{"id":"3c051195-f258-4bf0-b735-1efea71a7025","arxiv_id":"2507.22530","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new dataset and network for segmenting hepatic vasculature in high-resolution hepatectomy videos, reporting the best scores on the new benchmark.","lead":"This paper introduces a new annotated dataset of liver blood vessels in high-resolution surgical videos and a video segmentation network, HRVVS, that uses pretrained visual autoregressive features plus a memory decoder. It reports the best segmentation scores on this dataset, although the gains over the closest baseline are modest and the evaluation protocol needs scrutiny.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The split unit in Section 3.1 is unspecified; if frames rather than videos are randomly split, temporally adjacent near-duplicate frames leak into both training and test and can inflate HRVVS's Table 1 margins.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern I find: the unspecified split unit in Section 3.1. The paper's contribution is a real dataset and a coherent video segmentation architecture, and I have no objection to the method itself. However, the strongest claim is the quantitative superiority over state-of-the-art methods, and that claim rests entirely on the integrity of the train/test separation. A frame-level random split of 11,442 frames drawn from only 35 videos would place near-duplicate frames in both training and test, and the temporal memory mechanism in HRVVS would be advantaged by such leakage. The reported margins are also small in absolute terms and come without error bars or significance tests, so even modest leakage could change the ranking. The concrete test I propose directly settles the concern by checking the split file and recomputing the comparison under a video-level split. This does not move the reader's conditional verdict; it sharpens the condition: the paper should be accepted only if the split is verified to be video-level and the numbers reproduce. I therefore keep the verdict unchanged.","tokens_in":8646,"tokens_out":4009,"duration_ms":54853,"concrete_test":"Obtain the released dataset and split script and check whether train/validation/test membership is assigned over video IDs or over individual frame paths. Then retrain HRVVS and SALI under a strict video-level split using the same hyperparameters, epochs, and evaluation protocol, and recompute the five metrics in Table 1. If the Jaccard gap over SALI shrinks below roughly one point or reverses, the stated 'significantly outperforms' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the Table 1 comparison showing HRVVS ahead of SALI on Jaccard, Dice, F-measure, and E-measure. That comparison is only valid if the test set is independent of the training set in the temporal sense. Section 3.1 states only 'The data is randomly split into training, validation, and test sets with a ratio of 7:1:2.' With 35 videos and 11,442 frames, the natural unit being split is ambiguous. If the split is at frame level, roughly 70% of each video's frames go into training, and almost every test frame is temporally adjacent to a training frame. Consecutive frames in a surgical video at 1080x1920 are near-duplicates, so the test set would not measure generalization to unseen videos. HRVVS is a video model with a memory bank and cross-frame propagation, so it is precisely the kind of method that would benefit disproportionately from such leakage, and the reported gains over SALI are small in absolute terms (Jaccard 0.5405 vs 0.5239; no error bars or significance tests are reported). The architecture and dataset contribution are plausible, but the headline comparison cannot be trusted until the split unit is confirmed to be video-level and the metrics are reproduced under that split.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Hepa-SEG, described as the first high-resolution hepatic vasculature segmentation dataset for surgical videos, containing 35 hepatectomy videos and 11,442 high-resolution frames with frame-by-frame annotations. The authors also propose HRVVS, a video segmentation network with a dual-branch encoder that uses a frozen pretrained visual autoregressive (VAR) model as residual priors, a multi-view spatiotemporal interaction module (MSIM), and a dynamic weight fusion module (DWFM) with a memory bank. The method is compared against nine image- and video-level segmentation baselines on five metrics, and Table 1 reports improvements over the strongest baseline SALI on four of five metrics, with the S-measure exception attributed to LDNet.","tokens_in":8895,"tokens_out":5180,"duration_ms":60905,"significance":"If the empirical evaluation is valid, the paper makes two useful contributions: a new, publicly planned benchmark for a clinically relevant but understudied task, and a plausible architectural recipe that demonstrates transfer of a pretrained VAR model to surgical video segmentation. The dataset contribution is particularly valuable because no prior high-resolution video hepatic vasculature segmentation dataset exists. However, the headline quantitative claim rests entirely on a single random split with no error bars or significance tests, and the split unit is not specified, so the comparison cannot yet be regarded as established.","major_comments":[{"comment":"The split protocol is underspecified in a load-bearing way. The text says 'The data is randomly split into training, validation, and test sets with a ratio of 7:1:2' but does not state whether the unit of splitting is the video or the frame. With 35 videos and 11,442 frames, a frame-level random split would place temporally adjacent, near-duplicate frames of the same video into both training and test sets. Because HRVVS is explicitly a video model with a memory bank and multi-frame propagation, such leakage would disproportionately inflate its results in Table 1. Please specify the split unit; if it was frame-level, re-run all experiments with a video-level split (e.g., report which video IDs are in each set) and report the resulting metrics. This is essential for the central claim that HRVVS outperforms state-of-the-art methods.","section":"Sec. 3.1"},{"comment":"No standard deviations, confidence intervals, or significance tests are reported anywhere, yet the abstract states that HRVVS 'significantly outperforms' the state-of-the-art methods. The margins over the strongest baseline SALI are small in absolute terms (Jaccard 0.5405 vs. 0.5239; Dice 0.6532 vs. 0.6424), and the ablation table suffers from the same absence of repeated runs. Please report mean and standard deviation over multiple seeds or, preferably, per-video metrics with a paired significance test over the 35 videos. Without this, the 'significantly outperforms' claim is not supported.","section":"Table 1 / Sec. 3.2"},{"comment":"The baseline identity of 'ISNet' is inconsistent and must be corrected for reproducibility. The text in Sec. 3.2 refers to 'ISNet [33]' as a high-resolution segmentation method, but reference [33] is 'Isnet: Shape matters for infrared small target detection' (CVPR 2022), whereas Table 1 lists 'ISNet [19] ECCV 22', with reference [19] being 'Highly accurate dichotomous image segmentation' (ECCV 2022). These are different models. Please align the citation and specify exactly which ISNet variant was used in the comparison, and rerun the baseline with the intended model if the mismatch changes the reported numbers.","section":"Sec. 3.2 and References"}],"minor_comments":[{"comment":"Equations (4) and (5) introduce fusion weights α, β, γ, and δ, but the paper does not state whether these are learned or fixed hyperparameters. Please specify and, if fixed, report their values.","section":"Sec. 2.4"},{"comment":"The memory bank uses 'exponential downsampling rates' for historical frames, but the rates are never defined. Please give the precise schedule or formula.","section":"Sec. 2.3"},{"comment":"The ablation 'basic' model is not described. Please define the architecture that excludes the VAR branch, MSIM, and DWFM so that the reader can interpret the incremental contributions.","section":"Table 2"},{"comment":"As this is a new dataset, the paper should report annotation details: the number of annotators, their clinical expertise, the annotation tool, and inter-annotator agreement. Without these, the 'high quality frame-by-frame annotated' claim is difficult to assess.","section":"Sec. 3.1"},{"comment":"The abstract says 'Extensive experiments on surgical video datasets' (plural), but experiments are conducted only on Hepa-SEG. Please change to 'surgical video dataset' or add additional datasets.","section":"Abstract"},{"comment":"The notation {L^{mism}_m} appears to be a typo for {L^{msim}_m}; the module abbreviation is also written as both 'Multi-view Spatiotemporal Interaction Module' and 'Multi-scale Integration Module' in the ablation discussion. Please unify the terminology.","section":"Fig. 2 and Sec. 2.3"}],"recommendation":"major_revision","confidential_remarks":"The core architecture and dataset proposal are plausible, and I see no indication of intentional data leakage; the central problem is that the current evaluation protocol is ambiguous and underpowered. A re-run with video-level splitting and repeated trials would likely resolve the main concern. The ISNet citation discrepancy should also be checked to confirm that the intended baseline was actually used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main reason to look at this paper is the dataset. Hepa-SEG, 35 hepatectomy videos with frame-level manual annotations at 1080x1920, is a genuinely new resource and the first public benchmark for hepatic vasculature segmentation in surgical video. That alone carries weight. The method is a reasonable engineering effort: pretrained VAR features as residual priors into a Swin encoder, plus a memory decoder with attention-based fusion. The ablation shows the full model does best, and DWFM appears to be the main driver. Nothing revolutionary, but it is not a toy either.\n\nThe soft spot is the one the reader flagged: the split. Section 3.1 says 'The data is randomly split into training, validation, and test sets with a ratio of 7:1:2.' It does not say videos are split. With 35 videos and 11,442 frames, a frame-level split would place near-duplicate consecutive frames into both training and test, directly inflating the video-level metrics. HRVVS is a memory-based video model, so it would benefit disproportionately from that leakage. The reported gains over SALI are also small in absolute terms (Jaccard 0.5405 vs 0.5239, Dice 0.6532 vs 0.6424), and no standard deviations or significance tests appear anywhere. Until the authors confirm the split is video-level and reproduce the numbers under that condition, the headline claim should not be taken at face value. I do not think this is fraud—likely sloppy reporting—but it is load-bearing.\n\nTwo smaller issues. First, the baseline training protocol is underdescribed: were all baselines trained with the same sliding window, input resolution, and loss? That matters for a fair comparison. Second, there is a citation inconsistency for ISNet: the text cites reference [33] (a different ISNet paper on infrared targets) while Table 1 lists ISNet as [19] (the dichotomous segmentation paper). Minor, but symptomatic of rushed finalization.\n\nWho is this for? Anyone working on surgical video segmentation or medical video datasets. The dataset, once released, will be a useful benchmark even if the method is not. The paper deserves a serious referee, but the review must demand clarification of the split unit, re-evaluation if needed, error bars, and the actual code and data. Right now it is a conditional accept at best, not a reject outright.","headline":"The dataset is a real contribution; the method is plausible but the headline numbers are not trustworthy until the split unit is clarified.","tokens_in":9441,"tokens_out":2240,"would_cite":false,"duration_ms":27695,"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 introduces the first high-resolution hepatic vasculature video segmentation dataset, Hepa-SEG, and a network, HRVVS, that reports state-of-the-art performance on it.","keywords":["hepatic vasculature segmentation","surgical video","high-resolution segmentation","visual autoregressive model","video object segmentation","Hepa-SEG dataset","memory decoder","hepatectomy"],"falsifier":"Re-run Table 1 with a strict video-level split, holding out entire videos rather than individual frames; if HRVVS's Jaccard lead over SALI shrinks from +3.16% to within noise, the current split was leaking temporal correlation and inflating the reported gains.","tokens_in":8456,"feed_emoji":"🩺","tokens_out":7195,"duration_ms":83469,"temperature":0.7,"pith_summary":"This paper introduces the first dataset and segmentation network for hepatic vasculature in high-resolution surgical video. The Hepa-SEG dataset contains 35 hepatectomy videos and 11,442 manually annotated 1080x1920 frames, and the proposed HRVVS network combines a frozen visual autoregressive model as a residual prior with a dynamic memory decoder. On that dataset, HRVVS is reported to outperform nine state-of-the-art image and video segmentation methods, reaching a Jaccard index of 0.5405, Dice of 0.6532, S-alpha of 0.7878, F-measure of 0.6769, and E-measure of 0.8711. If the benchmark is sound, this gives surgeons and researchers the first measurable target for live vessel localization during liver resection.","feed_headline":"Liver-vessel video segmentation sets new benchmark: Jaccard 0.54","feed_subtitle":"New dataset of 35 hepatectomy videos and a memory decoder push past SALI on four metrics.","key_machinery":"The core mechanism is the hierarchical autoregressive residual prior: a frozen VAR model's multi-scale unified-quantization features are projected by adapters and added into the downsampling layers of a Swin-based multi-view encoder, so generation-style priors compensate for information lost as resolution decreases. The decoder side hinges on a dynamic memory mechanism: MSIM updates current local and global features via multi-head cross-attention with exponentially downsampled historical frames from a memory bank, and DWFM fuses four local views patch-wise using weights that combine current global, previous-frame, and historical weights, with historical weights updated by an exponential moving average. These two mechanisms carry the argument that both within-frame detail and cross-frame continuity are needed for accurate vessel segmentation.","core_discovery":"The central claim is that hepatic vasculature segmentation in high-resolution hepatectomy videos can be solved by a dual-branch encoder in which a pretrained visual autoregressive model (VAR) injects multi-scale residual features into a Swin-based multi-view encoder, combined with a dynamic memory decoder that propagates selected global features across frames. The paper asserts this is the first work on this task and that HRVVS significantly outperforms state-of-the-art image and video segmentation methods on the introduced Hepa-SEG dataset, beating the strongest baseline SALI on Jaccard, Dice, F-measure, and E-measure, with only S-alpha lower than LDNet. The reported gains are attributed to the VAR residual priors reducing information degradation during downsampling and to the Multi-view Spatiotemporal Interaction Module (MSIM) and Dynamic Weights Fusion Module (DWFM) reducing redundant cross-frame transmission while preserving fine vessel detail.","pith_inferences":["If the reported gains survive a video-level split, the same frozen-autoregressive-prior-plus-adapter recipe could be tested on other high-resolution surgical video segmentation tasks where annotated frames are scarce, such as ureter or nerve delineation.","The dynamic memory weighting in DWFM is domain-agnostic; ablating it on an established public video object segmentation benchmark would show whether temporal weight smoothing rather than the surgical domain drives the improvement over video baselines.","The paper does not report per-frame inference time or memory, so measuring those on a clinical GPU would settle whether the gains translate into real-time intraoperative guidance rather than only offline accuracy."],"forward_implications":["If the reported numbers hold under proper splitting, HRVVS provides the first state-of-the-art baseline on Hepa-SEG, giving future work a concrete target for hepatic vasculature segmentation in surgical video.","The VAR residual prior mechanism shows that a frozen generative image model can supply multi-scale context to a segmentation encoder, reducing information loss during downsampling without full finetuning.","The dynamic memory decoder demonstrates a way to propagate temporal information across long surgical videos while limiting redundant transmission, which is directly relevant to other long video segmentation tasks.","The public release of Hepa-SEG means the dataset itself becomes a benchmark, allowing other methods to be compared under identical conditions on a previously missing task."],"supporting_citations":[{"why":"Supplies the pretrained visual autoregressive model whose multi-scale features are injected as residual priors into the encoder.","marker":"[23]"},{"why":"Provides the Swin Transformer hierarchical backbone that the multi-view branch builds on.","marker":"[16]"},{"why":"SALI is the strongest video baseline and the main comparison that HRVVS is reported to beat on four metrics.","marker":"[10]"},{"why":"Vivim is a recent medical video segmentation baseline used in the comparison table.","marker":"[31]"},{"why":"Introduces the SUN-SEG video polyp segmentation dataset and algorithm that frame the video segmentation setting for medical images.","marker":"[12]"},{"why":"PraNet is one of the image-level polyp segmentation baselines in the comparison.","marker":"[7]"},{"why":"Med-SAM2 is a SAM2-based medical video segmentation baseline compared in Table 1.","marker":"[35]"},{"why":"Supplies the five evaluation metrics used to score all methods on Hepa-SEG.","marker":"[17]"}],"fun_headline_variants":["Autoregressive priors sharpen liver-vessel video segmentation","New dataset and memory decoder for liver vessel segmentation","VAR priors reduce information loss in surgical video segmentation","HRVVS: high-res liver vessel segmentation with autoregressive priors","Memory decoder and VAR residuals advance hepatectomy video analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark is valid only if the random 7:1:2 split described in Section 3.1 is applied at the video level or otherwise breaks temporal correlation, so that test frames are not near-duplicates of training frames.","fun_headline_variants_meta":{"raw":{"variants":["Autoregressive priors sharpen liver-vessel video segmentation","New dataset and memory decoder for liver vessel segmentation","VAR priors reduce information loss in surgical video segmentation","HRVVS: high-res liver vessel segmentation with autoregressive priors","Memory decoder and VAR residuals advance hepatectomy video analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001447,"raw_usage":{"total_tokens":5823,"prompt_tokens":937,"completion_tokens":4886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":4804}},"tokens_in":553,"tokens_out":4886,"duration_ms":38999,"temperature":1.0,"reasoning_tokens":4804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:33:41.718294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 1 with a strict video-level split, holding out entire videos rather than individual frames; if HRVVS's Jaccard lead over SALI shrinks from +3.16% to within noise, the current split was leaking temporal correlation and inflating the reported gains.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the five evaluation metrics used to score all methods on Hepa-SEG."}],"review_version":1}