature aggregation, deformation, and unpooling layers then refines the point positions. The final centerline points are converted into a distance map that is fused with the coarse segmentation to produce a refined mask. The method is evaluated on HaN-Seg, HNCTA, ASOCA, and ImageCAS, with three base segmentors (softDice, clDice, DSCNet), reporting segmentation Dice/clDice, topological errors, HD95, and centerline F1/HD/Chamfer/beta0, plus qualitative CPR visualizations.","tokens_in":15995,"tokens_out":4575,"duration_ms":44783,"significance":"If its central claims are validated, DeformCL would be a practically useful representation: it outputs centerlines directly, supports feature aggregation along tubular curves, and improves segmentation connectivity. The manuscript has clear strengths: the code is released; the pipeline is specified in enough detail to reimplement; experiments cover four datasets and multiple backbones; ablations test template control-point count, loss components, interpolation, and interaction mechanism; and a Swin UNETR backbone is included. The clinical CPR visualization is a relevant and informative qualitative contribution. The main weakness is that the topology of the final centerline is fixed by the coarse-segmentation skeleton, which limits the claimed noise robustness and natural connectivity and makes the topological improvements in the main tables hard to attribute to the deformable representation.","major_comments":[{"comment":"The deformation and unpooling operations only move existing points and insert midpoints on existing edges; no operation adds or removes branches. Because the initial template is derived from the coarse segmentation skeleton, the final centerline topology is predetermined by the coarse segmentation. Therefore the method cannot recover a vessel branch that is absent from the coarse segmentation, nor can it prune a spurious branch that the coarse segmentation contains. This directly weakens the claims of 'noise robustness' and 'natural connectivity' and is consistent with the fragmented predictions on occluded vessels shown in Supplementary Fig. 13. To support the central claim, the authors should quantify, on the test sets, how often ground-truth branches are missing from the coarse segmentation and report branch-level recall/precision of the final centerlines relative to the coarse skeleton. Without such analysis, the topological gains in Table 1 cannot be separated from the quality of the initial skeleton.","section":"Sec. 3.2, Eqs. (4)-(5), Supplementary Fig. 13"},{"comment":"The centerline comparison is not apples-to-apples because the baseline methods are segmentation models that were not trained for centerline extraction, and the procedure used to derive their centerlines from the predicted masks is not described. A common protocol (same skeletonization algorithm, same tolerance, same matching criterion) is essential for F1, HD95, and Chamfer comparisons. Moreover, the claim of consistent centerline improvement is contradicted by the HaN-Seg row: Ours(DSCNet) reports HD95=8.222, which is worse than DSCNet's 7.462. The authors should specify the centerline generation procedure for all methods, release the evaluation script, and either train the baselines with a centerline-aware head or restrict the centerline claims to comparisons among methods that produce centerlines natively.","section":"Table 2"},{"comment":"No standard deviations, confidence intervals, or significance tests are reported. This is particularly important because the test sets are small (11 images for HaN-Seg and 7 for ASOCA), and several reported improvements are around 1 Dice point or less (e.g., Ours(clDice) on ImageCAS improves Dice by only 0.07). Without variance estimates, the claim of 'effectiveness and superiority' is not statistically supported. The authors should report per-subject variability or fold-based intervals and apply paired tests to the Dice, clDice, Betti, and centerline metrics where small numbers of branches can drive the average.","section":"Tables 1 and 2"},{"comment":"The ablation of loss functions reports only segmentation Dice and clDice, yet the text argues that SDF loss and regularization loss improve centerline centrality and smoothness. These claims are qualitative and based on Fig. 7. The ablation should report the centerline metrics (F1, HD95, Chamfer, beta0) for each loss configuration, because the paper's central contribution is the centerline representation and its downstream effect on segmentation. Without these numbers, the contribution of L_sdf and L_reg to centerline quality is not quantitatively established.","section":"Sec. 4.5, Table 4"}],"minor_comments":[{"comment":"There are several typos: 'strcutures' in the SDF Loss paragraph, 'neightbour' in Eq. (8), 'analusis' in the caption of Table 5, and 'planner' should be 'planar' in Sec. 3.1.","section":"Throughout"},{"comment":"The domain of the distance map is declared as R^{H*D*W}, but the input volume is described as R^{H*W*D}; please make the dimension order consistent.","section":"Eq. (6)"},{"comment":"The overview figure is dense and hard to read at print size; the stage weights and loss labels should be enlarged or split into a separate figure.","section":"Figure 3"},{"comment":"The sentence 'We have not included PointScatter [31] and Topoloss [9] in the comparison as it is designed for 2D images' has an ambiguous pronoun; clarify that both methods are designed for 2D and are not applicable to 3D volumes under the reported GPU memory constraints.","section":"Sec. 4.4"},{"comment":"The Transformer architecture is not fully specified in the main text; please state the number of layers, heads, and hidden dimensions, or refer the reader to a precise supplementary section.","section":"Sec. 3.2, Eq. (2)-(4)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a medical-imaging journal and the code release is commendable. The main risk is overclaiming topological advantages from the deformable representation when the topology is inherited from the coarse segmentation. The fixed-topology limitation should be addressed head-on with branch-level analysis rather than avoided. The centerline evaluation protocol and the absence of statistical significance reporting also need attention before the paper can be accepted. I would not require new datasets, but the requested analyses are necessary to substantiate the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid method-level paper, not a breakthrough. The new thing is applying the deformable-template idea from mesh generation (Pixel2Mesh, Voxel2Mesh) to centerline graphs for 3D vessel extraction, with an adaptive per-image template built by skeletonizing the coarse segmentation, then cascaded transformer deformation and unpooling stages trained with local chamfer, SDF, and regularization losses. That combination is new, the pipeline is specified well enough to reimplement, and the code is released. The ablations actually test the design choices—template control points, loss components, interaction module, interpolation method—and the extra compute is modest.\n\nThe main soft spot is the one the stress-test identifies: deformation only moves points and unpooling only inserts midpoints on existing edges. The graph topology, including branches, is fixed by the coarse segmentation skeleton. The method cannot prune a spurious branch or recover a missed one. So the 'noise robustness' and 'natural connectivity' claims are overstated, and the topological gains in Table 1 are largely inherited from the coarse skeleton rather than produced by the deformable representation. The authors' own failure case—fragmented predictions on nearly occluded vessels in Supplementary Fig. 13—confirms this. It is a real limitation, but it does not sink the paper; the method still refines geometry, improves centerline placement, and helps segmentation. It should be stated as a limitation instead of being waved away.\n\nTwo other things. First, the centerline comparison in Table 2 is not apples-to-apples: the baselines are segmentation models that were not trained to extract centerlines, and the paper never says how their outputs were converted to centerlines. Second, GraphMorph [40], the closest prior graph-morphing work from the same group, is cited but not compared. No standard deviations are reported anywhere, and some individual metrics move little or slightly negative (e.g., clDice on ImageCAS for the clDice-based variant is +0.01, and HD95 is worse in a couple of cases). These are fixable with a revision.\n\nThe central argument—that a continuous deformable centerline representation can serve as an effective head for vessel segmentation—holds up well enough to take seriously. The paper deserves a serious referee. With fair baselines, a stated topology limitation, and error bars, it would be a strong conference paper. I'd send it to review.","headline":"DeformCL is a well-specified method-level advance in deformable centerline representation, but its topology claims are overstated because the graph cannot add or remove branches; still worth a serious referee.","tokens_in":16541,"tokens_out":2573,"would_cite":true,"duration_ms":25728,"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":"Deformable centerline graphs outperform per-pixel masks for 3D vessel extraction.","keywords":["deformable centerline","vessel extraction","3D medical image","graph representation","tubular structure","centerline segmentation","deformable template","topology preservation"],"falsifier":"Compute per-branch centerline recall of the coarse segmentation and of DeformCL on a dataset with partially occluded vessels; if a branch entirely absent from the coarse mask appears in the final prediction, the deformation can create branches, whereas if none ever appear, the method's recall is capped by the coarse mask's completeness.","tokens_in":15630,"feed_emoji":"🫀","tokens_out":7435,"duration_ms":61312,"temperature":0.7,"pith_summary":"This paper argues that the standard approach to vessel extraction in 3D medical images—classifying every voxel into a mask—is the root cause of fragmented, noisy, and topologically wrong vessel predictions. It introduces DeformCL, a continuous representation in which vessel centerlines are stored as a graph of points connected by edges, and trains a cascaded network that deforms a rough initial template to fit the true vessels. The authors report consistent gains over strong mask-based baselines on four datasets, with the largest improvements in connectivity-sensitive metrics such as clDice and Betti number errors. Because the network outputs structured centerlines directly, the result can be used for curved planar reformation without a separate skeletonization post-processing step. The paper is trying to establish that a connected, deformable representation is inherently better matched to curvilinear anatomy than discrete mask prediction.","feed_headline":"Deformable centerline graphs beat masks for 3D vessel extraction","feed_subtitle":"A centerline graph, refined by cascaded deformation, lifts connectivity and topology scores on four datasets.","key_machinery":"The machinery is the deformable centerline graph: a set of 3D points connected by edges, initialized per image from the coarse segmentation, and refined by a cascade of deformation stages. Each stage combines trilinear feature sampling, a Transformer over the graph for global context, an MLP with tanh activation for bounded point movement, and unpooling that inserts midpoints between connected points. Three auxiliary losses shape the deformation: a local Chamfer loss restricted to patches around ground-truth points, a signed-distance-function loss that pulls points toward the vessel center, and a regularization loss penalizing long edges. The segmentation head then converts the deformed centerlines into a mask through a distance map fused with the coarse prediction.","core_discovery":"The central claim is that vessel extraction should be posed as deforming a connected centerline graph rather than classifying voxels. The paper introduces DeformCL, where a graph G = {V, ε} has 3D centerline points as nodes and edges that encode adjacency. An adaptive template, built by skeletonizing a coarse segmentation, reconstructing a graph via minimum spanning tree, and subsampling to a few control points, is refined through L cascaded stages. Each stage samples hierarchical feature maps at the point locations, aggregates features along the graph with a Transformer, predicts a bounded offset, and unpooling doubles the point density. The final centerlines are converted to a segmentation by computing a voxel-to-centerline distance map and fusing it with the coarse mask. The paper reports that this continuous representation with natural connectivity reduces fractures, filters local false positives, and improves topological metrics across HaN-Seg, HNCTA, ASOCA, and ImageCAS datasets.","pith_inferences":["A direct consequence not tested in the paper: since deformation never adds branches, centerline recall cannot exceed the recall of the coarse segmentation; a version that grows new edges during deformation would be needed for fully occluded vessels.","The same deformable-centerline principle could be applied to other curvilinear structures—neuronal arbors, airways, or road networks—where connectivity, centrality, and noise tolerance matter more than voxel-wise accuracy.","The paper's preference for rough initial templates (four control points) suggests the deformation is an active-contour-like process; measuring sensitivity to random template perturbations would reveal how much of the performance comes from the template prior versus the learned deformation."],"forward_implications":["Adding DeformCL as an extra head improves Dice, clDice, and Betti number errors of softDice, clDice, and DSCNet baselines on all four datasets.","The predicted centerlines are structurally complete and smooth enough to be used directly for curved planar reformation, without the usual mask-to-skeleton conversion.","Because the representation is continuous, noise robustness follows from the top-down deformation, reducing scattered false positives compared with bottom-up per-pixel classification.","The added computational cost is modest: about 0.12M extra parameters and 0.06 seconds per image over the 3D UNet backbone."],"supporting_citations":[{"why":"3D U-Net backbone that produces the coarse segmentation and hierarchical feature maps used by the template generator and the deformation stages.","marker":"[3]"},{"why":"3D thinning algorithm that converts the coarse segmentation mask into discrete skeleton points for template generation.","marker":"[15]"},{"why":"Strategy for reconstructing a graph from skeleton points via minimum spanning tree, used to build the initial centerline template.","marker":"[39]"},{"why":"Pixel2Mesh establishes the deformable-template paradigm of progressively deforming an initial shape toward a target.","marker":"[32]"},{"why":"Voxel2Mesh extends deformable templates to 3D medical volumes and motivates the cascaded deformation of an initial centerline template.","marker":"[34]"},{"why":"clDice defines a topology-sensitive similarity and loss, and serves as both a baseline and a metric for connectivity evaluation.","marker":"[28]"},{"why":"DSCNet baseline with dynamic snake convolution and topological constraints, one of the models DeformCL is added to.","marker":"[23]"},{"why":"DDT baseline for tubular-structure segmentation via distance transforms, against which DeformCL is compared.","marker":"[33]"},{"why":"Transformer layers provide the global feature aggregation along the centerline graph during each deformation stage.","marker":"[30]"},{"why":"V-Net's softDice is a baseline and the source of the Dice loss used for coarse and refined segmentation.","marker":"[19]"}],"fun_headline_variants":["DeformCL turns vessel extraction into deformable graph learning","Centerline graphs deform to beat mask-based vessel segmentation","Vessel extraction via deformable centerline graphs outperforms masks","DeformCL: learning deformable vessel centerlines, not voxel masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every vessel branch that matters must already be present in the coarse segmentation, because the deformation and unpooling layers move points along existing edges and never create new branches.","fun_headline_variants_meta":{"raw":{"variants":["DeformCL turns vessel extraction into deformable graph learning","Centerline graphs deform to beat mask-based vessel segmentation","Vessel extraction via deformable centerline graphs outperforms masks","DeformCL: learning deformable vessel centerlines, not voxel masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1544,"prompt_tokens":914,"completion_tokens":630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":530,"tokens_out":630,"duration_ms":5788,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:12:31.417503+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute per-branch centerline recall of the coarse segmentation and of DeformCL on a dataset with partially occluded vessels; if a branch entirely absent from the coarse mask appears in the final prediction, the deformation can create branches, whereas if none ever appear, the method's recall is capped by the coarse mask's completeness.","supporting_citations":[{"cited_title":"3d u-net: learn- ing dense volumetric segmentation from sparse annota- tion","cited_arxiv_id":null,"evidence_quote":"3D U-Net backbone that produces the coarse segmentation and hierarchical feature maps used by the template generator and the deformation stages."},{"cited_title":"Building skeleton models via 3-d medial surface axis thin- ning algorithms.CVGIP: Graphical Models and Image Pro- cessing, 56(6):462–478, 1994","cited_arxiv_id":null,"evidence_quote":"3D thinning algorithm that converts the coarse segmentation mask into discrete skeleton points for template generation."},{"cited_title":"Topology-preserving automatic labeling of coronary arteries via anatomy-aware connection classifier","cited_arxiv_id":null,"evidence_quote":"Strategy for reconstructing a graph from skeleton points via minimum spanning tree, used to build the initial centerline template."},{"cited_title":"V oxel2mesh: 3d mesh model genera- tion from volumetric data","cited_arxiv_id":null,"evidence_quote":"Voxel2Mesh extends deformable templates to 3D medical volumes and motivates the cascaded deformation of an initial centerline template."},{"cited_title":"cldice-a novel topology-preserving loss function for tubular structure seg- mentation","cited_arxiv_id":null,"evidence_quote":"clDice defines a topology-sensitive similarity and loss, and serves as both a baseline and a metric for connectivity evaluation."},{"cited_title":"Dynamic snake convolution based on topo- logical geometric constraints for tubular structure segmenta- tion","cited_arxiv_id":null,"evidence_quote":"DSCNet baseline with dynamic snake convolution and topological constraints, one of the models DeformCL is added to."},{"cited_title":"Deep distance transform for tubular structure segmentation in ct scans","cited_arxiv_id":null,"evidence_quote":"DDT baseline for tubular-structure segmentation via distance transforms, against which DeformCL is compared."},{"cited_title":"Attention is all you need.Advances in neural information processing systems, 30, 2017","cited_arxiv_id":null,"evidence_quote":"Transformer layers provide the global feature aggregation along the centerline graph during each deformation stage."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation","cited_arxiv_id":null,"evidence_quote":"V-Net's softDice is a baseline and the source of the Dice loss used for coarse and refined segmentation."}],"review_version":1}