{"id":"06efdc39-9501-4681-a3bf-fa5ecc62877b","arxiv_id":"2508.17658","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors build the first point cloud-based coronary artery completion dataset and show that their TSRNet baseline outperforms existing point cloud and voxel completion methods on three benchmarks.","lead":"This paper introduces a new benchmark dataset, PC-CAC, for reconnecting broken blood-vessel-like structures represented as 3D point clouds, and a neural network, TSRNet, that fills the gaps. A generalist might care because point cloud completion could make automated analysis of coronary arteries more robust in severe cases like blockages or narrowing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PC-CAC's reference point clouds are automatic segmentation skeletons, not expert-corrected anatomy, so the benchmark lacks a validated clinical anchor.","rationale":"The reader's weakest assumption is exactly the one that carries the paper's central claim: the completeness ground truth for PC-CAC is an automatic segmentation converted to point clouds, and the fractures are simulated random deletions. My concern is the same and is load-bearing because the benchmark's two advertised contributions--a new clinically derived dataset and a baseline that 'consistently outperforms' SOTA--are only meaningful if the reference point cloud is an accurate proxy for the anatomy to be reconnected. The paper provides no validation of this proxy: no inter-rater study, no comparison against manual annotations, and no test on real pathological discontinuities (Secs. 4.1, 5). I also note that coronary segments are generated as centerlines (Alg. 1, Table 1), so even the geometry being completed is one-dimensional; this reinforces, rather than replaces, the reference-standard concern. The recommended verdict remains CONDITIONAL: the paper's engineering contributions are plausible and the empirical comparisons are extensive, but the central claim should be accepted only after an external validation of the ground truth. This stress-test pass therefore does not move the reader's verdict, so the appropriate setting is UNCHANGED.","tokens_in":18395,"tokens_out":8560,"duration_ms":95728,"concrete_test":"Take 50 PC-CAC test CTAs and have two independent radiologists manually annotate the complete coronary artery centerline and lumen radius (or manually correct the automatic segmentation) to create an expert reference point cloud for each case. Recompute Table 2's CD, F1, and Fidelity for TSRNet and the three strongest baselines (SeedFormer, PoinTr, SnowflakeNet) against this expert reference, and report inter-rater agreement. If the relative ranking changes, or TSRNet's mean CD advantage over SeedFormer narrows beyond one standard deviation, the segmentation-derived ground truth is not a neutral benchmark and the claim must be re-scoped to 'segmentation-proxy centerline completion' rather than clinically validated tubular reconstruction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PC-CAC is constructed from the output of an automatic coronary-artery segmentation (Qi et al., 2023), cleaned morphologically and reduced to a centerline for coronary arteries (Alg. 1 and Sec. 4.1); no expert correction or independent anatomical verification is reported. The 'complete' point cloud is therefore not an independent reference standard: it is a skeleton produced by the same algorithmic family the authors develop. Any systematic bias in that segmentation is embedded in both the input and the ground truth, so the completion scores in Table 2 measure fidelity to the segmentation pipeline's reconstruction of the vessel tree, not to the patient's true coronary anatomy. Additionally, representing coronaries as centerline-only points omits vessel caliber, so the dataset supports curve reconnection rather than detail-preserved 3D tubular reconstruction, which is what the clinical framing in the Introduction claims. Because the headline claim--a first clinically derived coronary completion benchmark and a consistently superior baseline--depends on this reference standard, the current evaluation cannot certify anatomical or clinical validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PC-CAC, a point cloud-based coronary artery completion dataset derived from real CT data, and proposes TSRNet, a network for reconnecting fractured tubular structures. The dataset and TSRNet are evaluated on PC-CAC and two converted public datasets (PC-ImageCAS, PC-PTR) against generic point cloud completion baselines and two voxel-based methods. The authors report that TSRNet achieves the best or second-best results on most metrics across the three datasets, and they provide ablations of the main components and hyperparameters. The dataset and code are publicly released.","tokens_in":18598,"tokens_out":5796,"duration_ms":57788,"significance":"If the benchmark were independently validated, this would be a useful contribution: a large-scale (427 patients, 3,416 cases) public benchmark for a clinically motivated reconnection task, with a strong baseline and broad comparisons against seven point cloud completion methods and two voxel-based methods. The paper also ships a public codebase and dataset link, which supports reproducibility. However, the significance is substantially tempered by the fact that the ground truth is the authors' own automatic segmentation output without expert correction or independent anatomical validation, and by internal inconsistencies in the reported loss threshold and ablation setup. The central clinical claim—that the benchmark and method capture 'detail-preserved' completion of true coronary anatomy—is not yet supported.","major_comments":[{"comment":"The PC-CAC ground truth is generated from automatic segmentation results (nnU-Net and the authors' own DSN, Qi et al. 2023) and then skeletonized to a centerline for the coronary arteries. No expert correction or independent anatomical verification is reported. The completion task is therefore defined as reconstructing the segmentation pipeline's output, not the patient's true coronary anatomy. This is load-bearing for the clinical framing in the Introduction and for the 'detail-preserved' claim, because the representation discards vessel caliber and the ground truth embeds any systematic bias of the segmentation. I recommend that the authors either obtain expert-verified ground truth for at least a test subset, or explicitly reframe the benchmark as 'segmentation-output reconnection' and temper claims about clinical validity.","section":"Sec. 4.1 and Algorithm 1"},{"comment":"The implementation details state that the loss threshold epsilon is set to 6 × 10^-3, but Table 4 reports an ablation over epsilon with values 3.1, 2.9, and 2.7, and row 7 (epsilon = 2.7) is declared best. If the Table 4 values are implicitly scaled by 10^-3, then the default (6 × 10^-3) differs from the reported best (2.7 × 10^-3). If they are not scaled, the units are inconsistent. The paper does not clarify which epsilon was actually used for the main results in Table 2. This ambiguity directly affects the reproducibility of the loss function and the interpretation of the ablation study, so it must be resolved.","section":"Sec. 4.3 vs. Table 4"},{"comment":"The text states that the hyperparameter ablation is conducted 'on the PC-CAC dataset', but the best row (Row 7) reports CD=2.395, F1=96.83, Fidelity=2.062, which exactly match the TSRNet results for PC-ImageCAS in Table 2, not the PC-CAC results in the same table. This indicates a mismatch between the table's data and its stated evaluation dataset. As written, the ablation conclusions are not reproducible and the experimental setup is unclear. The authors should correct either the text or the table and re-run the ablation on the appropriate dataset.","section":"Sec. 5.1.1 and Table 4"},{"comment":"The paper's abstract claims TSRNet 'consistently outperforms state-of-the-art approaches across multiple evaluation metrics', and Sec. 5.1.1 states it achieves 'state-of-the-art performance in CD, F1, and Fidelity Error' on all three datasets. This is contradicted by Table 2 on PC-ImageCAS, where SeedFormer achieves F1=97.47 and TSRNet achieves 96.83; the text later acknowledges 'second-best F1'. The authors should either weaken the headline claim to reflect metric-wise and dataset-wise exceptions, or provide a clear statistical argument for why the method is still considered consistently superior (e.g., paired significance tests).","section":"Sec. 5.1.1 and Table 2"}],"minor_comments":[{"comment":"The text refers to 'Table 5.2.2' while the actual table is labeled 'Table 5'; the cross-reference should be corrected.","section":"Sec. 5.2"},{"comment":"Typo: 'where fice representative cases' should be 'where five representative cases'.","section":"Sec. 5.2.2"},{"comment":"The evaluation metric 'Fidelity Error' is also used as a loss term (Sec. 3.3). This coupling should be explicitly acknowledged, since optimizing the loss directly minimizes the evaluation metric and may inflate its reported advantage.","section":"Sec. 4.3"},{"comment":"The qualitative figures show red-marked error regions, but the criteria for marking are not defined. A brief description of how error regions were determined would improve interpretability.","section":"Fig. 4 and Fig. 5"},{"comment":"The PC-ImageCAS and PC-PTR conversion details are summarized only in Table 1. It would be helpful to state explicitly whether the same fracture-simulation algorithm (Algorithm 1) is applied to these public datasets, and with which min/max break ratios beyond the point-removal percentages mentioned.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The benchmark's ground truth is derived from the authors' own segmentation method (Qi et al., 2023), with the first author appearing on both this manuscript and the segmentation paper. This is not itself misconduct, but it means the 'clinical validity' claim rests on an internal reference standard. I would advise the editor to request the authors to either add an independent validation (e.g., expert-annotated test subset) or clearly scope the benchmark as measuring reconnection quality relative to automatic segmentation outputs. The novelty claim of being 'the first point cloud-based tubular structure reconnection dataset' should also be checked against Weng et al.'s PTR dataset and the existing cardiac point-cloud completion work of Beetz et al., which the paper cites but does not position against in the contribution list."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper gives the community a real asset: a new coronary artery point cloud dataset (PC-CAC) with simulated fractures, plus point cloud versions of ImageCAS and PTR, and a baseline network that consistently beats the existing point cloud completion methods on those benchmarks. The dataset is large (427 patients, 3,416 cases), and the authors release code and data. That is a legitimate contribution, and the empirical work is honest enough that the main claim—TSRNet outperforms the baselines on CD, F1, and fidelity on these three datasets—looks defensible.\n\nWhat is genuinely new is the task formulation: point cloud-level reconnection of fractured tubular structures. The network itself is an assembly of known modules (core point selection, set abstraction + transformer, coarse-to-fine refinement, CD + fidelity loss), but it is adapted sensibly to the problem and the ablations show each piece helps.\n\nThe soft spots are real, but not fatal.\n\nFirst, the ground truth is generated from the authors' own segmentation pipeline (Qi et al. 2023, nnU-Net), with no expert correction or independent validation reported. If that segmentation is biased, the bias is baked into both input and target, so the scores measure faithfulness to the segmentation, not to the true coronary anatomy. This is the load-bearing assumption, and the paper does not adequately defend it.\n\nSecond, the coronaries are represented as centerlines (Algorithm 1), not surfaces. The task for those vessels is graph reconnection, not detail-preserved 3D reconstruction. The phrase \"detail-preserved\" in the title and abstract overstates what the benchmark actually measures. Only the aorta gets surface points.\n\nThird, there is an internal inconsistency: Section 4.3 says the loss threshold epsilon is set to 6e-3, while Table 4 reports the best result with epsilon=2.7. This needs to be resolved before publication.\n\nFourth, the comparison with voxel-based methods (DRTT, VSR-Net) is under-specified. Those are volumetric methods, but the metrics are point-cloud distances. How were the inputs and outputs converted? Without that detail, the \"we beat voxel methods\" claim is hard to scrutinize.\n\nNone of this sinks the paper. But the authors should either validate the ground truth against an expert-corrected subset or soften the clinical framing. They should also be explicit that the coronary completion is centerline reconnection, not vessel surface reconstruction.\n\nMy take: send it to peer review. The benchmark is useful, the baseline is reproducible, and the issues are fixable with clearer claims and a validation appendix.","headline":"Useful new point cloud completion benchmark for coronary vessels, but the ground truth is segmentation output and the coronaries are centerlines, so the 'detail-preserved' claim overreaches.","tokens_in":19123,"tokens_out":3371,"would_cite":true,"duration_ms":35721,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that reconnecting fractured coronary arteries is best posed as point cloud completion, and it introduces a clinical benchmark and a network, TSRNet, that beat voxel-based repair and prior point-cloud methods on three…","keywords":["point cloud completion","tubular structure reconnection","coronary artery","PC-CAC dataset","benchmark","topology preservation","medical image analysis","deep learning"],"falsifier":"Take a set of coronary CTAs with expert-annotated complete vessel trees, remove vessel segments at locations of genuine stenosis or occlusion rather than random components, and ask TSRNet to reconnect them; if its completion error on those real clinical fractures is no better than generic point-cloud completion, the PC-CAC synthetic fractures are not representative of the clinical problem.","tokens_in":18220,"feed_emoji":"🫀","tokens_out":8713,"duration_ms":78970,"temperature":0.7,"pith_summary":"This paper argues that fractured tubular structures, especially coronary arteries, can be reconnected more accurately and efficiently when the task is posed as point cloud completion rather than voxel repair. To support that argument, it builds PC-CAC, a point cloud benchmark derived from real clinical coronary CTAs, and introduces TSRNet, a completion network whose components target the thin, imbalanced, topologically complex geometry of vessels. The authors report that TSRNet outperforms prior point cloud completion models and two voxel-based reconnection methods on PC-CAC and on point-cloud conversions of two public datasets. If the benchmark faithfully represents clinical discontinuities, the work gives downstream diagnosis and hemodynamic simulation a reusable dataset and baseline for restoring vessel continuity.","feed_headline":"Point-cloud model reconnects fractured coronary arteries better","feed_subtitle":"A new benchmark and network restore vessel continuity that segmentation loses at stenoses and occlusions.","key_machinery":"The load-bearing object is TSRNet plus the dataset-construction pipeline. The pipeline converts clinical coronary segmentations into a point cloud by extracting the aorta surface with marching cubes and farthest point sampling, extracting coronary centre lines, normalizing, and then deleting entire connected point-components at random to create fractured inputs while keeping the intact cloud as ground truth. The network's components are a Core Points Selection module that samples separately in dense and sparse regions with priority for endpoints and isolated points; a TransSA module that alternates set abstraction and transformer self-attention; a two-stage dense refinement that reconstructs a coarse cloud then a fine cloud; and a global-to-local loss that applies Chamfer Distance and Fidelity Error globally, then re-applies both only to sparse regions once global loss falls below a threshold. That loss schedule is what shifts training from overall shape alignment to fine vascular detail.","core_discovery":"The central claim is that vessel reconnection can be redefined as a point cloud completion problem, and that a dedicated dataset and network make that formulation effective. PC-CAC contains 3,416 point clouds from 427 patients, each with 4,096 points (3,072 on the aorta surface, 1,024 on coronary branch centre lines), with 10 to 30 percent of coronary points removed in up to eight synthetic fracture patterns per patient to simulate stenosis, occlusion, and tortuous-region breakage. TSRNet combines a core-point selection module that balances dense and sparse regions, a transformer set-abstraction feature extractor, two progressive refinement stages, and a global-to-local loss mixing Chamfer Distance and Fidelity Error. On PC-CAC, PC-ImageCAS, and PC-PTR the paper reports the lowest Chamfer Distance and Fidelity Error in every case and the highest F1 on two of the three datasets, with the largest margin on PC-PTR fidelity error (0.102 versus 3.666 for the second-best model). The authors take this as evidence that point-cloud-level completion preserves thin tubular detail better than voxel-level repair or generic point-cloud completion.","pith_inferences":["The synthetic fractures are random deletions of connected point components in segmentation-derived clouds; real occlusions also corrupt image intensity, so a test on CTA regions with true stenosis may show different failure modes than the benchmark captures.","Because the ground truth is the authors' own segmentation output, part of the reported gap over other methods may reflect how closely TSRNet reproduces that particular segmentation's topology; an independent expert-annotated completeness standard would separate reconnection skill from dataset bias.","The unusually large fidelity-error margin on PC-PTR suggests the method's density-aware sampling and sparse-region loss are especially valuable when thin structures dominate; this predicts strong transfer to retinal vessels, airways, or other slender tubular networks.","The authors state plans to pair point clouds with original image data; that pairing could be used to resolve ambiguous long-gap reconnections by consulting image appearance, a direction their current geometry-only model cannot test."],"forward_implications":["Point-cloud representation cuts the reconnection problem from over a million voxels to about four thousand points, making completion fast enough to serve as a preprocessing step in clinical pipelines.","The reconnected point cloud can be converted directly into a sub-voxel mesh, so downstream hemodynamic simulation and anatomical reconstruction do not need to return to voxel space.","Because TSRNet does not rely on explicit fracture endpoint detection, it avoids a failure mode of random-walk reconnection methods that require local endpoint matching.","The method remains stable as fracture severity rises from 10% to 30%, and at bifurcations it reconstructs all intersecting branches, suggesting the model is not merely filling short gaps.","Evaluating the same network on coronary and pulmonary datasets indicates the approach transfers across different tubular anatomies, not just one vessel tree."],"supporting_citations":[{"why":"Supplies the clinical coronary segmentation results from which the PC-CAC point clouds and ground truth are derived.","marker":"(Qi et al., 2023)"},{"why":"Provides the nnU-Net segmentation framework used alongside Qi et al. to produce the vessel segmentations that become PC-CAC point clouds.","marker":"(Isensee et al., 2021)"},{"why":"Provides the ImageCAS coronary segmentation dataset that is converted into the PC-ImageCAS benchmark.","marker":"(Zeng et al., 2023)"},{"why":"Provides the pulmonary tree repairing dataset converted into PC-PTR and a voxel-domain reconnection context.","marker":"(Weng et al., 2023)"},{"why":"SeedFormer is the patch-seed architecture that inspires TransSA and is the strongest point-cloud baseline on several reported metrics.","marker":"(Zhou et al., 2022)"},{"why":"Defines the l1 Chamfer Distance used as the primary completion metric and global loss component.","marker":"(Fan et al., 2017)"},{"why":"Defines the Fidelity Error used to evaluate input preservation and as the local loss component.","marker":"(Yuan et al., 2018)"},{"why":"Provides PointNet and farthest point sampling used in the Core Points Selection module and feature extraction.","marker":"(Qi et al., 2017a)"},{"why":"PoinTr is a transformer-based completion baseline that TSRNet outperforms and whose geometry-aware design is related.","marker":"(Yu et al., 2021)"}],"fun_headline_variants":["New benchmark and network mend fractured coronary arteries","TSRNet outperforms on new point-cloud coronary completion benchmark","Point-cloud dataset and model reunite broken coronary arteries","Restoring vessel continuity: a new point-cloud benchmark and network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole benchmark assumes that the authors' segmentation of clinical CTAs is an accurate complete ground truth, and that randomly deleting connected chunks of that segmentation mimics the discontinuities a clinician would see in real stenosis or occlusion.","fun_headline_variants_meta":{"raw":{"variants":["New benchmark and network mend fractured coronary arteries","TSRNet outperforms on new point-cloud coronary completion benchmark","Point-cloud dataset and model reunite broken coronary arteries","Restoring vessel continuity: a new point-cloud benchmark and network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2484,"prompt_tokens":1008,"completion_tokens":1476,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":1411}},"tokens_in":624,"tokens_out":1476,"duration_ms":10148,"temperature":1.0,"reasoning_tokens":1411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:02:00.742958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of coronary CTAs with expert-annotated complete vessel trees, remove vessel segments at locations of genuine stenosis or occlusion rather than random components, and ask TSRNet to reconnect them; if its completion error on those real clinical fractures is no better than generic point-cloud completion, the PC-CAC synthetic fractures are not representative of the clinical problem.","supporting_citations":[{"cited_title":", author Wu, C","cited_arxiv_id":null,"evidence_quote":"Provides the ImageCAS coronary segmentation dataset that is converted into the PC-ImageCAS benchmark."},{"cited_title":", author Yang, J","cited_arxiv_id":null,"evidence_quote":"Provides the pulmonary tree repairing dataset converted into PC-PTR and a voxel-domain reconnection context."},{"cited_title":", author Cao, Y","cited_arxiv_id":null,"evidence_quote":"SeedFormer is the patch-seed architecture that inspires TransSA and is the strongest point-cloud baseline on several reported metrics."},{"cited_title":", author Su, H","cited_arxiv_id":null,"evidence_quote":"Defines the l1 Chamfer Distance used as the primary completion metric and global loss component."},{"cited_title":", author Khot, T","cited_arxiv_id":null,"evidence_quote":"Defines the Fidelity Error used to evaluate input preservation and as the local loss component."}],"review_version":2}