{"id":"fef65876-4d85-45bc-83b8-69edb7efbb46","arxiv_id":"2507.15223","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A hierarchical part-based generative model that separates global vessel tree topology from local segment geometry achieves state-of-the-art graph fidelity on three vascular datasets.","lead":"Researchers built a three-stage AI model that generates realistic 3D blood vessel networks by first sketching the global branch tree, then creating each vessel segment, then assembling them. The approach beats existing methods on structural fidelity metrics and could improve synthetic data for medical imaging and surgical planning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The maximum-spanning-tree preprocessing in §4 drops any non-tree cycle; for the Circle of Willis (CoW) dataset this makes the method structurally unable to reproduce the true vascular topology, so the 'topological consistency' claim is only measured within a tree representation.","rationale":"The central claim is bold: 'first successful application of a part-based generative approach for 3D vessel modeling, setting a new benchmark.' The three-stage decomposition (global tree, local segments, assembly) is internally coherent, and the code is released, which is positive. However, the correctness of the claim depends on the premise that real vessel networks are adequately represented as binary trees. Section 4's 'maximum spanning tree' construction is where this premise enters: any cycle in a vascular graph is explicitly destroyed. This is not a minor implementation detail; it determines the representable class of the whole model. The CoW dataset is a particularly clean test because the Circle of Willis contains a ring; if the preprocessing deletes one connection, the model cannot synthesize the observed topology, and the graph metrics used in Table 1 are computed on a representation that has already erased the cycle. Without checking Betti numbers before and after preprocessing, the reported 'topological consistency' is circular relative to the tree assumption. I agree with the reader's weakest_assumption; my recommended verdict is unchanged because the reader already flagged the condition that would make the paper acceptable (explicitly scoping the claim to tree-like networks or handling cycles).","tokens_in":7557,"tokens_out":5334,"duration_ms":60191,"concrete_test":"For every CoW test sample, compute the first Betti number (cyclomatic number, β = |E| − |V| + 1 for a connected graph) of the raw skeleton graph extracted in §4 before and after the maximum spanning tree construction. Also count nodes with degree ≥ 3. If any sample has β_raw > 0 and β_key = 0 (or any multifurcation is split), then rerun the Table 1 graph metrics by adding the deleted cycle edge back into the generated graph (using the nearest reconstructed segment) and compute GWD/Deg./Spec. against the original full skeleton. If the reported GWD degrades sharply or the generated graph cannot represent the loop, the 'topological consistency' claim is unsupported for CoW.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 states: 'we adopt an adaptive mapping [30] method to construct a key graph and build a maximum spanning tree.' A maximum spanning tree of a graph with a cycle must delete at least one edge. Vessel networks such as the Circle of Willis (CoW) contain an anastomotic ring (the posterior communicating arteries), so the ground-truth skeleton cannot be a tree. By construction, the key graph becomes a tree, Stage 1 generates only binary trees, and Stage 3 assembles a tree. The graph metrics (Deg., Spec., GWD) in Table 1 are then computed on the model's generated tree skeletons and on the baseline skeletons, but if the same spanning-tree preprocessing is applied to all compared graphs, the missing cycle is hidden from every metric. Thus the headline claim of 'topological consistency' compares trees against trees, not against the actual cyclic vascular topology. A related underspecification exists for multifurcations: the RAE/transformer architecture in Stage 1 has only 'left/right child' branches, so a node with degree >3 would need to be split into multiple binary nodes, and the paper does not describe how this is done. The central claim of 'first successful application of a part-based generative approach for 3D vessel modeling' is contingent on the tree assumption; if the CoW ring is discarded, the method cannot be called realistic on that dataset.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage hierarchical part-based generative model for 3D blood vessel networks. Stage 1 uses a recursive variational autoencoder to generate a binary-tree ``key graph'' from skeleton bifurcation and terminal points. Stage 2 uses a transformer-based variational autoencoder to synthesize vessel segments conditioned on a geometric descriptor (length, straight-line distance, curvature, and depth). Stage 3 assembles the generated segments into a complete vessel skeleton according to the key graph and reconstructs the final mesh. The method is evaluated on ImageCAS, VascuSynth, and the Circle of Willis (CoW) datasets against VesselVAE, PointDiffusion, and TreeDiffusion, using point-based metrics (JSD, Chamfer distance) and graph-based metrics (degree MMD, Laplacian spectrum MMD, Graph Wasserstein distance). The authors report leading results on most graph metrics but substantially worse Chamfer distances than PointDiffusion, and they claim this is the first successful part-based generative approach for 3D vessel modeling.","tokens_in":7809,"tokens_out":4551,"duration_ms":49638,"significance":"If the underlying tree representation is accepted, the hierarchical part-based decomposition is a plausible and potentially useful contribution: it explicitly separates global branching topology from local segment geometry, and the three-stage pipeline is clearly described. The release of code and the use of public datasets support reproducibility. However, the central claim of ``realistic 3D blood vessel'' generation and ``topological consistency'' is weakened by the maximum-spanning-tree preprocessing, which discards cyclic topology present in the CoW dataset and forces all generated structures to be binary trees. The quantitative support is also mixed: the method trails PointDiffusion by a large margin on Chamfer distance on every dataset, so the headline ``superior performance'' statement is stronger than the evidence. With a clearly scoped claim (tree-shaped vascular networks) and a clarified evaluation pipeline, the work would be a useful addition; as written, the broad realism and topology claims need revision.","major_comments":[{"comment":"The paper's central claim of topological consistency is compromised by the maximum spanning tree preprocessing. In §4, the authors state that they ``adopt an adaptive mapping [30] method to construct a key graph and build a maximum spanning tree.'' For the Circle of Willis (CoW) dataset, the ground-truth vascular network contains an anastomotic ring (the posterior communicating arteries); a maximum spanning tree of a cyclic graph necessarily deletes at least one edge. Stage 1 then generates only binary trees, and Stage 3 assembles a tree. Consequently, the generated vessels cannot reproduce the cyclic topology of CoW, and the graph metrics in Table 1 (Deg., Spec., GWD) compare tree representations rather than the true vascular topology. The claim that the model achieves ``topological consistency'' is therefore only valid within a tree approximation. Please either (a) restrict the claims and method to tree-shaped vascular networks and document the fraction of test cases that are non-tree, or (b) add a cycle-aware evaluation (e.g., Betti number or cycle count computed on original skeletons) and a baseline comparison that does not rely on MST reduction.","section":"§4 Data Preparation"},{"comment":"The binary-tree representation is underspecified for non-binary branch points. The decoder in Eqs. (2)--(3) uses NodeCLS to predict whether left and/or right children exist, so any node with more than two children (trifurcations or higher multifurcations) must be split into multiple binary nodes. The paper does not describe how such splits are performed, how the geometric descriptor C (length, distance, curvature, depth) is assigned to the resulting sub-segments, or how these artificial splits affect the assembled skeleton. Since real vascular networks can contain such branch points, this missing specification affects the generality of the method and the interpretation of the reported graph metrics.","section":"§3 Stage 1; §4 Data Preparation"},{"comment":"The statement that the model ``achieves competitive performance across most tasks'' should be qualified by the large Chamfer distance gap. On CD, the proposed method is worse than PointDiffusion by a factor of roughly 20 on ImageCAS (24.4 vs 1.1), about 5 on VascuSynth (34.8 vs 7.0), and about 15 on CoW (15.1 vs 1.0). Since CD is a standard geometric reconstruction metric, the paper needs to explain why the graph metrics should be prioritized, and the ``superior performance'' claim in the abstract and conclusion is too strong as written.","section":"§4 Table 1"},{"comment":"The computation of the graph-based metrics is not fully specified. It is unclear whether the skeletons of all methods (including PointDiffusion's point clouds and TreeDiffusion's implicit fields) are derived with the same skeletonization and maximum spanning tree preprocessing as the proposed method, or whether the ground-truth graphs are also reduced to trees before computing Deg., Spec., and GWD. This is essential for a fair comparison and for assessing whether the reported improvements in topological consistency come from the generative model itself or from a preprocessing step applied uniformly to all methods. Please provide the precise preprocessing pipeline for each baseline and the details of the metric computation.","section":"§4 Metrics"}],"minor_comments":[{"comment":"There are several typographical issues: ``frame work'' in the abstract, ``T able 1'' in the table caption, and ``Instage 1'' in the Implementation Details section.","section":"Abstract; Table 1; Implementation Details"},{"comment":"The caption states that the comparison uses ``TreeDiffusion's best-performing samples,'' but the criterion for selecting these samples is not defined; please state the selection procedure to avoid the appearance of cherry-picking.","section":"Figure 5"},{"comment":"The claim ``first successful application of a part-based generative approach'' is stronger than the evidence; consider softening it to ``a first part-based approach'' given the mixed quantitative results and the tree-only scope.","section":"Abstract; Contributions"},{"comment":"The related work section introduces reference [15] only in a list of baselines; a brief discussion of how the proposed key-graph representation differs from the graph-generation approach in [15] would help position the contribution.","section":"Related Work"},{"comment":"The sentence ``We are the first study to address this challenging dataset'' for ImageCAS is vague; please clarify whether this refers to generative modeling of vessels, segmentation, or something else, and verify the claim against the literature.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the maximum-spanning-tree preprocessing is, in my reading, valid and load-bearing for the paper's main claim. The method is internally consistent, and the code release is a plus, but the ``realistic'' and ``topological consistency'' claims need to be scoped to tree-shaped vasculature or supported by cycle-aware evaluation. The Chamfer distance results also undercut the ``superior performance'' phrasing. These issues are fixable within the manuscript's scope, so major revision seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2507.15223. First, the core idea is genuinely new: nobody has modeled 3D vessels as a combination of a global binary-tree key graph and locally generated tubular segments. That decomposition is clean, the three-stage pipeline (RVAE for topology, transformer VAE for segments, assembly) is well-specified, and the code is public. Second, the paper's claim to 'realistic' generation on the Circle of Willis dataset is undermined by its own preprocessing. Section 4 says it builds a maximum spanning tree from the skeleton. CoW contains an anastomotic ring (posterior communicating arteries); a spanning tree must delete an edge, so the model can never produce that cycle. The graph metrics in Table 1 are then computed on the generated trees and on baseline skeletons that likely went through the same spanning-tree preprocessing, so the missing ring is invisible to every metric. That is not just a performance gap; it changes the claim from 'topological consistency' to 'tree-consistency.' The paper never acknowledges this limitation.\n\nWhere the paper is solid: for tree-like vessels—coronary arteries, VascuSynth—the part-based approach makes sense and the graph metrics (degree MMD, spectrum, GWD) show clear gains over VesselVAE and TreeDiffusion. The separation of global topology from local geometry is well-motivated, and the qualitative results look structurally coherent.\n\nSoft spots, in order of worry. (1) The CoW cycle problem above is load-bearing for the realism claim on that dataset. (2) The paper cites Prabhakar et al. [15] but does not compare against it, even though it is the most relevant vessel-graph diffusion baseline. (3) No error bars anywhere; given the small test sets, some of the metric differences may not be significant. (4) The abstract says 'superior performance over existing methods,' but Table 1 shows PointDiffusion wins Chamfer distance by an order of magnitude on all datasets, and TreeDiffusion beats the method on spectrum MMD on VascuSynth. That is overstatement.\n\nBottom line: worth a serious referee. The idea is a useful addition to the vessel-generation toolbox, provided the CoW limitation is addressed or the claims are narrowed to acyclic vasculature. I would send it to review with a request for clarification.","headline":"Novel part-based vessel generator, but the spanning-tree preprocessing hides the Circle of Willis ring and the realism claim is overstated.","tokens_in":8368,"tokens_out":3306,"would_cite":true,"duration_ms":35613,"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":"A hierarchical part-based framework that separates global tree topology from local segment geometry yields more topologically consistent 3D blood vessel networks than existing generative methods.","keywords":["vasculature","vessel generation","3D shape modeling","hierarchical structure","part-based method","recursive variational autoencoder","transformer VAE","key graph"],"falsifier":"Run the method on a vascular dataset that contains anatomically known loops or trifurcations (for example, the circle of Willis with its communicating arteries); if the extracted maximum spanning tree removes those cycles and the generated samples never contain a loop or triplet junction, then the claim that the model preserves topological consistency is refuted.","tokens_in":7352,"feed_emoji":"🫀","tokens_out":5895,"duration_ms":56634,"temperature":0.7,"pith_summary":"Blood vessel networks are hard to generate because they are globally tree-like and locally curved and irregular. This paper argues that existing point-cloud and implicit-field generative models blur those two levels, so they produce disconnected or anatomically odd vessels, especially on complex real-world coronary data. To fix that, the authors split the problem: one stage generates the branching skeleton as a binary tree (the 'key graph'), another generates each vessel segment as a 3D curve from a geometric descriptor, and a final stage assembles segments onto the graph. They report that this three-stage part-based approach beats the baselines on graph-based measures of structural fidelity and topological consistency, while remaining competitive on point-cloud metrics. If correct, it establishes a new application area for part-based generative modeling: vascular anatomy where topology matters.","feed_headline":"Part-based model keeps branching intact in 3D vessel generation","feed_subtitle":"Separating global tree shape from local segment curves beats baselines on topological fidelity across three datasets.","key_machinery":"The key graph is the central object: a binary tree whose nodes carry spatial coordinates, a direction vector, and the geometric descriptor $C=[\\ell,\\delta,\\kappa,\\rho]$, with each edge corresponding to one vessel segment. The encoding and decoding machinery is a recursive variational autoencoder (RVAE) that aggregates child hidden states into parents with an MLP, classifies node presence during decoding, and is regularized by a KL term; segments are generated by a Transformer VAE whose tokens are skeleton points $[x,y,z,r]$ and whose condition is $C$. Assembly walks the decoded graph depth-first, attaching each generated segment by scaling, translation, and two rotation steps so that segment direction matches the node's orientation, then reconstructs the surface from the skeleton using the per-point radii.","core_discovery":"The central claim is that a vascular network is better generated as a composition of parts than as a whole shape: first a binary-tree 'key graph' that fixes the global branching topology, then independent vessel segments generated as ordered 3D point sequences, then an assembly stage that places each segment onto the graph by scaling, translation, and rotation. The authors represent each key-graph node by its 3D coordinate, the local direction $[n_x,n_y,n_z]$, and a geometric descriptor $C=[\\ell,\\delta,\\kappa,\\rho]$ (length, endpoint distance, curvature, tree depth), and train a recursive variational autoencoder on this tree structure while a Transformer-based variational autoencoder generates each segment conditioned on $C$. On ImageCAS coronary arteries, VascuSynth synthetic trees, and processed Circle-of-Willis meshes, the method claims the best graph-based metrics (degree and Laplacian-spectrum discrepancy, graph Wasserstein distance), which are the metrics that score topological consistency; on point-cloud metrics it is competitive but not always best.","pith_inferences":["If the decomposition generalizes, the same key-graph-plus-segments recipe could be applied to other tree-like anatomical structures such as airways, bile ducts, or plant roots, where global topology matters as much as local shape.","The reliance on a maximum spanning tree means the method will silently drop vascular cycles or trifurcations; a natural test is to quantify how much topology is lost in preprocessing and whether the generative model can ever reproduce it.","One could swap the RVAE for a different hierarchical generator or the Transformer VAE for a latent diffusion per segment; the paper's contribution is the decomposition, not the specific architectures, so those swaps are plausible extensions."],"forward_implications":["Generated vessels preserve the binary branching hierarchy better than point-cloud and implicit neural-field baselines, as measured by degree and Laplacian-spectrum discrepancies.","The part-based decomposition allows complex real-world datasets like ImageCAS to be modeled without the topological breakdowns seen in baseline reconstructions.","Because local segments are generated independently and assembled, the framework can scale to vessels with many bifurcations without the fidelity degradation VesselVAE suffers.","Skeleton-based assembly with predicted radii yields tubular meshes that avoid the holes point-cloud reconstructions produce.","This is the first part-based generative approach for 3D vascular modeling, opening that direction for future work."],"supporting_citations":[{"why":"Supplies the adaptive key-graph mapping and the skeleton-to-mesh reconstruction used in preprocessing and final assembly.","marker":"[30]"},{"why":"TreeDiffusion, the implicit neural-field diffusion baseline for anatomical trees that the method must beat on generation quality.","marker":"[17]"},{"why":"VesselVAE, the recursive variational autoencoder baseline for branching vascular structures that the method extends and compares against.","marker":"[6]"},{"why":"PointDiffusion, the point-cloud diffusion baseline used as the state-of-the-art point-cloud comparison.","marker":"[11]"},{"why":"Defines the Graph Wasserstein distance used to score reconstruction quality of skeleton graphs.","marker":"[2]"},{"why":"ImageCAS, the real-world coronary artery dataset used as the most challenging test.","marker":"[28]"},{"why":"Supplies the Jensen-Shannon divergence and Chamfer distance metrics for point-cloud evaluation.","marker":"[1]"}],"fun_headline_variants":["Part-based vessel gen: tree topology first, tubes second","Splitting global tree from local segments improves vessel models","Global tree, local tubes: new 3D vessel generation","Graph first, segments later: part-based vessel synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a real vascular network can be faithfully represented as a binary tree; the preprocessing builds a maximum spanning tree, so any genuine loop or multi-way junction that the spanning tree discards is absent from every generated vessel.","fun_headline_variants_meta":{"raw":{"variants":["Part-based vessel gen: tree topology first, tubes second","Splitting global tree from local segments improves vessel models","Global tree, local tubes: new 3D vessel generation","Graph first, segments later: part-based vessel synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1742,"prompt_tokens":932,"completion_tokens":810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":743}},"tokens_in":548,"tokens_out":810,"duration_ms":8965,"temperature":1.0,"reasoning_tokens":743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:38:00.819117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a vascular dataset that contains anatomically known loops or trifurcations (for example, the circle of Willis with its communicating arteries); if the extracted maximum spanning tree removes those cycles and the generated samples never contain a loop or triplet junction, then the claim that the model preserves topological consistency is refuted.","supporting_citations":[{"cited_title":"In: International Symposium on Bioinformatics Research and Applications","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive key-graph mapping and the skeleton-to-mesh reconstruction used in preprocessing and final assembly."},{"cited_title":"In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024","cited_arxiv_id":null,"evidence_quote":"TreeDiffusion, the implicit neural-field diffusion baseline for anatomical trees that the method must beat on generation quality."},{"cited_title":"In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"VesselVAE, the recursive variational autoencoder baseline for branching vascular structures that the method extends and compares against."},{"cited_title":"Computerized Medical Imaging and Graphics 109, 102287 (2023)","cited_arxiv_id":null,"evidence_quote":"ImageCAS, the real-world coronary artery dataset used as the most challenging test."},{"cited_title":"In: Proceedings of the 35th International Conference on Machine Learning (ICML)","cited_arxiv_id":null,"evidence_quote":"Supplies the Jensen-Shannon divergence and Chamfer distance metrics for point-cloud evaluation."}],"review_version":1}