{"id":"be3d7844-9884-439b-b231-b56d427cbf65","arxiv_id":"2507.21328","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GLCP jointly learns vessel segmentation, skeleton maps, and local discontinuity maps, improving accuracy and connectivity on 2D and 3D tubular structure benchmarks.","lead":"A new deep learning framework, GLCP, improves segmentation of thin blood vessels by teaching the model to spot and repair small breaks using auxiliary skeleton and discontinuity maps. It could make medical image analysis more reliable for tasks like diagnosing eye disease and planning surgery.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Discontinuity pseudo-labels are generated online from the model's own predictions (Eq. 2), so the reported topology gains are not isolated from self-training noise or extra-head regularization; without ablating the label source, the central claim remains conditional.","rationale":"The central claim is that the proposed discontinuity prediction and DAR improve accuracy and topological continuity. The mechanism that makes this method novel—the adaptive endpoint-guided pseudo-labeling—is the least supported link. Because the pseudo-labels come from the model's own predictions, any improvement from the auxiliary head could reflect the intended effect of learning where fragmentation occurs, a generic regularizer that penalizes isolated endpoints, or an artifact of the threshold selecting false-positive endpoints and DAR amplifying the main prediction. The paper's ablations separate modules but not label sources, and no code or checkpoints are provided to test reproducibility. This does not mean the method is wrong; it means the current evidence is compatible with several simpler explanations. I agree with the reader that the conditional verdict is appropriate, with the pseudo-label source as the key open question. The proposed experiment distinguishes these explanations directly.","tokens_in":184,"tokens_out":6102,"duration_ms":97651,"concrete_test":"Retrain GLCP on STARE and TopCoW-binary with three label conditions: (a) online pseudo-labels as in the paper; (b) static pseudo-labels generated once from the trained nnUNet baseline; (c) randomized discontinuity labels with the same spatial density. Use five seeds per condition and report mean plus/minus standard deviation for Dice, clDice, and Betti error, with paired significance tests where sample sizes allow. If condition (b) or (c) matches (a) within one standard deviation, the dynamic self-referential labeling is not essential to the claimed gains; if (a) clearly dominates both, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 builds the ground-truth discontinuity masks from endpoints of the current predicted skeleton and the GT skeleton. Eq. (1) computes each predicted endpoint's distance to the nearest GT endpoint; Eq. (2) thresholds these distances at mean+std. Thus every training iteration uses labels that are a function of the model's own output, with no teacher-student detachment, no frozen pseudo-label condition, and no report of how the selected endpoint set changes as training proceeds. If early predictions are fragmented, the selected points are mostly false-positive endpoints, so the discontinuity head can improve Dice simply by learning to suppress isolated predictions; the DAR module then uses sigma(Fd) as an attention map, making the refinement partly a self-amplified error map. Table 3 ablates the presence of the Dis head but never varies the label-generation strategy (e.g., static labels from a pretrained network, labels from GT alone, or random labels with matched density), so the Betti-error reduction is confounded with extra capacity and regularization. With test sets of 10 (STARE), 18 (TopCoW), and 4 (CCA) images and no significance testing, the reported gains, while plausible, do not yet establish that the global-to-local connectivity mechanism itself is responsible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLCP, a multi-head framework for tubular structure segmentation. It augments a shared nnUNet or SwinUNETR backbone with a skeleton head and a discontinuity head, plus a lightweight refinement module DAR. The discontinuity head is trained from masks constructed online by detecting endpoints in the current predicted skeleton and thresholding their distance to ground-truth skeleton endpoints (Eqs. 1-2), then expanding selected points into cube windows. A KL consistency loss aligns the skeleton head with skeletons derived from the main segmentation output, and DAR uses skeleton and discontinuity probability maps as attention to refine the prediction. Experiments on STARE, CCA, and TopCoW compare Dice, clDice, Betti error, and HD against several loss functions and architectures, with ablations in Table 3. The authors claim superior accuracy and connectivity across 2D and 3D tubular segmentation tasks.","tokens_in":8731,"tokens_out":4597,"duration_ms":51112,"significance":"If the claimed improvements hold, the framework is a practical contribution: it integrates into nnUNet by adding only four lightweight convolution layers, is evaluated on 2D and 3D datasets including a multi-class setting, and consistently lowers Betti errors in the reported comparisons. The authors also test transfer to SwinUNETR, which supports generalizability. The idea of explicitly localizing discontinuities through endpoint mining is well motivated and clearly presented. However, because the auxiliary discontinuity labels are generated from the model's own output during training, the reported gains are not yet isolated from self-training effects or extra head capacity, and the small test sets without error bars or significance tests limit the strength of the central claim.","major_comments":[{"comment":"The ground-truth discontinuity mask is generated from the current predicted segmentation: predicted skeleton endpoints are compared to ground-truth skeleton endpoints, and endpoints with distance greater than mean plus standard deviation are selected. The figure caption states that SE and EDM are used only for training, but this does not address the more important concern that the pseudo-labels are a function of the model's own output at every iteration. The paper does not ablate the label-generation strategy (e.g., labels from a frozen pretrained network, labels derived from GT alone, or random labels with matched density), nor does it report how the selected endpoint set evolves as training proceeds. Consequently, the Betti error reductions in Table 3 (e.g., STARE from 4.60 to 3.00, TopCoW-multi from 0.57 to 0.39) are confounded with extra head capacity and implicit regularization, so the attribution of the improvement to the global-to-local connectivity mechanism itself is not yet established.","section":"Section 2.1, Eqs. (1)-(2); Section 3.3, Table 3"},{"comment":"The test sets are small (STARE has 10 test images, TopCoW has 18, CCA has 4) and the paper reports no standard deviations, multiple seeds, or significance tests. Several reported differences are small in absolute terms, such as STARE Dice 83.67 versus 83.39 and clDice 87.44 versus 87.11, and TopCoW-binary clDice 95.58 versus 95.57. With single runs, these differences could easily be within run-to-run variation, especially for Betti errors computed on only four CCA images. The authors should provide at least multiple-seed means with standard deviations or statistical significance tests, or temper the 'superior' and 'outperforms' wording accordingly.","section":"Section 3.3, Tables 1-2"},{"comment":"Several free parameters are fixed without sensitivity analysis: the weight factors alpha and beta are both set to 0.5, the cube window size Ri is set to one-eighth of the patch size, and the dynamic threshold in Eq. (2) is fixed as mean plus standard deviation, with DBSCAN parameters unspecified. Since the threshold and window size directly determine the auxiliary discontinuity supervision, the reported improvements may rely on a favorable choice of these parameters. A sensitivity study over reasonable ranges, at least for the threshold formula and the loss weights, would demonstrate robustness and make the method easier to adopt on new datasets.","section":"Section 3.2, Eq. (2)"}],"minor_comments":[{"comment":"In the definition of the cube window, the third coordinate condition is written as '|y - pz_i|' but should presumably be '|z - pz_i|'; please correct this typo.","section":"Eq. (3)"},{"comment":"The symbol '⊗' is not defined; specify whether it denotes element-wise multiplication or some other tensor operation, since it is used in both the consistency loss and the refinement formulation.","section":"Eq. (4) and Eq. (5)"},{"comment":"The dataset name is spelled 'TopCoW' in the abstract and most of the text but 'ToPCoW' in Tables 1 and 2; please unify the spelling.","section":"Tables 1-2 and text"},{"comment":"The text says Table 1 compares five other state-of-the-art loss functions, but the table also includes a multi-decoder framework and a transformer-based architecture; please adjust the wording to match the table content.","section":"Section 3.3"},{"comment":"For the 3D datasets, please state whether Betti errors and HD are computed per-case and then averaged or computed on aggregated volumes, so that the metrics are reproducible.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the self-referential pseudo-label mechanism in Section 2.1: the discontinuity head is trained from the model's own predicted endpoints, and the absence of an ablation on the label source leaves the central attribution claim unproven. I would ask the authors to add a detached-teacher or static-label baseline and to report multi-seed results, especially for the small test sets, before the paper can be recommended for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a decent, careful method paper. The genuinely new bit is the auxiliary discontinuity head that generates labels from predicted skeleton endpoints, plus a light dual-attention refinement. It reports consistent gains across 2D and 3D datasets and two backbones, with ablations showing each component helps. I think the core direction is sound.\n\nWhat is good: the problem is well-motivated (fragmentation in tubular structures), and the endpoint-guided label generation is a clean idea. The evaluation uses relevant topology metrics (clDice, Betti error) and compares against several recent losses. The ablation in Table 3 suggests the discontinuity task is the main driver of the Betti error reduction, and the DAR module adds a bit. That is useful evidence.\n\nSoft spots, in order of importance. First, the discontinuity ground truth is computed online from the model's own predicted skeleton (Eq. 2). The stress-test note is correct: there is no frozen-label condition, no comparison to static or GT-only labels. So part of the improvement could be self-training dynamics or just extra regularization rather than the specific label design. Table 3 only toggles the head on/off; it does not vary the label source. That is a real gap for a paper whose central claim is about the mechanism. Second, the test sets are very small (4 CCA, 10 STARE, 18 TopCoW) and there are no error bars or significance tests; a couple of the reported differences are within what you would expect from noise. Third, hyperparameters like the threshold formula and window size are fixed without sensitivity analysis. None of this is damning, but it keeps the paper at the level of plausible and promising rather than rigorously established.\n\nI would note the code is not yet out despite the promise, which makes replication harder; that is minor if the paper is accepted elsewhere with code.\n\nWho gets value from this: anyone working on topology-preserving segmentation or auxiliary-task training for tubular structures. It deserves a serious referee, not a desk reject. I would send it out, mainly asking for the label-source ablation and some variability assessment.\n\nBest.","headline":"A solid within-subfield method paper that deserves review, with one self-supervised label-source gap that should be probed before publication.","tokens_in":9226,"tokens_out":2305,"would_cite":true,"duration_ms":25437,"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 claims that structural fragmentation in vessel segmentation is best reduced by making the model predict its own break-prone spots—redundant endpoints in its predicted skeleton—and then refine the mask with those discontinuity…","keywords":["Tubular structure segmentation","Vascular segmentation","Connectivity preservation","Discontinuity map","Skeleton prediction","Dual-attention refinement","nnU-Net","Betti error"],"falsifier":"Train GLCP with the discontinuity masks replaced by labels derived only from ground-truth skeletons (or with the dynamic threshold shifted to mean, mean plus 0.5 standard deviations, and mean plus two standard deviations); if the reported Betti-error gains persist across these variations, the self-referential pseudo-labeling is robust, and if they collapse, the improvement is an artifact of the self-generated targets.","tokens_in":8261,"feed_emoji":"🫀","tokens_out":9027,"duration_ms":93006,"temperature":0.7,"pith_summary":"This paper tries to establish that structural fragmentation in tubular structure segmentation—the broken, gap-riddled vessel masks that plague vascular image analysis—is best fixed by teaching the model to predict where breaks occur, not just by imposing global topology losses. It proposes GLCP, which adds three parallel heads to a standard segmentation network: one for the global segmentation, one for a skeleton map, and one for a local discontinuity map whose training targets are mined from the model's own predicted skeleton endpoints. A self-supervised consistency loss aligns the skeleton derived from the segmentation output with the skeleton head's prediction, and a dual-attention refinement module fuses the skeleton and discontinuity maps into the final mask. On retinal (STARE), coronary CTA (CCA), and Circle of Willis (TopCoW) datasets, the method reports higher Dice and clDice and lower Betti errors than several state-of-the-art topology losses, on both 2D and 3D data.","feed_headline":"New training targets cut vessel-segmentation breaks","feed_subtitle":"A multi-head framework learns where vessel skeletons break and refines the map, beating baseline losses on retinal, coronary, and brain…","key_machinery":"The central mechanism is an endpoint-guided discontinuity mining strategy that converts skeletons into training targets for a dedicated head: endpoints are extracted from predicted and ground-truth skeletons via a fixed convolution kernel, distances between them are computed, endpoints with distance greater than the mean plus one standard deviation are selected, DBSCAN merges nearby candidates into clusters, and the chosen points are expanded into local cube windows to form the discontinuity ground truth. The Interactive Multi-head Segmentation (IMS) module—three parallel heads on a shared backbone—produces segmentation, skeleton, and discontinuity maps, with a gradient-truncated symmetric KL-divergence consistency loss between the segmentation-derived skeleton and the skeleton head's output. The Dual-Attention-based Refinement (DAR) module then uses sigmoid-activated skeleton and discontinuity maps as attention weights added to the segmentation feature to produce the final refined map.","core_discovery":"The central claim is that local discontinuity awareness, rather than additional global topological constraints, is the missing ingredient for connectivity in tubular structures. To make this concrete, the paper introduces a discontinuity prediction head whose training targets are mined from the network's own current predictions: skeleton endpoints of the predicted mask that lie far from the ground-truth skeleton (more than mean plus one standard deviation of all such distances) are labeled as discontinuities, clustered by DBSCAN, and expanded into small cube regions that the head must learn to flag. A parallel skeleton head learns the global vessel tree, and a symmetric KL-divergence loss with gradient truncation forces the skeleton derived from the segmentation output and the skeleton head's output to agree. The refined segmentation then linearly combines the segmentation with skeleton and discontinuity attention maps. The paper reports that this combination outperforms existing loss-based and multi-task baselines across all datasets, with the discontinuity task contributing the largest share of the topology improvement in ablations.","pith_inferences":["Because the discontinuity labels are generated from the model's own predicted skeleton endpoints, the framework implicitly assumes the model's current errors are visible in its skeleton; an untested alternative would replace these self-derived labels with oracle labels from the ground-truth skeleton to see whether the reported gains shrink.","The predicted discontinuity map at inference could be repurposed as an uncertainty signal for downstream review, flagging exactly where the predicted vessel tree is fragile.","The dynamic threshold (mean plus one standard deviation) is a single design choice; a sensitivity analysis across thresholds (e.g., mean, mean plus 0.5 standard deviations, mean plus two standard deviations) would show whether the method's performance depends on the exact statistics or is robust to them."],"forward_implications":["On STARE, CCA, and TopCoW (binary and multi-class), adding GLCP to nnU-Net V2 and SwinUNETR improves Dice and clDice over the respective baselines and over clDice, cbDice, and skeleton-recall loss additions.","Betti errors (sum of $\\beta_0$ and $\\beta_1$) drop on every tested dataset, indicating fewer disconnected fragments and fewer spurious holes in the predicted vessel topology.","Ablation shows the discontinuity prediction task is the dominant contributor to the topology improvement, with the skeleton task, consistency loss, and DAR refinement each adding smaller gains.","The method transfers to a transformer-based backbone (SwinUNETR), so the benefit is not specific to the convolutional nnU-Net architecture."],"supporting_citations":[{"why":"STARE retinal vessel dataset used for 2D evaluation of the method.","marker":"[5]"},{"why":"Defines the Betti error metric used to quantify topology continuity.","marker":"[6]"},{"why":"Multi-decoder skeleton/edge prediction framework that GLCP's IMS is compared with and inspired by.","marker":"[7]"},{"why":"Supplies the nnU-Net baseline architecture and default loss configuration that GLCP extends.","marker":"[9]"},{"why":"Skeleton recall loss baseline that GLCP compares against.","marker":"[11]"},{"why":"Centerline-boundary Dice loss baseline that GLCP compares against.","marker":"[19]"},{"why":"Provides the clDice metric and loss baseline and the skeleton-based topology-preservation idea GLCP builds on.","marker":"[20]"},{"why":"Supplies skeletonization used to extract skeletons from ground truth and predicted masks for endpoint detection.","marker":"[22]"},{"why":"TopCoW brain CTA benchmark providing 3D binary and multi-class evaluation.","marker":"[24]"},{"why":"CCA coronary artery CTA dataset used for 3D evaluation.","marker":"[25]"}],"fun_headline_variants":["Local break detection fixes vessel segmentation gaps","Spotting breaks improves tubular network segmentation","Discontinuity-aware training mends vessel segmentation splits","Vessel segmentation: learn local breaks, fix global connectivity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The auxiliary discontinuity labels are produced from the model's own predicted skeletons via a dynamic endpoint-distance threshold, assuming those self-generated targets reliably mark real fragmentation and that training on them improves generalization instead of reinforcing the model's current errors.","fun_headline_variants_meta":{"raw":{"variants":["Local break detection fixes vessel segmentation gaps","Spotting breaks improves tubular network segmentation","Discontinuity-aware training mends vessel segmentation splits","Vessel segmentation: learn local breaks, fix global connectivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1379,"prompt_tokens":942,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":558,"tokens_out":437,"duration_ms":5337,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:52:21.359845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GLCP with the discontinuity masks replaced by labels derived only from ground-truth skeletons (or with the dynamic threshold shifted to mean, mean plus 0.5 standard deviations, and mean plus two standard deviations); if the reported Betti-error gains persist across these variations, the self-referential pseudo-labeling is robust, and if they collapse, the improvement is an artifact of the self-generated targets.","supporting_citations":[{"cited_title":"Advances in neural information processing systems32 (2019)","cited_arxiv_id":null,"evidence_quote":"Defines the Betti error metric used to quantify topology continuity."},{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Multi-decoder skeleton/edge prediction framework that GLCP's IMS is compared with and inspired by."},{"cited_title":"In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G","cited_arxiv_id":null,"evidence_quote":"Skeleton recall loss baseline that GLCP compares against."},{"cited_title":"PeerJ 2, e453 (2014) GLCP: Global-to-Local Connectivity Preservation 11","cited_arxiv_id":null,"evidence_quote":"Supplies skeletonization used to extract skeletons from ground truth and predicted masks for endpoint detection."},{"cited_title":"ArXiv pp","cited_arxiv_id":null,"evidence_quote":"TopCoW brain CTA benchmark providing 3D binary and multi-class evaluation."},{"cited_title":"IEEE Transactions on Medical Imaging (2024)","cited_arxiv_id":null,"evidence_quote":"CCA coronary artery CTA dataset used for 3D evaluation."}],"review_version":1}