{"id":"f54f8909-5789-45cb-9c68-b526c4980690","arxiv_id":"1908.05343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A graph convolutional network predicting radial vertex locations on a tubular mesh segments coronary artery lumens in cardiac CT, outperforming a same-sized fully connected baseline.","lead":"This paper uses graph convolutional networks to predict the positions of mesh vertices outlining coronary arteries in CT scans, producing smooth artery surface models. It tests the method on 78 artery segments from a public benchmark and shows it beats a fully connected baseline in overlap and surface distance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GCN-over-MLP advantage is asserted from pooled mean metrics; without per-segment paired statistics or patient-clustered error bars, the central claim is not yet established.","rationale":"I read the paper as a solid, clearly described application of GCNs to coronary lumen segmentation, with a fair parameter-matched baseline and evaluation on a public challenge framework. The reader's conditional verdict seems right, but I would anchor it on a different weak point. The centerline dependence flagged by the reader is a genuine limitation and is explicitly acknowledged in the Discussion; however, because both the GCN and the MLP use the same automatically extracted centerlines, it does not directly threaten the comparative claim that GCN layers help. What most directly threatens that claim is the absence of any uncertainty quantification around the reported differences. With only 78 segments clustered in 18 patients, pooled means can hide large per-patient variability, and no code or per-segment predictions are available to check this. A paired, patient-clustered statistical test would settle the question. If the differences survive that test, the central claim is credible; if not, the conditional verdict should be reconsidered. I therefore keep the verdict conditional while shifting the burden of proof to the statistical comparison.","tokens_in":5929,"tokens_out":5048,"duration_ms":53650,"concrete_test":"Use the leave-one-patient-out predictions to compute per-segment paired DSC, MSD, and HD for the GCN and MLP, then test the paired differences with a patient-level clustered bootstrap or a Wilcoxon signed-rank test on the 78 segment pairs. If the 95% CI for the mean difference includes zero for DSC or MSD, or the paired p-value exceeds 0.05, the claimed improvement over the baseline is not statistically supported. As a secondary check, rerun both networks with 5 different seeds and report mean plus/minus SD over seeds to confirm the effect size exceeds training noise.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's central claim is that adding mesh interaction via GCN layers improves overlap and accuracy over an MLP baseline. Table 1 supports this only with pooled mean DSC/MSD/HD over 78 segments from 18 patients (DSC 0.75 vs 0.67 healthy, 0.73 vs 0.69 diseased; MSD 0.25 vs 0.32 mm and 0.28 vs 0.31 mm; HD 1.53 vs 1.59 mm and 1.86 vs 1.84 mm). No standard deviations, confidence intervals, or paired significance tests are reported, and the leave-one-patient-out design means segments from the same patient are correlated. The improvements in MSD are small relative to voxel size (in-plane resolution 0.29-0.43 mm), so measurement noise could plausibly account for them. The centerline limitation acknowledged in the Discussion is real, but it affects the GCN and MLP arms equally and is therefore not the decisive weak point for the comparative claim. The load-bearing unverified step is statistical: are the observed GCN-vs-MLP differences reproducible across patients and seeds? The manuscript does not provide the per-segment values needed to check this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph convolutional network (GCN) for coronary artery lumen segmentation in cardiac CT angiography. A tubular surface mesh is constructed around an automatically extracted centerline, with each vertex described by a fixed angle and a predicted radial distance. Image features are extracted along rays at each vertex, and a five-layer GCN with a mean aggregator predicts the radial distance, propagating information between neighboring mesh vertices. The method is evaluated on 78 coronary artery segments from 18 CCTA volumes of the public Coronary Artery Stenoses Detection and Quantification Evaluation Framework using leave-one-patient-out cross-validation. The authors report DSC of 0.75/0.73, MSD of 0.25/0.28 mm, and HD of 1.53/1.86 mm for healthy/diseased segments, and compare against an MLP baseline with the same number of parameters and training procedure to support the claim that mesh information improves segmentation overlap and accuracy.","tokens_in":6188,"tokens_out":2515,"duration_ms":27937,"significance":"If the central claim is established, the work is a useful contribution to coronary lumen segmentation: it addresses the need for contiguous, sub-voxel, mesh-based segmentations for downstream stenosis assessment and CFD, and it demonstrates that graph convolutions on a tubular mesh can serve as an intrinsic smoothness mechanism, avoiding post-processing such as graph cuts or CRFs. The experimental design has notable strengths: evaluation on a public challenge framework, leave-one-patient-out cross-validation, automatic centerline extraction, and a controlled ablation where the GCN and MLP have identical trainable parameter counts and training setup. However, the reported quantitative support for the central GCN-over-MLP claim is limited to pooled mean metrics without measures of variability or statistical testing, and the patient-level correlation of segments is not accounted for. The dataset is also small (18 volumes), so reproducibility across patients and training seeds remains unquantified.","major_comments":[{"comment":"The central comparative claim that the GCN improves segmentation over the MLP baseline is not supported by inferential statistics. The reported DSC, MSD, and HD values are pooled means over 78 segments from only 18 patients, with no standard deviations, confidence intervals, or paired significance tests. Because leave-one-patient-out cross-validation is used, segments from the same patient are correlated, so treating all 78 segments as independent is inappropriate. Moreover, the MSD differences (0.07 mm for healthy and 0.03 mm for diseased segments) are smaller than the in-plane voxel size (0.29–0.43 mm), so measurement noise could plausibly account for them. Please provide per-segment or per-patient paired results, a paired test such as Wilcoxon signed-rank or a patient-clustered bootstrap, and state whether multiple training seeds were used.","section":"§4, Table 1"},{"comment":"The abstract and Section 5 state that mesh information improves segmentation 'accuracy', but the Hausdorff distance results do not consistently support this: in diseased segments the GCN has HD 1.86 mm versus 1.84 mm for the MLP, i.e. slightly worse. The claim should be qualified to overlap (DSC) and mean surface distance, or the HD comparison should be tested and reported separately. As written, the blanket statement overstates the evidence in Table 1.","section":"§4, Table 1"},{"comment":"The Discussion acknowledges that 'the exact location of this coronary artery centerline can lead to noticeable differences in segmentation accuracy.' This sensitivity is not quantified in the paper. Although the centerline is a shared input to both the GCN and MLP arms, so it does not by itself undermine the relative comparison, it does affect the absolute accuracy claims and the practical utility of the method. Please report the magnitude of this sensitivity or provide evidence that the reported results are stable with respect to small centerline perturbations.","section":"§5, Discussion"}],"minor_comments":[{"comment":"The text says the inclusion of GCN layers leads to 'substantially higher overlap (DSC) and better accuracy (MSD)'; this wording is appropriate only for DSC and MSD, not for HD, and could be adjusted to avoid the inconsistency noted in the major comments.","section":"§4, Table 1"},{"comment":"The abstract and Section 2 say that surface meshes enclosing the full coronary artery tree were automatically extracted, but the evaluation and training are described per segment. Please clarify whether the method is applied to the full tree and then scored per segment, or trained per segment, so that the unit of analysis is unambiguous.","section":"§2-§4"},{"comment":"The description of training says each network was trained for 50,000 iterations, but no information is given about random seed(s) or the number of repetitions. A brief statement about seed sensitivity would help assess reproducibility, especially given that the reported differences are small.","section":"§4"},{"comment":"Minor typographical slips include 'Table 1 lists results the three expert observers' (missing 'of') and the rendering of one author's surname as 'Iˇsgum'; these should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid short-paper-style contribution with a fair controlled ablation, but the central claim currently rests on pooled means without error bars or significance testing on a small, clustered dataset. This is fixable within the manuscript's scope by adding per-segment/per-patient paired statistics and seed information, so I recommend major revision rather than rejection. I would also suggest the authors carefully qualify the HD-related wording, since Table 1 shows a slightly worse diseased HD for the GCN."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a legitimate and cleanly executed application of GCNs to coronary lumen surface mesh prediction, with a fair MLP baseline that keeps parameter count and training procedure identical. Second, the central claim—that adding mesh information via GCN layers improves accuracy—rests on pooled mean metrics from 78 segments in 18 patients, with no error bars or paired significance tests. The evidence is suggestive, not conclusive.\n\nWhat's actually new: tubular mesh vertex regression has been around (Lugauer, Freiman), and GCNs have been used for airways and cortex, but the combination—regressing radial vertex positions on a coronary lumen mesh with a GCN—is new. The controlled comparison is a real strength: same data, same 14,567 parameters, same training procedure, so the difference isolates the graph neighborhood aggregation. The method is also practically attractive: no post-processing smoothing, and it produces contiguous sub-voxel meshes that compare well with prior work and human experts on the public challenge.\n\nWhere it's soft. The GCN-over-MLP advantage is reported as means over segments, but segments from the same patient are correlated, and no per-segment values, confidence intervals, or paired tests are given. The MSD differences (0.25 vs 0.32, 0.28 vs 0.31 mm) are small relative to in-plane resolution (0.29–0.43 mm), so measurement noise could matter. The Hausdorff distance is actually slightly worse for GCN in diseased segments (1.86 vs 1.84), so the \"more accurate\" claim is really carried by DSC and MSD. The acknowledged centerline sensitivity is real, but it affects both arms equally, so it doesn't undermine the comparative claim. The dataset is small (18 volumes), which is normal for this challenge, but it means the reproducibility of the GCN-vs-MLP gap across patients and seeds is untested. No code is released, which makes it hard to check the statistical robustness yourself.\n\nNet: a solid conference-level contribution for the MICCAI-type audience. It deserves a serious referee: the method is clearly described, the evaluation is fair as far as it goes, and the missing statistics are a tractable revision, not a fundamental flaw. I'd ask for per-segment numbers, patient-clustered error bars, and ideally code or external validation. I wouldn't desk-reject it.","headline":"A clean, fairly evaluated GCN application to coronary lumen meshing whose central claim over the MLP baseline needs per-segment statistics before I'd trust the gap.","tokens_in":6679,"tokens_out":2020,"would_cite":false,"duration_ms":18837,"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":"Coronary lumen segmentation improves when the network also sees the mesh graph.","keywords":["graph convolutional networks","coronary CT angiography","coronary lumen segmentation","tubular surface mesh","centerline extraction","medical image segmentation","cardiac imaging","deep learning"],"falsifier":"Take a CCTA image and displace the extracted centerline by 0.5 mm in random directions, then rerun the trained GCN; if the Dice similarity coefficient or mean surface distance degrades materially, the reported accuracy is largely an artifact of centerline placement rather than a robust lumen-surface estimate.","tokens_in":5757,"feed_emoji":"❤️","tokens_out":6633,"duration_ms":63094,"temperature":0.7,"pith_summary":"This paper tries to show that a graph convolutional network operating directly on a tubular surface mesh can segment the coronary artery lumen from cardiac CT angiography with sub-voxel accuracy and without post-processing smoothing. The authors model the vessel wall as a fixed tube around an automatically extracted centerline, and the network regresses one radial distance per mesh vertex from image intensity rays combined with features of neighboring vertices. On 78 coronary segments from a public evaluation framework, the method reaches a Dice similarity coefficient of 0.75 in healthy and 0.73 in diseased segments, with mean surface distances around 0.25 to 0.28 mm. The key empirical claim is that adding GCN layers improves overlap and accuracy over an equally sized multi-layer perceptron that sees the same rays but no mesh neighbors, producing smoother and more accurate surfaces.","feed_headline":"Coronary lumen meshes improve when the network sees the mesh","feed_subtitle":"Sharing vertex information across the surface mesh, the GCN beats an equally sized network without mesh links.","key_machinery":"The load-bearing object is a tubular surface mesh $G(V,E)$ built around an automatically extracted coronary centerline: each vertex is anchored to a centerline point and a fixed angle $\\varphi_v$ in a cross-sectional plane, so the only free parameter per vertex is the radial distance $r_v$. A five-layer graph convolutional network, using the element-wise mean aggregator of GraphSAGE, propagates hidden features across the mesh edges and regresses $r_v$ from 32 image-intensity samples along a ray. The network is trained with the cubed-distance loss $(1/|V|)\\sum_{v\\in V} |r_v^3 - f(x_v)^3|$. This machinery carries the argument because neighborhood propagation across the fixed mesh graph is what produces smooth, contiguous surfaces without explicit regularization or post-processing; removing it, as in the MLP baseline, lowers overlap and accuracy.","core_discovery":"The central discovery, on the paper's own terms, is that the inclusion of mesh information through GCN layers improves coronary lumen segmentation over a baseline without mesh interaction. Instead of classifying voxels, the network solves a regression problem for each vertex of a tubular surface mesh: given a fixed centerline, each vertex is pinned to a fixed angle in a cross-sectional plane, and the network predicts the single radial distance $r_v$ to the lumen wall. Input features are image intensities sampled along rays, and five GCN layers propagate hidden representations across the mesh graph so that each vertex prediction is informed by vertices up to five steps away. This yields Dice similarity coefficients of 0.75/0.73 and mean surface distances of 0.25/0.28 mm on healthy and diseased segments, and a controlled comparison with a multi-layer perceptron using the same parameters shows substantially higher overlap and accuracy for the GCN.","pith_inferences":["Inference: Because the graph is fixed by the centerline and angular sampling, the same trained network could in principle generate meshes at different resolutions by re-sampling the graph without retraining; the paper does not test this.","Inference: The mean aggregator ignores spatial relations between neighbors; adding edge features that encode angular or geodesic offsets, a direction the authors flag, could make the network sensitive to local curvature and stenosis shape.","Inference: The explicit admission that centerline location causes noticeable accuracy differences suggests a testable design: jointly optimize centerline extraction and radial-distance regression rather than treating them as separate stages."],"forward_implications":["If the central claim holds, coronary lumen segmentation for stenosis detection and blood-flow simulation can be produced directly as contiguous sub-voxel surface meshes, without graph-cut smoothing or explicit calcium removal.","The controlled MLP comparison shows that neighborhood propagation on the mesh, not just the ray features, is responsible for the improvement in overlap and mean surface distance.","Because the method directly outputs a surface rather than voxel labels, the architecture itself supplies the smoothness that earlier tubular-mesh methods needed additional penalties or post-processing to enforce.","The authors identify centerline dependence as a shared limitation, so accuracy on a new scan is bounded by the quality of the automatically extracted centerline."],"supporting_citations":[{"why":"Supplies the public challenge framework with the 78 annotated coronary segments, expert reference meshes, and evaluation metrics used in all experiments.","marker":"[5]"},{"why":"Provides the deep-learning centerline extraction that defines the tubular graph for every image in the study.","marker":"[13]"},{"why":"Is the earlier boosting-tree lumen-segmentation baseline that the GCN is compared against and that required graph-cut smoothing.","marker":"[10]"},{"why":"Is the tubular-mesh segmentation baseline that enforces smoothness by penalizing jumps between neighboring vertices.","marker":"[2]"},{"why":"Supplies the GraphSAGE element-wise mean aggregator that defines the GCN layer and its neighborhood propagation.","marker":"[3]"},{"why":"Introduces the graph convolutional network formulation that the architecture is built on.","marker":"[4]"}],"fun_headline_variants":["Mesh-aware GCN sharpens coronary artery segmentation","Graph networks leverage mesh links for coronary lumen","Coronary meshes gain accuracy with graph convolution","Mesh connectivity boosts GCN coronary segmentation","GCNs plot coronary walls better with mesh context"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes the coronary artery centerline is accurate enough that constraining each mesh vertex to a fixed angle and a single radial distance around that centerline can capture the true lumen surface; the paper itself notes that the exact centerline location can cause noticeable differences in accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Mesh-aware GCN sharpens coronary artery segmentation","Graph networks leverage mesh links for coronary lumen","Coronary meshes gain accuracy with graph convolution","Mesh connectivity boosts GCN coronary segmentation","GCNs plot coronary walls better with mesh context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000944,"raw_usage":{"total_tokens":4036,"prompt_tokens":954,"completion_tokens":3082,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":3012}},"tokens_in":570,"tokens_out":3082,"duration_ms":19457,"temperature":1.0,"reasoning_tokens":3012,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:15:57.148556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a CCTA image and displace the extracted centerline by 0.5 mm in random directions, then rerun the trained GCN; if the Dice similarity coefficient or mean surface distance degrades materially, the reported accuracy is largely an artifact of centerline placement rather than a robust lumen-surface estimate.","supporting_citations":[{"cited_title":"Med Image Anal 17(8), 859–876 (2013)","cited_arxiv_id":null,"evidence_quote":"Supplies the public challenge framework with the 78 annotated coronary segments, expert reference meshes, and evaluation metrics used in all experiments."},{"cited_title":"Med Image Anal 51, 46–60 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the deep-learning centerline extraction that defines the tubular graph for every image in the study."},{"cited_title":"In: International MICCAI Workshop on Medical Computer Vision","cited_arxiv_id":null,"evidence_quote":"Is the earlier boosting-tree lumen-segmentation baseline that the GCN is compared against and that required graph-cut smoothing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the tubular-mesh segmentation baseline that enforces smoothness by penalizing jumps between neighboring vertices."},{"cited_title":"In: Adv Neural Inf Process Syst (NIPS)","cited_arxiv_id":null,"evidence_quote":"Supplies the GraphSAGE element-wise mean aggregator that defines the GCN layer and its neighborhood propagation."},{"cited_title":"In: Int Conf Learn Represent (ICLR) (2017)","cited_arxiv_id":null,"evidence_quote":"Introduces the graph convolutional network formulation that the architecture is built on."}],"review_version":1}