{"id":"f0b6d87e-eee0-4c4a-91d9-d85dcc317261","arxiv_id":"2504.13599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dual-branch 3D network with a vision graph encoder achieves leading Dice, Hausdorff distance, and precision results on the ASOCA and ImageCAS coronary CTA datasets.","lead":"ViG3D-UNet is a 3D neural network that combines ordinary convolution with graph-based processing to segment coronary arteries in CT scans. The authors report improved accuracy and fewer broken branches than four established baselines on two public datasets, though the connectivity gain is inferred from indirect metrics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Connectivity claim rests on unvalidated proxy metrics; no direct measurement of vascular connectivity or significance testing is reported.","rationale":"The reader's weakest_assumption correctly identifies that precision and HD95 are not valid direct measures of vascular connectivity, and that no significance testing accompanies the reported gains. My review reaches the same conclusion after independently checking the results tables, the metric definitions, and the discussion sections. The improvements over the strongest baseline (STUNet) are numerically small and within the noise indicated by the large reported standard deviations; without direct topology metrics it is impossible to know whether the method actually reduces breaks or endpoint omissions. Because the paper is otherwise internally plausible—the architecture is a reasonable extension of vision GNNs, and the ablation studies are consistent—the appropriate verdict is CONDITIONAL: the connectivity claim should be re-examined once direct metrics and significance tests are provided. This does not change the reader's verdict, so I recommend UNCHANGED. I agree with the reader's assessment that the concern is not about disagreement with graph-based methods, but about the gap between the evidence and the claim.","tokens_in":15688,"tokens_out":4380,"duration_ms":42635,"concrete_test":"On the same ImageCAS and ASOCA test folds, compute direct connectivity metrics: (1) number of 3D connected components in the predicted coronary mask versus ground truth, (2) fraction of ground-truth branch endpoints that lie within a clinically relevant tolerance (e.g., 2 mm) of the predicted mask, (3) total recovered centerline length, and (4) Betti-0 error. Compare ViG3D-UNet against nnU-Net and STUNet using paired Wilcoxon signed-rank tests on these metrics. If ViG3D-UNet does not show statistically significant improvement in component count or endpoint coverage, the connectivity claim is falsified. As a secondary check, correlate precision and HD95 with these direct metrics to see whether the asserted proxies are valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that ViG3D-UNet 'surpassed competing methods in maintaining vascular segmentation connectivity' (abstract; also Section IV.D)—is supported only by precision, HD95, and visual inspection, not by any direct connectivity metric. Section IV.A.3 states 'Precision indicates the algorithm's ability to maintain connectivity,' but precision measures false-positive rate, not topology: a prediction can be highly precise yet contain broken branches or omitted distal segments. HD95 measures boundary distance, which can be inflated by a single distant false positive or a missing branch tip, but it does not quantify component count, branch continuity, or endpoint coverage. Moreover, the reported recall of ViG3D-UNet is lower than STUNet on both datasets (ASOCA 79.65 vs 82.31; ImageCAS 80.79 vs 83.51), which is consistent with missing vascular branches—a direct connectivity failure. The magnitude of the claimed improvements is also small relative to the reported standard deviations: for ImageCAS, DSC is 81.72 vs 81.54 (STUNet) with std ~5, and HD95 is 15.21 vs 17.60 with std ~18.5; no paired significance test is provided. The paper's own limitations mention limited training data and a 'blind search of feature distances' in graph construction, further weakening the inference that the architecture specifically improves connectivity. Thus the central claim is not empirically established by the presented evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ViG3D-UNet, a 3D U-shaped segmentation network whose encoder combines a 3D CNN branch with a 3D vision graph neural network (ViG3D) branch, fuses the two feature streams with channel attention, and employs a 'paperclip-shaped' offset decoder. The method is evaluated on two public coronary CTA datasets, ASOCA and ImageCAS, against nnU-Net, STUNet, UNETR, and SwinUNETR. The authors claim that ViG3D-UNet achieves state-of-the-art accuracy and, as the central contribution, superior vascular connectivity, supported by precision, HD95, and visual examples. Ablation studies examine the graph neighbor count K, the number of ViG3D blocks per scale, and the contribution of each architectural module.","tokens_in":15894,"tokens_out":5107,"duration_ms":43732,"significance":"If the connectivity claim were established, this would be a useful empirical contribution: it offers an end-to-end graph-based alternative to two-stage graph segmentation pipelines and reports competitive accuracy on two public datasets against strong baselines. The architectural idea is clearly presented, and the use of public datasets plus a fair baseline setup is commendable. However, the paper's central claim—connectivity improvement—is supported only by proxy metrics, the reported accuracy gains over the strongest baseline are small, and no significance testing is provided. With direct topological metrics and proper statistical evaluation, the work could be a solid contribution; in its current form the headline claim is overstated.","major_comments":[{"comment":"The central claim that ViG3D-UNet 'surpassed competing methods in maintaining vascular segmentation connectivity' is not supported by the reported metrics. Precision (Eq. 11) counts voxel-level false positives, and HD95 (Eq. 9) measures surface distance; neither quantifies topology. A prediction can be highly precise and have small HD95 while still containing broken branches, disconnected components, or missing distal segments. The lower recall of ViG3D-UNet relative to STUNet on both datasets (ASOCA 79.65 vs 82.31; ImageCAS 80.79 vs 83.51) is consistent with omitted branches, which is a direct connectivity failure. I request direct connectivity metrics over the full test sets: number of connected components per coronary tree, fraction of ground-truth branch endpoints covered, centerline overlap or Dice on centerlines, and/or Betti-number error. The anecdotal visual evidence in Figures 6 and 7 should also be replaced or supplemented by these quantitative topology measurements.","section":"Section IV.A.3 and Section IV.D, Table V"},{"comment":"No variance or significance testing is reported for the comparison against baselines. Table II shows the proposed method's standard deviations are large relative to the reported differences (e.g., ImageCAS DSC 81.72 ± 5.09 vs STUNet 81.54; HD95 15.21 ± 18.53 vs STUNet 17.97). Since the improvements over the strongest baseline are 0.57 DSC and 2.39 HD95 on ImageCAS, the reader cannot determine whether these differences are statistically reliable. Please report per-fold or per-case results and apply paired tests (e.g., Wilcoxon signed-rank or paired t-test on per-case DSC and HD95) across the five folds or on the test set, and make the figures show these comparisons explicitly.","section":"Section IV.D, Table V and Fig. 5"},{"comment":"The hyperparameters K=7 and the per-scale ViG3D block counts (2, 4, 16, 2) are selected by experiments on the ImageCAS dataset, which is also the dataset used for the final head-to-head comparison. This makes the reported ImageCAS numbers a selection outcome rather than an unbiased evaluation. The authors should either perform nested validation on ImageCAS or demonstrate that the same hyperparameters are optimal on ASOCA before claiming generalizable connectivity improvement. Additionally, the paper's own limitation—'a blind search of feature distances is involved in the current spatial graph representation'—acknowledges that the graph construction is not informed by vessel structure, which weakens the mechanistic claim that the graph module specifically encodes vascular connectivity.","section":"Section IV.A.2 and Section IV.C"},{"comment":"The experimental protocol is described inconsistently. Section IV.A.1 states that ASOCA is 'strategically divided into 30 samples for training and 10 samples for testing' and ImageCAS into 800/200, while Section IV.A.2 says 'A five-fold cross-validation strategy was used for experiments on both datasets.' These statements cannot both describe the identical protocol. Please clarify whether the reported numbers are from a fixed split or from an aggregation of five folds, and if five-fold cross-validation was used, describe how the folds were constructed and why the dataset descriptions mention fixed splits.","section":"Section IV.A.1 and Section IV.A.2"}],"minor_comments":[{"comment":"The column header 'Recall↓' is incorrect; recall is a higher-is-better metric and should be marked with an upward arrow.","section":"Table V"},{"comment":"In the sentence 'For each voxel vi, its neighboring voxels N (vj) were identified using the K-nearest neighbors method,' the notation N(vj) should be N(vi).","section":"Section III.A, after Eq. (3)"},{"comment":"The text says the impact of different numbers of ViG3D blocks is shown in Table IV, but the ViG3D unit scaling results are in Table III; Table IV contains the structural ablations.","section":"Section IV.C"},{"comment":"The sentence 'Experiments of a five-fold cross-validation strategy on the ASOCA and ImageCAS datasets, respectively' is grammatically incomplete.","section":"Section IV.B"},{"comment":"The figure captions concatenate model names without spaces (e.g., 'nnUNet STU-Net Swin-UNETRUNETR ViG3D-UNetGT'); these should be separated for readability.","section":"Figures 6 and 7"},{"comment":"The table title contains a stray closing parenthesis: 'WITH THE STATE-OF-THE-ART ALGORITHMS . )'.","section":"Table V title"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear architecture and a fair baseline setup on public data, but the headline claim is connectivity improvement, and that claim is not currently measured. The main fix is straightforward in principle: add direct topology metrics (connected components, endpoint coverage, centerline overlap) and paired significance tests. If the authors cannot show a connectivity benefit with such metrics, the paper should be reframed around the accuracy results and the connectivity claim dropped. The inconsistent split/cross-validation description also needs correction before any revision can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a reasonable new combination and the paper is honest about its limitations, but the abstract's connectivity claim is not actually tested. The experiments use precision and HD95 as proxies for vascular continuity, and no direct topology metric appears anywhere. The DSC gains over STUNet are about 0.2 on ImageCAS and actually negative on ASOCA, all within the reported standard deviations.\n\nWhat's new: a 3D vision graph branch running in parallel with a CNN encoder, fused by channel attention, with a paperclip-shaped offset decoder that concatenates fusion features to the first four decoder layers and texture features to the last two. This is a genuinely different combination from the cited graph-based vessel segmentation works, which are mostly two-stage cascades. The method is evaluated on two public datasets (ASOCA, ImageCAS) against four standard baselines, and the ablations isolate the contribution of each module. The paper is also transparent about its limitations: limited data, blind graph search, and it reports standard deviations.\n\nThe soft spot is the central claim. Section IV.A.3 asserts that precision 'indicates the algorithm's ability to maintain connectivity,' which is not true in general; a high-precision prediction can still have broken branches. HD95 measures boundary distance, not topology. The recall numbers actually go the other way: STUNet has higher recall on both datasets (ASOCA 82.31 vs 79.65; ImageCAS 83.51 vs 80.79), which is more consistent with ViG3D-UNet missing distal branches. The improvements that do exist are small relative to the standard deviations, and no paired significance test is reported. The K=7 and block-count ablations were done on the same dataset used for the final comparison, so there is some selection bias, though this is common in architecture papers. The visual comparisons in Figs. 6 and 7 are suggestive but not quantified.\n\nBottom line: this is a plausible incremental architecture and the experiments are not cherry-picked, but the headline contribution ('connectivity-aware segmentation') is not empirically established. The fix is straightforward: report connected-component counts, branch coverage, centerline continuity (e.g., overlap with ground-truth centerlines before and after thinning), and paired significance tests for the key metrics. Code availability would also help.\n\nWho is it for: people working on tubular structure segmentation and graph-based medical image architectures; it's a useful empirical datapoint and a reasonable starting point for follow-up work. I'd send it to review, but I'd ask for a direct connectivity evaluation before acceptance.","headline":"Architecturally plausible and honestly reported, but the connectivity headline is not actually measured: the evaluation uses precision and HD95 as proxies, and the gains over STUNet are within the reported noise.","tokens_in":16515,"tokens_out":2521,"would_cite":false,"duration_ms":22655,"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 3D vision-graph U-Net produces coronary artery segmentations that stay connected while matching or beating four baselines on two CTA datasets.","keywords":["coronary artery segmentation","3D vision graph neural network","vascular connectivity","computed tomography angiography","U-Net","graph representation","medical image segmentation","tubular structure segmentation"],"falsifier":"Compute the number of connected components, the fraction of ground-truth branch endpoints covered by the prediction, or the centerline continuity on the ASOCA and ImageCAS test folds; if ViG3D-UNet does not improve these direct measures over the baselines at matched thresholds, the central connectivity claim is falsified.","tokens_in":15444,"feed_emoji":"🫀","tokens_out":11728,"duration_ms":87542,"temperature":0.7,"pith_summary":"The paper proposes ViG3D-UNet, a U-shaped segmentation network that adds a 3D vision graph neural network branch alongside a convolutional encoder, so that vessels are treated as connected graphs of voxels during training rather than only as textured volumes. The authors aim to show that this graph representation reduces the fragmentation and missing distal branches that plague coronary artery segmentation from CTA while preserving, or even improving, overlap accuracy. On the large ImageCAS dataset they report the best DSC (81.72), HD95 (15.21), IoU (69.40), precision (83.16), and recall (80.79) among four competing baselines; on the smaller ASOCA dataset they report the best HD95 (13.39) and precision (90.39) and second-best DSC (84.23). A sympathetic reader would care because coronary segmentation feeds reconstruction and stenosis quantification, where a few breaks can change the diagnosis.","feed_headline":"A graph branch keeps coronary vessels connected","feed_subtitle":"ViG3D-UNet reports fewer vessel breaks and best HD95 on ASOCA and ImageCAS versus four baselines.","key_machinery":"The load-bearing object is the ViG3D block, a volumetric graph-aggregation unit. A stem reduces an input patch to $\\frac{H}{4} \\times \\frac{W}{4} \\times \\frac{D}{4}$, each voxel feature becomes a graph node, and each node is connected to its $K=7$ nearest neighbors; the block then applies max-relative graph convolution, $g(\\cdot) = [x_i, \\max(\\{x_j - x_i \\mid j \\in \\mathcal{N}(x_i)\\})]$, followed by an MLP update and a feed-forward network, stacked in a pyramid of 2, 4, 16, and 2 blocks across scales. The graph aggregation is what transmits same-branch information between nearby vascular voxels, while the parallel CNN branch supplies texture; channel attention fuses the two, and the paperclip-shaped offset decoder concatenates early CNN texture features into the last decoder layers and fusion features into the first four layers to avoid redundant 3D computation.","core_discovery":"In the paper's own terms, the central discovery is that embedding a 3D Vision GNN module as one branch of a parallel encoder, fusing its topological features with CNN texture features through channel attention, and decoding with a paperclip-shaped offset decoder produces volumetric coronary artery segmentations that maintain connectivity across branches and at endpoints better than standard CNN and transformer baselines. The evidence is the score table: on ImageCAS the method has the best DSC, HD95, IoU, precision, and recall among the four compared methods, with DSC 81.72 against 81.54 for the strongest baseline and HD95 15.21 against 17.60 for the self-configuring U-Net baseline; on ASOCA it has the best HD95 and precision and second-best DSC (84.23, within 0.57 of the strongest baseline). The authors interpret the lower HD95 and higher precision, together with qualitative figures showing fewer breaks, as evidence of superior connectivity.","pith_inferences":["A direct topological evaluation—connected-component count, fraction of ground-truth branch endpoints recovered, or centerline continuity—would be a sharper test of the connectivity claim, and the paper does not report such metrics.","If the connectivity gain is real, the same graph-in-U-Net design should transfer to other sparse tubular structures such as airways, retinal vessels, and cerebral arteries, since the paper tests only coronary arteries.","The optimal neighbor count $K=7$ and the stacked ViG3D configuration were tuned on ImageCAS; a testable extension is whether the best $K$ scales with voxel anisotropy or branch thickness across datasets.","The lower standard deviation on the small ASOCA dataset hints that graph aggregation regularizes the model under limited training data, which could be checked by training on progressively smaller training subsets."],"forward_implications":["On ImageCAS, ViG3D-UNet reports the best DSC (81.72), HD95 (15.21), IoU (69.40), precision (83.16), and recall (80.79) among the four compared baselines.","On ASOCA, it reports the best HD95 (13.39) and precision (90.39) and second-best DSC (84.23), suggesting the graph prior helps when training data are scarce.","Ablation studies attribute the gain to all three components: removing the ViG3D module lowers DSC by 0.61 and raises HD95 by 2.39, removing channel attention lowers DSC by 0.51, and removing the offset decoder lowers DSC by 1.16 and raises HD95 by 3.13.","Because the pipeline is end-to-end, it avoids the two-stage graph-construction-then-refinement cascade of earlier graph-based vessel segmentation, which should simplify training."],"supporting_citations":[{"why":"Supplies the self-configuring U-Net training pipeline and serves as the primary CNN baseline on both datasets.","marker":"[7]"},{"why":"Provides the scalable U-Net baseline with comparable model scale whose scores ViG3D-UNet is claimed to beat on ImageCAS and nearly match on ASOCA.","marker":"[17]"},{"why":"Supplies the transformer-encoder baseline compared on both datasets.","marker":"[15]"},{"why":"Supplies the Swin-transformer baseline compared on both datasets.","marker":"[9]"},{"why":"Provides the ASOCA challenge dataset of 40 cardiac CTA volumes used for the small-scale evaluation.","marker":"[38]"},{"why":"Provides the ImageCAS dataset of 1000 CTA volumes used for the large-scale evaluation.","marker":"[39]"},{"why":"Introduces the Vision GNN graph representation that the ViG3D module adapts to 3D, including the over-smoothing concern used to justify the choice of K.","marker":"[27]"},{"why":"Supplies the max-relative graph convolution operator used in the 3D graph aggregation.","marker":"[35]"}],"fun_headline_variants":["3D graph module stitches broken coronary vessels in scans","Graph neural net reduces vessel breaks in coronary imaging","Connectivity-aware segmentation via graph branch improves continuity","Volumetric graph representation boosts vascular connectivity","Graph-enhanced U-Net maintains coronary vessel continuity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The connectivity claim rests on treating precision and HD95 as measures of vascular connectivity; if those two metrics do not track true branch continuity, the paper has not actually shown better connectivity.","fun_headline_variants_meta":{"raw":{"variants":["3D graph module stitches broken coronary vessels in scans","Graph neural net reduces vessel breaks in coronary imaging","Connectivity-aware segmentation via graph branch improves continuity","Volumetric graph representation boosts vascular connectivity","Graph-enhanced U-Net maintains coronary vessel continuity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1538,"prompt_tokens":947,"completion_tokens":591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":520}},"tokens_in":563,"tokens_out":591,"duration_ms":6196,"temperature":1.0,"reasoning_tokens":520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:04:17.539635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the number of connected components, the fraction of ground-truth branch endpoints covered by the prediction, or the centerline continuity on the ASOCA and ImageCAS test folds; if ViG3D-UNet does not improve these direct measures over the baselines at matched thresholds, the central connectivity claim is falsified.","supporting_citations":[{"cited_title":"The attention mechanism [19] was introduced for more effective aggregation of contextual information","cited_arxiv_id":null,"evidence_quote":"Provides the scalable U-Net baseline with comparable model scale whose scores ViG3D-UNet is claimed to beat on ImageCAS and nearly match on ASOCA."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Swin-transformer baseline compared on both datasets."},{"cited_title":"Graph convolution based cross-network multiscale feature fusion for deep vessel segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the ASOCA challenge dataset of 40 cardiac CTA volumes used for the small-scale evaluation."},{"cited_title":"Adaptive context selection for polyp segmentation,","cited_arxiv_id":null,"evidence_quote":"Introduces the Vision GNN graph representation that the ViG3D module adapts to 3D, including the over-smoothing concern used to justify the choice of K."}],"review_version":1}