{"id":"eb724de7-658c-4957-84f9-7bce70af4002","arxiv_id":"2508.01785","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A point-based Couinaud segmentation model that replaces explicit liver vessel priors with a 3D dynamic graph reasoning module, achieving competitive Dice and ASD on MSD and LiTS.","lead":"This paper presents a point-based deep learning method for Couinaud liver segmentation that does not need manually drawn liver vessel priors, learning anatomical structure implicitly via a graph reasoning module. It reports competitive Dice and surface distance scores on two public CT datasets compared to four existing point-based methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported margins over baselines rest on a single, undisclosed split with no error bars or significance tests; on LiTS the gap to the vessel-prior baseline is below one Dice point, so the claimed superiority is not yet supported.","rationale":"The reader's weakest_assumption names both annotation quality and split representativeness. I agree that the split is load-bearing, but I would foreground statistical robustness: even with perfect annotations, a single unreproducible split with no significance testing cannot support the claimed margin, especially on LiTS where the gap is under one Dice point. Annotation validity is a shared precondition for all methods and, if anything, a bias toward vessel-prior annotations would disadvantage the proposed method, making its result conservative. In contrast, the absence of uncertainty quantification directly threatens the comparison that the central claim rests on. I therefore partially agree with the reader and keep the CONDITIONAL verdict, adding a concrete reproducibility check that would settle whether the observed margins are meaningful. The mechanism claim (that the graph module specifically encodes anatomical structure) is also worth probing, but the statistical robustness of the empirical comparison is the more immediate gate: if the numbers do not survive re-splitting, the performance claim fails regardless of mechanism.","tokens_in":8761,"tokens_out":8106,"duration_ms":105792,"concrete_test":"Run the full comparison under five random 10/3/7 splits (or five-fold cross-validation) on both MSD and LiTS, computing per-case Dice for every method and reporting means with 95% confidence intervals, plus a paired significance test (e.g., Wilcoxon signed-rank) between the proposed method and each baseline. If the LiTS Ours-vs-Zhang difference is not significant (p>0.05) or the confidence intervals overlap, the head-to-head claim should be softened to 'comparable' rather than 'exceed'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the proposed method is competitive with, and in the conclusion is said to exceed, four point-based baselines. However, the evidence for this claim in Tables 1 and 2 is a single point estimate per method: no error bars, no per-case variance, and no significance tests are reported. The paper specifies only a 10/3/7 training/validation/test split ratio (Section 4, Implementation details) without disclosing the actual subject assignment or random seed, so the results cannot be reproduced and could depend on an unrepresentative split. The margins on LiTS are especially thin: 79.56 vs. 78.62 Dice (0.94 points) against the vessel-prior baseline [28]. Given that per-subject Dice variability in liver segmentation is typically substantial, a difference below one point could easily arise from sampling noise. The standard deviations reported in the ablation study (Table 3), e.g., 80.41 ± 10.74, are across the eight Couinaud segments, not across test subjects, so they do not quantify the uncertainty of the headline comparison. If the split is re-randomized and the LiTS margin disappears, the conclusion 'we exceed' in the Discussion would be unsupported, and the abstract's more modest 'competitive' claim would still require validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a point-based network for Couinaud liver segmentation from CT that avoids explicit liver-vessel priors. The method builds on AGCNN and adds two components: grid feature embeddings that bridge point features into a voxel grid, and a 3D dynamic graph reasoning module that learns affinities among a small set of neighboring voxels. The authors evaluate on the MSD and LiTS datasets using Couinaud annotations inherited from prior works, compare against four point-based baselines, report Dice and average surface distance, and provide an ablation study. The central claim is that the method is competitive with, and in the conclusion is said to exceed, the four baselines without requiring explicit vessel structure priors.","tokens_in":9074,"tokens_out":6004,"duration_ms":73060,"significance":"If the empirical claims are supported, the contribution is useful: removing the need for manually defined vessel priors while retaining anatomical context through learned graph affinities is a sensible direction for point-based Couinaud segmentation. Strengths include the use of two public datasets, comparison with four relevant baselines, a clear ablation design, qualitative failure-case analysis, and released code. The ablation demonstrates that the grid-feature bridge is necessary for the graph module to work. However, the current evidence is not sufficient for the stated 'exceed' conclusion: the headline comparisons rest on a single undisclosed split without error bars or significance tests, and the annotations are inherited from prior work with no independent validation, including the same first author's earlier method that is also the strongest baseline.","major_comments":[{"comment":"The main comparison is a single point estimate per method on one train/validation/test split. The paper only reports the split ratio (10/3/7), not the exact subject assignment or random seed, and Tables 1 and 2 contain no error bars, confidence intervals, or significance tests. On LiTS, the margin over the vessel-prior baseline [28] is 0.94 Dice points; without per-subject variance and a paired statistical test, this margin is not evidence of superiority. Please report per-case mean and standard deviation over multiple random splits, or at least bootstrapping over test subjects, and disclose the exact split and seed. The standard deviations in Table 3 are not accompanied by a statement of what they are computed over and do not by themselves validate the headline comparison.","section":"Section 4, Implementation details; Tables 1–2"},{"comment":"The Couinaud ground-truth labels are inherited from Tian et al. [21] and Zhang et al. [28] rather than generated and validated for this study. No annotation protocol, inter-observer agreement, or independent anatomical verification is reported. Because [28] is also the strongest baseline and is a prior work by the same first author, there is a risk of circularity if those annotations were produced by a method trained on subjects that overlap the evaluation set, or if the labels encode a vessel model that [28] matches by construction. Please describe the provenance of the annotations, report inter-rater reliability on a subset, and verify that no subject overlap exists between the annotation-generation procedure and the training/validation/test split used here.","section":"Section 4, Dataset description"},{"comment":"The conclusion states that the method 'exceed[s] PointNet, PointNet++, AGCNN and Zhang et al.'s method', whereas the abstract claims only that the method is 'competitive'. The reported evidence supports at most the weaker claim. If the statistical and annotation-validation concerns above are not addressed, the conclusion should be softened to reflect the actual strength of the evidence.","section":"Section 5, Discussion and conclusion"}],"minor_comments":[{"comment":"The notation '33(= 27) voxels' is ambiguous and should read '3^3 (= 27) voxels'.","section":"Section 3.2"},{"comment":"There are typos: 'precictions' should be 'predictions', and 'heaptic' should be 'hepatic'.","section":"Section 4.2 and Figure 4 caption"},{"comment":"The exact source, license, and availability of the Couinaud annotations from [21] and [28] should be stated, since the reader currently cannot obtain the same labels to reproduce the experiments.","section":"Section 4, Dataset description"},{"comment":"The inference-time and GFLOPs numbers would be more useful if the software versions of PyTorch, Torchmetrics, and MONAI were specified, as these can affect runtime measurements.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an incremental but relevant contribution. The main weakness is empirical: the headline claim of exceeding four baselines is not supported by the reported statistics, and the inherited annotations create a self-reference risk that should be explicitly addressed. I would encourage the editor to request a revised version with multi-split evaluation, significance testing, and annotation validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Drop the vessel priors and add a 3D dynamic graph reasoning module: that's the whole paper, and it's a reasonable simplification. The new combination is real—nobody else in the point-based Couinaud line has removed explicit priors and learned affinities implicitly in this way. The authors also do the right things: they compare against four point-based baselines, release code, and include an honest failure case where tumors hide anatomical landmarks.\n\nThe soft spot is the evidence for the central empirical claim. Tables 1 and 2 report single Dice/ASD numbers per method on a single 10/3/7 split with no disclosed subject assignment, no error bars, no significance tests. On LiTS the margin over the vessel-prior baseline is 0.94 Dice points; that could easily be sampling noise. The standard deviations in the ablation are computed across the eight segments, not across test subjects, so they don't speak to the comparison. On MSD the margin is larger, about 4 Dice points, which is more encouraging, but still unquantified across runs or splits.\n\nA second concern is the annotation provenance. The ground truth comes from the same first author's prior paper [28] (and Tian et al. [21]), and there's no independent validation or inter-rater agreement. That doesn't make the method wrong, but it means the benchmark could be biased in a subtle way, especially when the comparison includes [28] itself.\n\nThe method itself is coherent. The ablations show both the grid embeddings and the graph reasoning contribute, and the reasoning module is doing something meaningful in the qualitative figures. The 'we exceed' statement in the discussion is stronger than the evidence in the tables; the abstract's 'competitive' is the right word.\n\nThis is a modest, useful increment for the liver segmentation crowd. It deserves a serious referee, but only if the authors are asked to provide multiple random splits or bootstrapped confidence intervals, and to clarify how the annotations were produced. If the LiTS margin disappears under re-splitting, the conclusion should be softened.","headline":"Useful architectural simplification, but the single unvalidated split and thin LiTS margin mean the 'exceed' claim is not yet supported.","tokens_in":9549,"tokens_out":2823,"would_cite":false,"duration_ms":32680,"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":"This paper claims that point-based Couinaud segmentation can match or beat vessel-prior methods without any explicit vessel prior, by learning anatomical structure through a 3D dynamic graph reasoning module.","keywords":["Couinaud segmentation","3D graph reasoning","point-based segmentation","liver CT","implicit anatomical information","dynamic graph message passing","vessel-free liver segmentation"],"falsifier":"Re-run the full method and its ablation on the same public data with the same inherited annotations and the same 10/3/7 split: if the 80.41 MSD and 79.56 LiTS average Dice are not reproduced, or if removing the graph reasoning module fails to lower Dice by roughly two points, the claim that implicit graph-based anatomy drives the gain is not supported.","tokens_in":8594,"feed_emoji":"🩺","tokens_out":11921,"duration_ms":119329,"temperature":0.7,"pith_summary":"This paper sets out to show that point-based Couinaud segmentation of liver CT can be done accurately without the manual liver vessel prior that earlier point-based methods required. Couinaud segmentation divides the liver into eight functionally independent segments whose borders come from hepatic and portal vein planes, and it is used in planning liver resection and ablation. The proposed model learns that vessel structure implicitly: at each of four sampling levels it revoxelizes the point features and applies a dynamic graph reasoning module that computes affinities among neighboring voxels, so no explicit vessel labels are needed at inference. On the MSD and LiTS benchmarks it reports the highest average Dice among four compared point-based baselines, 80.41% versus 76.00% on MSD and 79.56% versus 78.62% on LiTS, with competitive average surface distance. The practical payoff is that surgical planning could skip the time-consuming vessel-annotation step without giving up segment accuracy.","feed_headline":"No vessel priors needed: graph point model tops Couinaud segment Dice","feed_subtitle":"A 3D graph module learns liver anatomy implicitly, beating four point-based baselines on MSD and LiTS.","key_machinery":"The load-bearing mechanism is the dynamic graph reasoning module $G_r(f(p_i))$ applied at each of four down-sampled point levels. The model first revoxelizes the point features into a $[M_i\\times M_i\\times M_i]$ grid and processes them with two 3D residual convolutions, producing a grid feature embedding $f(p_i)$ that bridges point space and grid space. In that grid, the module computes affinities over a fixed subset of $3^3=27$ voxels selected by learnable 3D offsets, with the update $$\\hat{p}_i = \\mathrm{softmax}\\big(q_i\\cdot \\$\\Delta$(k_i,\\delta_i)+\\mathrm{pos}_i\\big)\\cdot \\$\\Delta$(v_i,\\delta_i),$$ where $q_i,k_i,v_i$ are self-attention projections and $\\Delta(\\cdot,\\delta_i)$ is a deformable unfold that adapts keys and values to the offsets. The output is devoxelized back to point coordinates by interpolation. This mechanism carries the claim because it replaces the externally supplied vessel prior with affinities that the network itself learns from textural and positional patterns around the vessel planes.","core_discovery":"The central claim is that the anatomical information needed for Couinaud segmentation can be learned from the CT intensities themselves instead of being provided as a prior. The authors add a 3D dynamic graph reasoning module on top of voxelized point features; the module selects a small set of voxels with learnable 3D offsets and computes self-attention affinities among them, which together encode implicit anatomical relations between liver regions. With only eight segment labels as supervision, this module yields an average Dice of 80.41 on MSD and 79.56 on LiTS, higher than the vessel-prior point-based baseline at 76.00 and 78.62, and the best average surface distance on LiTS among the four baselines. The ablation study shows both the grid feature embeddings and the graph reasoning module are needed: removing the graph module drops average Dice by about two points on each dataset, and removing the grid embeddings drops it further. The paper's intended message is therefore that explicit vessel priors are not necessary for competitive point-based Couinaud segmentation, because a learned graph can supply the anatomy.","pith_inferences":["A natural next test is whether the learned 3D offsets align across patients with the actual hepatic and portal vein planes; if they do, the module could double as a weakly supervised vessel-anatomy detector, which the paper does not claim.","The same skip-prior recipe may transfer to other anatomy-defined partition tasks where boundaries are invisible planes, such as bronchopulmonary segments or brain vascular territories, though the paper only studies liver Couinaud segments.","Because the paper identifies large tumors as the main failure mode, an extension would be to condition the graph reasoning on tumor regions or suppress them before affinity computation; this is an inference, not part of the paper's experiments."],"forward_implications":["Liver surgery planning can skip the manual vessel-prior step: the proposed model matches or exceeds a vessel-prior point-based baseline while trained only on segment labels.","The method keeps native CT resolution: points are placed at each voxel's physical coordinates, so volumes with axial spacing up to 7.5 mm (MSD) and down to 1.0 mm (LiTS) are handled without resampling to a fixed grid.","The graph module is doing real work: ablating $G_r(f(p))$ lowers average Dice by roughly two points on both datasets, so the reported gain depends on the learned affinities, not on the backbone alone.","Segment boundaries stay close to the expected vessel planes in the qualitative results, but the method is not robust to large tumors that erase landmarks such as the left hepatic vein or left portal vein, as the paper's failure analysis shows."],"supporting_citations":[{"why":"Supplies the four-level point-processing backbone and offline ball-query neighbor sets that the proposed architecture extends; also one of the four baselines.","marker":"[25]"},{"why":"Provides the dynamic graph reasoning with adaptive sampling and self-attention that the paper extends from 2D to 3D; this is the core of the implicit anatomy module.","marker":"[27]"},{"why":"Supplies the voxelization and coordinate-based interpolation used to build grid feature embeddings and to map graph outputs back to point space.","marker":"[11]"},{"why":"Defines the vessel-prior point-based method that the paper must beat and supplies the Couinaud annotations used for the LiTS dataset.","marker":"[28]"},{"why":"Supplies the Couinaud annotations used for the MSD dataset; the reported evaluation numbers rest on these labels.","marker":"[21]"},{"why":"Provides the deformable unfold layer that adapts keys and values under the learnable 3D offsets in the affinity update.","marker":"[30]"},{"why":"One of the four point-based baselines compared in the evaluation; its per-segment results set part of the performance bar.","marker":"[17]"},{"why":"One of the four point-based baselines compared in the evaluation; its per-segment results set part of the performance bar.","marker":"[18]"}],"fun_headline_variants":["Graph module learns liver anatomy, beats vessel-prior baselines","No vessel priors: graph-based point model tops Couinaud Dice","Implicit anatomy via graph reasoning improves Couinaud measures","Point-based Couinaud without vessel priors using learned graphs","Graph-reasoned points outperform vessel priors in liver segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing premise is that the Couinaud ground-truth labels inherited from prior work, together with the random 10/3/7 train/validation/test split, are accurate and representative enough that the reported Dice margins reflect the method rather than annotation bias or an unlucky split.","fun_headline_variants_meta":{"raw":{"variants":["Graph module learns liver anatomy, beats vessel-prior baselines","No vessel priors: graph-based point model tops Couinaud Dice","Implicit anatomy via graph reasoning improves Couinaud measures","Point-based Couinaud without vessel priors using learned graphs","Graph-reasoned points outperform vessel priors in liver segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1385,"prompt_tokens":942,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":356}},"tokens_in":558,"tokens_out":443,"duration_ms":4255,"temperature":1.0,"reasoning_tokens":356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:22:18.441769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full method and its ablation on the same public data with the same inherited annotations and the same 10/3/7 split: if the 80.41 MSD and 79.56 LiTS average Dice are not reproduced, or if removing the graph reasoning module fails to lower Dice by roughly two points, the claim that implicit graph-based anatomy drives the gain is not supported.","supporting_citations":[{"cited_title":"IEEE Transac- tions on Pattern Analysis and Machine Intelligence45(8), 9374–9392 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the four-level point-processing backbone and offline ball-query neighbor sets that the proposed architecture extends; also one of the four baselines."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic graph reasoning with adaptive sampling and self-attention that the paper extends from 2D to 3D; this is the core of the implicit anatomy module."},{"cited_title":"Advances in neural information processing systems32 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the voxelization and coordinate-based interpolation used to build grid feature embeddings and to map graph outputs back to point space."},{"cited_title":"Computers in Biology and Medicine182, 109202 (2024) Title Suppressed Due to Excessive Length 11","cited_arxiv_id":null,"evidence_quote":"Defines the vessel-prior point-based method that the paper must beat and supplies the Couinaud annotations used for the LiTS dataset."},{"cited_title":"In: International workshop on machine learning in medical imaging","cited_arxiv_id":null,"evidence_quote":"Supplies the Couinaud annotations used for the MSD dataset; the reported evaluation numbers rest on these labels."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Provides the deformable unfold layer that adapts keys and values under the learnable 3D offsets in the affinity update."},{"cited_title":"Advances in neural information processing systems 30 (2017)","cited_arxiv_id":null,"evidence_quote":"One of the four point-based baselines compared in the evaluation; its per-segment results set part of the performance bar."}],"review_version":1}