{"id":"41a51e1a-7327-45a6-af7e-9b49c0847e50","arxiv_id":"2411.15778","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A deep-learning and image-processing pipeline that improves liver vessel skeletonization, separates portal from hepatic venous trees, and introduces a new multi-class liver vessel dataset and morphometry analysis, though with an unaddressed case-count inconsistency (77 claimed, 73 counted).","lead":"This paper presents an automatic pipeline for 3D liver vessel analysis: a learned skeletonization network, a heuristic algorithm that separates the portal and hepatic venous trees, and anatomical branch labeling with morphometry. It also announces a new public multi-class liver vessel dataset built from LiTS and IRCAD cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'clean multi-class segmentation' claim rests on excluding 9/62 LiTS failures and a 77-case total that does not add up from the paper's own numbers; until the separation algorithm is evaluated on all cases and the dataset count is reconciled, the headline overstates support.","rationale":"The reader's weakest assumption correctly identifies the separation algorithm's heuristic nature and the exclusion of 9 failure cases as the soft spot. I agree that this is the most load-bearing concern: the central value proposition is a high-quality multi-class dataset, and the reported improvement in segmentation comes from training on labels produced by that heuristic algorithm. The paper itself provides no quantitative error metric for the separation, and the surgeon validation is limited to the retained subset. The case-count discrepancy (77 claimed vs 73 derivable from the paper's own numbers) is a concrete, checkable sign that the dataset claim is not fully supported. This does not mean the pipeline is without value; the honest negative result on ClDice and the release of a partially validated dataset are useful. But the abstract's unconditional language should be tempered, and the requested checks would determine whether the concern is a fixable reporting issue or a deeper selection-bias problem. I therefore keep the verdict conditional, with conditions focused on full-case evaluation and dataset-count reconciliation.","tokens_in":8179,"tokens_out":3740,"duration_ms":34693,"concrete_test":"Re-run the separation algorithm on all 62 LiTS cases without exclusion, and compare the automatic portal/hepatic labels against the surgeon-corrected labels voxel-by-voxel in the conflict regions; if corrected labels for the 9 excluded cases do not exist, have two surgeons independently label those cases and measure Dice and volume of disagreement. If the automatic labels agree with the corrected labels in dense distal regions at a rate comparable to the retained 53 cases, the exclusion is not hiding a systematic failure; if agreement is substantially lower, or if the final dataset count is not 77, the 'clean low-error 77-case dataset' claim must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 2.2 assigns conflict voxels by connected-component proximity and a fixed 60-degree branch-angle rule (step 9, citing Murray [14]). Section 3.2 reports that 9 of 62 LiTS cases had 'larger errors due to the heavy interconnections of the original segmentation' and were excluded, leaving 53 LiTS volumes; together with 20 IRCAD volumes this gives 73 cases, not the 77 claimed in the Abstract and Section 3.3. Because Table 2 attributes the main segmentation improvement to training on these automatically separated labels, the value of the pipeline depends on those labels being correct. The surgeon validation covers only the 53 retained LiTS cases, with no quantitative per-tree error on the excluded cases or on the final dataset. The abstract's unconditional 'clean... low error' and '77 cases' claims therefore go beyond what is demonstrated: they hold only for a selected subset, and the count is internally inconsistent. This is load-bearing because if the 9 excluded cases are typical of dense distal anatomy, the learned segmentation improvement may reflect selection bias rather than a generally applicable separation algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end pipeline for 3D liver vessel analysis: an nnU-Net-based skeletonization network (NeuralSkel) trained to predict Lee's skeletonization and used in a ClDice-style loss; an algorithm that separates single-label vessel segmentations into portal and hepatic trees using connected components, skeleton analysis, and a 60-degree branch-angle heuristic; automatic anatomical branch labeling; and a new public dataset called LIRCAD. The results show that NeuralSkel approximates Lee's skeleton more closely than morphological soft-skeletonization methods, but that using it inside the ClDice loss does not improve overall segmentation performance. Training on the automatically separated LIRCAD labels improves segmentation metrics by 2-3 points over the IRCAD-only baseline. The paper also reports surgeon-based validation of the separation and a morphometric analysis of the labeled trees.","tokens_in":8289,"tokens_out":4475,"duration_ms":39991,"significance":"If the claims hold, the paper would make three useful contributions: a reproducible skeletonization network for tubular structures, a practical tool for converting single-class vessel annotations into multi-class portal/hepatic labels, and a public dataset with anatomical branch labels and morphometric features. The explicit release of code and data and the surgeon-in-the-loop validation are strengths, as is the transparency about the negative ClDice result. However, the significance is tempered by two load-bearing issues: the skeletonization comparison is partly circular because NeuralSkel is trained on the same Lee skeletonization used as the evaluation reference, and the 'clean, low-error' dataset claim is demonstrated only on a selected subset, since 9 of 62 LiTS cases were excluded and the dataset count is internally inconsistent (77 claimed vs 73 derivable from the paper's own counts).","major_comments":[{"comment":"The abstract and §3.3 state that the new public dataset contains 77 cases, but §3.2 reports 53 LiTS volumes after validation plus 20 IRCAD volumes, which sum to 73. This discrepancy is load-bearing because the dataset size is a headline contribution. Please reconcile the numbers, explain the source of the additional 4 cases, or correct the count to 73.","section":"Abstract and §3.2/§3.3"},{"comment":"The abstract's claim of a 'clean multi-class segmentation of the vessels, validated by surgeons to provide low error' is supported only for the 53 retained LiTS cases: 9 of 62 cases were explicitly excluded because they presented 'larger errors' in dense interconnected regions. No quantitative error metric is reported for the excluded cases or for the full 62-case set, so the claim overstates the algorithm's generality, and the training improvement in Table 2 may be partly attributable to selection. Please evaluate the separation algorithm on all 62 cases, report per-case failure rates, or revise the claim to describe the subset where the heuristics hold.","section":"Abstract and §3.2"},{"comment":"NeuralSkel is trained to predict Lee's skeletonization, and Table 1 compares all skeletonization methods against Lee's output. Consequently, the reported 98.2% Surface Dice for NeuralSkel is a fidelity measure to the training target, not an independent assessment of centerline quality. The Discussion's statement that NeuralSkel gives 'much better centerline representations than the morphological operations' should be rephrased as 'closer to Lee's skeleton' unless an independent reference (e.g., manually traced centerlines or a different skeletonization algorithm) is used.","section":"§2.1 and Table 1"},{"comment":"The separation algorithm relies on a fixed 60-degree branch-angle threshold and connected-component proximity (step 9, citing Murray's law), and the Discussion acknowledges that these fail in dense distal regions. The validation consists of a surgeon's categorical classification ('flawless', 'small mixup', 'larger errors') rather than a quantitative per-voxel or per-tree error. Please provide quantitative metrics (e.g., assignment accuracy, volume overlap against a manual reference) on a per-case basis, including the excluded cases, so that the 'low error' claim is measurable and not just categorical.","section":"§2.2 and §3.2"}],"minor_comments":[{"comment":"The term 'Bersenham line' appears to be a typo for 'Bresenham line'.","section":"§2.2"},{"comment":"The manuscript contains placeholder links 'Code link' and 'Data link'; the final version should provide actual URLs and a clear data availability statement.","section":"Code and data availability"},{"comment":"The automatic anatomical labeling heuristics, such as fusing paths based on Couinaud segments, are described only briefly; a flowchart or pseudo-code would improve reproducibility.","section":"§2.3"},{"comment":"The step numbers in the text (2)-(12) do not clearly correspond to the numbered boxes in the figure; please align the numbering between text and figure.","section":"Figure 2"},{"comment":"The Surface Dice tolerance is fixed at 2 mm; please justify this value or state it explicitly as a tunable parameter.","section":"§3.1"},{"comment":"Training details for NeuralSkel (dataset, number of volumes, augmentation strategy, number of epochs) are not specified; adding these details would improve reproducibility.","section":"§2.1"},{"comment":"The claim of the 'largest public liver vessel dataset to date' should be substantiated by comparing with existing public datasets, such as LiTS vessel annotations, IRCAD, or other published vessel datasets.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical image analysis venue, and the authors should be commended for releasing code and data and for reporting the negative ClDice result. However, the dataset-size inconsistency and the unquantified exclusion of 9/62 LiTS cases are load-bearing and must be resolved before acceptance; the circularity concern with NeuralSkel can likely be addressed by rewording, but the dataset evaluation needs additional quantitative reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid workshop paper with a genuinely useful resource inside it, but the headline claims run ahead of the evidence. The LIRCAD dataset, if actually released, would be a real contribution: 20 IRCAD + 53 LiTS cases with separated portal/hepatic trees and anatomical branch labels. The separation algorithm is clever and the surgeon validation on the successful subset is a legitimately independent check. And credit where due: they report the negative result on NeuralSkel-in-ClDice plainly, which is more than many papers do.\n\nThe soft spots are proportionately significant. The abstract says 77 cases, but the paper's own numbers (53 LiTS + 20 IRCAD) give 73. That is not a rounding issue; it suggests the count was not checked. The 9 excluded LiTS cases were excluded because the separation failed on dense distal interconnections, so the 'clean... low error' statement in the abstract only applies to the 53 cases where the heuristics worked. The surgeon validation covers those, not the 9. That makes the dataset claim selective. The skeletonization benchmark is partially circular—NeuralSkel is trained on Lee's skeleton and then compared against Lee—but since the reported comparison is on a held-out split and the main segmentation result is that NeuralSkel does not help ClDice, the circularity is a minor issue, not the load-bearing one. Code and data links are placeholders, so the 'publicly shared' claim is currently unverified. Anatomical labeling accuracy is shown qualitatively; no numbers.\n\nWho is this for? Researchers working on liver vessel segmentation who want a multi-class dataset and a starting point for tree-separation heuristics. The morphometry pipeline is a nice extra. It deserves a serious referee, but the dataset count and the exclusion handling need to be fixed before publication. I'd send it to review, with the expectation of a revised version.","headline":"Useful pipeline and honest negative result, but the dataset count doesn't add up and the 'clean low-error' claim only covers the cases the algorithm handled.","tokens_in":8976,"tokens_out":1912,"would_cite":false,"duration_ms":16774,"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":"An automatic pipeline using a learned centerline network and connected-component heuristics separates liver CT vessel segmentations into surgeon-validated portal and hepatic trees, and releases a 77-case annotated dataset enabling…","keywords":["liver vessel segmentation","portal vein","hepatic vein","skeletonization","ClDice","connected components","morphometry","medical imaging dataset"],"falsifier":"Take the nine excluded LiTS cases, repair the single-class segmentation so the distal regions are connected, rerun the separation algorithm, and count voxels that surgeons still judge mislabeled; if the error remains concentrated in the dense distal periphery, the angle-and-proximity heuristic is the bottleneck rather than the initial segmentation.","tokens_in":7853,"feed_emoji":"🩻","tokens_out":7595,"duration_ms":61146,"temperature":0.7,"pith_summary":"The paper tries to establish that an end-to-end pipeline—a pretrained neural skeletonizer used inside a topology-preserving loss, followed by a connected-component and skeleton-based separation algorithm—can turn a single-class 3D segmentation of liver vessels into anatomically labeled portal and hepatic venous trees with low error. The authors show that their neural skeletonizer represents vessel centerlines much better than morphological thinning, with 98.2% surface Dice versus 86.0% and 81.8% for the morphological baselines, and they build a public dataset of 77 CT cases with separated and branch-labeled trees. A sympathetic reading of the results is that the bottleneck for accurate liver vessel models is not skeleton quality within the ClDice loss, since adding NeuralSkel there does not improve segmentation, but rather having a large multi-class training set, which their separation algorithm provides. This matters because surgery planning and disease assessment in liver cancer depend on knowing which vessel perfuses and which drains each part of the liver, and such annotated data are scarce.","feed_headline":"Neural skeletonizer separates liver's portal and hepatic vein trees","feed_subtitle":"A 77-case public dataset with branch labels follows; retraining on it lifts vessel segmentation scores by 2-3 points.","key_machinery":"Two mechanisms carry the argument. NeuralSkel is a pretrained centerline-prediction network built on the nnU-Net architecture; it is differentiable, so it can replace morphological skeletonization inside the ClDice loss, and it produces centerlines that stay close to the reference skeleton, reaching 98.2% surface Dice versus 86.0% for soft-skeletonization and 81.8% for soft-persistent-skeletonization. The separation algorithm is the second mechanism: it pastes an inferred two-class labeling onto the original segmentation, dilates it, isolates uncovered regions as connected components, assigns them to the nearest tree, skeletonizes components touching both trees, splits them into branches, and resolves each branch by an angle rule of less than 60 degrees, iterating until no conflict remains.","core_discovery":"The central claim is that a learnable skeletonizer, NeuralSkel, a pretrained nnU-Net that predicts centerlines from softmax segmentation volumes, gives markedly better centerline geometry than the morphological skeletonization used by prior differentiable losses, but the paper reports that plugging this better skeleton into the ClDice loss does not improve the final vessel segmentation. Instead, the discovery that carries the contribution is algorithmic: after pasting a model's inferred portal and hepatic labeling onto a single-class annotation, disputed regions are resolved by connected-component proximity and a 60-degree branch-angle criterion derived from Murray's law, yielding a two-tree separation that surgeons rated as flawless or mildly confused in 53 of 62 LiTS cases. On the resulting LIRCAD dataset, retraining the vessel segmentation model improves all metrics by two to three percentage points and reduces Hausdorff distance. The paper further claims that these separated trees can be automatically labeled into clinically named branches, including the main portal vein, right and left branches, and Couinaud segments, and then analyzed morphometrically.","pith_inferences":["The paper's own exclusion of nine LiTS cases suggests the separation heuristics are reliable only where the initial segmentation is clean; a testable extension is to run the algorithm on a segmentation with better distal connectivity and check whether those nine cases become resolvable.","Since skeleton quality improved but downstream ClDice segmentation did not, the topology signal that matters may already be captured by the simpler losses; pairing NeuralSkel with other topology-preserving losses or using it to supervise post-processing could yet translate its centerline accuracy into segmentation gains.","The 60-degree branch-angle rule is organ-agnostic in principle, so the same separation tool could be tried on other paired vascular trees, such as hepatic artery and bile ducts, where anatomy follows similar branching, with manual validation.","The morphometric features extracted here could be correlated with tumor location and liver regeneration outcomes in future work, but that clinical link is not established by this paper."],"forward_implications":["If NeuralSkel generalizes as a skeletonizer, any tubular-structure segmentation trained with ClDice can swap in a learned, more accurate centerline module without hand-tuning morphological kernels.","The LIRCAD dataset of 77 separated, branch-labeled CT cases gives the community training data that improves vessel segmentation by two to three percentage points over training on IRCAD alone.","Automatic anatomical branch labels make morphometric measurements such as length, radius, emergence angle, generation, and descendant counts available per branch, enabling quantitative comparisons of right versus left liver territories.","The separation algorithm reduces annotation burden: surgeons needed to fix only minor mix-ups in 18 of the 53 final cases rather than labeling tree membership from scratch.","Because the separation works from a single-class input, it can be reapplied to other single-class annotations of liver vessels, not just the in-house LiTS subset."],"supporting_citations":[{"why":"U-Net based skeletonization with bag of tricks that NeuralSkel builds on for centerline prediction.","marker":"[10]"},{"why":"Introduces ClDice, the differentiable skeletonization loss that the paper modifies by substituting neural skeletonization.","marker":"[7]"},{"why":"Lee's thinning algorithm supplies the reference centerline used as ground truth for training NeuralSkel and for downstream tree analysis.","marker":"[17]"},{"why":"LiTS provides the 62 public CT cases whose single-class vessel annotations are separated and re-released as LIRCAD.","marker":"[12]"},{"why":"nnU-Net is the self-configuring segmentation framework used for both the vessel segmentation model and NeuralSkel.","marker":"[18]"},{"why":"skan converts the skeleton into branches at bifurcations, which the separation algorithm needs before applying the angle rule.","marker":"[13]"},{"why":"Murray's law on branching angles supplies the 60-degree threshold used to decide whether a skeleton branch belongs to a tree.","marker":"[14]"},{"why":"IRCAD provides the 20 CT volumes with both portal and hepatic trees that seed the multi-class training data.","marker":"[19]"},{"why":"Tian et al.'s Couinaud segmentation network provides the liver-sector prediction used to name branches anatomically.","marker":"[16]"},{"why":"Surface Dice is one of the metrics used to compare skeletonization methods and to evaluate the downstream segmentation model.","marker":"[22]"}],"fun_headline_variants":["Liver vein split uses Murray's law branch-angle rule","New dataset and algorithm separate liver venous trees","Surgeon-validated liver vein separation method introduced","77-case liver dataset boosts vessel segmentation scores","Algorithmic trick separates portal and hepatic veins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that contested voxels can be assigned to the correct venous tree from connected-component proximity and a fixed 60-degree branch-angle rule; if dense, highly interconnected distal regions defeat that premise, as the paper itself reports for 9 of 62 LiTS cases, the claimed clean multi-class segmentation is only established for the cases that survive such heuristics.","fun_headline_variants_meta":{"raw":{"variants":["Liver vein split uses Murray's law branch-angle rule","New dataset and algorithm separate liver venous trees","Surgeon-validated liver vein separation method introduced","77-case liver dataset boosts vessel segmentation scores","Algorithmic trick separates portal and hepatic veins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1750,"prompt_tokens":1049,"completion_tokens":701,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":631}},"tokens_in":665,"tokens_out":701,"duration_ms":6613,"temperature":1.0,"reasoning_tokens":631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:54:57.290310+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the nine excluded LiTS cases, repair the single-class segmentation so the distal regions are connected, rerun the separation algorithm, and count voxels that surgeons still judge mislabeled; if the error remains concentrated in the dense distal periphery, the angle-and-proximity heuristic is the bottleneck rather than the initial segmentation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces ClDice, the differentiable skeletonization loss that the paper modifies by substituting neural skeletonization."},{"cited_title":"Isensee, P.F","cited_arxiv_id":null,"evidence_quote":"nnU-Net is the self-configuring segmentation framework used for both the vessel segmentation model and NeuralSkel."},{"cited_title":"A new Python library to analyse skeleton images confirms malaria parasite remodelling of the red blood cell membrane skeleton","cited_arxiv_id":null,"evidence_quote":"skan converts the skeleton into branches at bifurcations, which the separation algorithm needs before applying the angle rule."},{"cited_title":"THE PHYSIOLOGICAL PRINCIPLE OF MINIMUM WORK AP- PLIED TO THE ANGLE OF BRANCHING OF ARTERIES","cited_arxiv_id":null,"evidence_quote":"Murray's law on branching angles supplies the 60-degree threshold used to decide whether a skeleton branch belongs to a tree."},{"cited_title":"Soler, A","cited_arxiv_id":null,"evidence_quote":"IRCAD provides the 20 CT volumes with both portal and hepatic trees that seed the multi-class training data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Tian et al.'s Couinaud segmentation network provides the liver-sector prediction used to name branches anatomically."}],"review_version":1}