{"id":"f1ee6cdd-ab23-4899-8cf0-6d33fb680c44","arxiv_id":"2411.15251","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OVS-Net, a SAM-based vessel segmentation framework with a micro-vessel enhancement branch and morphology-correction post-processing, reports higher Dice and better connectivity than six SAM baselines and 17 expert models across 17 datasets.","lead":"This paper presents OVS-Net, a single deep-learning model that segments blood vessels across X-ray, fundus, and OCTA images from 17 datasets, combining a frozen SAM encoder with a CNN branch for small vessels and a U-Net post-processor that repairs breaks in vessel connectivity. A smart generalist might read it because a single generalizable vessel segmenter with clinically usable connectivity could replace many modality-specific tools in medical imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'surpass 17 expert models' claim depends on literature-imported scores that were not re-run under OVS-Net's protocol; if these scores are not comparable, the headline claim is unsupported.","rationale":"The reader identifies comparability of literature-imported expert scores as the weakest assumption, and I agree. This is the most load-bearing risk because the paper's abstract and Section IV.B/C present the expert benchmark as core evidence for 'superior segmentation accuracy' and 'generalization.' The SAM-based comparison is internally consistent and supports a weaker claim, but the expert-model comparison is not re-run and protocol differences are non-trivial. For instance, Table VI shows OVS-Net's lead over the best literature baseline on HRF is only 0.40 Dice, which is within plausible cross-paper variation for dataset splits and preprocessing. Additionally, the paper contains secondary numerical inconsistencies (ablation text reports a 1.31 Dice increase vs. 1.03 in Table VIII; external Dice improvement reported as 1.79 vs. 2.79 from Table II), which should be corrected but do not by themselves invalidate the architecture. The suggested concrete test—re-running expert baselines under OVS-Net's protocol—is feasible for common models like U-Net and CS-Net and would settle the concern. If the retest confirms OVS-Net's superiority, the paper's conditional acceptance is warranted; if not, the claim must be downgraded. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":19290,"tokens_out":8431,"duration_ms":75432,"concrete_test":"Independently re-run a representative subset of the expert models (at minimum U-Net, U-Net++, CS-Net, H2Former, and one coronary specialist such as SPIRONet) under the exact OVS-Net data splits, preprocessing (1024×1024 resize, intensity normalization), and evaluation protocol, then compare Dice/IoU on the same test sets. If OVS-Net still achieves the highest scores on all or most datasets, the comparability concern is resolved; if expert models exceed OVS-Net on the retested datasets, the headline claim should be revised to exclude unsupported 'surpasses 17 expert models' wording.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion of superior accuracy over expert methods (abstract, Section IV.C) is supported only by scores imported from references [55]–[59] and [46], with no re-running of those models under the authors' training or evaluation protocol. Table I shows OVS-Net uses train/val/test splits and 1024×1024 resizing, while the cited expert scores come from heterogeneous sources with different splits, preprocessing, and metric conventions. For example, STARE and DRIVE are commonly evaluated with standard splits or leave-one-out, whereas the authors use their own random splits (Table I), and expert models were trained on different fractions of the data. If these protocol differences move scores by more than the observed margins (e.g., OVS-Net leads HRF by 0.40 Dice over SuperVessel in Table VI), the claim of 'surpassing 17 expert models' is not established. The SAM-based comparison in Table II is more controlled, but the expert-model comparison is a central part of the paper's headline, so this comparability assumption is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OVS-Net, a SAM-based vessel segmentation framework with a frozen ViT encoder plus feature/spatial adapters for macro-vessel extraction, a ConvNeXt-based micro-vessel enhancement branch, mask-decoder feature fusion, and a U-Net post-processing network trained with MSE and clDice losses to repair disconnected vessels. The authors compile 17 multi-modality datasets (coronary X-ray, fundus, OCTA, pelvic iliac X-ray), train on 15, and evaluate on two held-out fundus datasets, comparing against six SAM-based methods retrained under the same protocol and against 17 expert models whose scores are imported from prior publications. They report the highest average Dice, IoU, and clDice with the lowest β0 among SAM-based methods, a 34.6% connectivity improvement, and claim superior accuracy and generalization over expert models. Ablations show incremental gains from the adapter, micro-vessel enhancement, feature connection, and post-processing modules.","tokens_in":19479,"tokens_out":4874,"duration_ms":47737,"significance":"If the controlled SAM-based comparison is taken at face value, the paper makes a useful contribution: a single structure-agnostic model that improves both overlap metrics and connectivity on multi-modality vessel segmentation, supported by a large compiled benchmark and ablations that attribute gains to specific modules. The planned release of code and data is valuable for reproducibility. However, the headline claim of surpassing 17 expert models rests on literature-imported scores with heterogeneous protocols and is not established. The paper's core technical contribution and its evaluation against retrained SAM baselines are sound, but the expert-model comparison needs to be either re-run under a common protocol or explicitly reframed as context rather than head-to-head superiority.","major_comments":[{"comment":"The central claim that OVS-Net surpasses 17 expert models is supported only by scores imported from references [55]–[59] and [46] without re-running those models under the same training and evaluation protocol. Table I shows that OVS-Net uses its own random splits and 1024×1024 resizing, whereas the cited expert scores come from heterogeneous pipelines with different splits, preprocessing, and metric conventions; the paper also states that only [57] reports standard deviations and that the best published Dice was selected for comparison. Because several reported margins are small (e.g., DRIVE Dice 81.93 vs 81.41, IOSTAR Dice 78.82 vs 77.16, HRF Dice 77.14 vs 76.74), protocol differences of this size could change the ranking. The \"surpassing 17 expert models\" claim is therefore not established. I recommend either re-running expert baselines under the OVS-Net protocol or restricting the superiority claim to the controlled SAM-based comparison and presenting the expert-model numbers as contextual literature benchmarks rather than as head-to-head evidence.","section":"Section IV.C, Tables III–VI"},{"comment":"The morphological correction post-processing network is a key module, and the ablation in Table VIII attributes the β0 reduction from 0.48 to 0.34 to it. However, the text does not specify how the \"fragmented\" training masks are generated: what corruption is applied, to what fraction of masks, and how the target mask is defined. The description \"we mask the fragmented areas of the vascular masks\" leaves the data-generation procedure underspecified. Without a precise algorithm or pseudocode, the post-processing network cannot be reproduced, and the reported connectivity gain cannot be independently verified. Please provide the full procedure, including any parameters for simulating disconnections.","section":"Section III.E"},{"comment":"The abstract and conclusion claim \"robust generalization across diverse imaging modalities,\" but the external validation (Tables II and VI) uses only two fundus datasets, HRF and IOSTAR, which are the same modality as several training datasets. This tests cross-dataset generalization within a seen modality, not generalization to an unseen modality. The internal test sets do cover four modalities, but the external evaluation does not support the cross-modality generalization claim as stated. Please either add an external dataset from a modality not present in training or temper the wording to \"generalization to unseen datasets of a seen modality.\"","section":"Section IV.B/C and Conclusion"}],"minor_comments":[{"comment":"The formula for β0 is typeset incorrectly: \"β0 = 1/N NX⟩=1 |CC (Xi) − CC (Yi)|\" should use a summation symbol over i, and the text should clarify that CC(Xi) and CC(Yi) denote the numbers of connected components in patch i.","section":"Section IV.A, β0 definition"},{"comment":"For OVS-Net on the Pelvic Iliac artery dataset, β0 is reported as 0.05±0.61; a standard deviation larger than the mean suggests a typo or a unit inconsistency, and all other β0 standard deviations are much smaller relative to their means.","section":"Table II, Pelvic Iliac row"},{"comment":"The text states OVS-Net achieves an average Dice of 79.19 on IOSTAR, but Table VI lists 78.82; one of these is incorrect and should be corrected.","section":"Section IV.C, Table VI"},{"comment":"The text says \"On all three fundus datasets, our method consistently achieves the highest average Dice and IoU scores,\" but Table IV shows CS-Net has a higher IoU on DRIVE (70.17 vs 69.43); this internal contradiction should be fixed.","section":"Section IV.C, Table IV"},{"comment":"Reference [56] duplicates reference [20] (both cite the Full-Resolution Network paper by Liu et al.); please remove the duplicate and renumber.","section":"References"},{"comment":"The caption says \"we use the standard deviation as the error bar standard,\" but the expert-model scores in Tables III–VI mostly have no standard deviations, so it is unclear what error bars are shown; clarify the source of the error bars or remove the claim.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The strongest part of the paper is the controlled comparison against retrained SAM-based methods and the ablation study; the weakest is the expert-model comparison, which is imported from the literature and selected for the best published number. If the authors can re-run a subset of expert baselines under the same protocol, or reframe the expert comparison as context rather than as a head-to-head superiority claim, the paper would be much stronger. The missing post-processing training details also need to be supplied for reproducibility. The paper's scope fits a methods journal, but the overclaiming on expert models should be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere is my read on arXiv:2411.15251. The paper is a serious engineering effort with a real but circumscribed evaluation problem. The core architecture is a sensible assembly of known parts: frozen SAM ViT with adapters, a ConvNeXt-tiny micro-vessel branch with CBAM and FPN, shallow-feature fusion into the mask decoder, and a U-Net post-processor trained with MSE and clDice. The compilation of 17 vessel datasets across four modalities is genuinely useful. The controlled part of the experiments is the best feature: six SAM-based methods retrained under the same protocol, with standard deviations reported, and OVS-Net wins on average Dice, IoU, clDice, and beta0. The ablation, despite a minor numerical slip, shows each module helps.\n\nThe load-bearing weakness is the comparison against '17 expert models.' Those scores are imported from prior papers with different splits, preprocessing, and metric conventions; the authors did not retrain or re-run any of them. They also select the best published Dice per dataset, which biases the comparison. On HRF the lead over SuperVessel is 0.40 Dice, a margin that could easily be explained by protocol differences. So the headline that OVS-Net surpasses 17 expert models is not established by the evidence as presented. The internal SAM comparison is the reliable claim, and it is a good one.\n\nOther soft spots: external generalization is tested on only two fundus datasets (HRF, IOSTAR), so 'multi-modality generalization' beyond the training domains is thin. The ablation text claims a 1.31-point Dice improvement from the micro-vessel module, but Table VIII shows 88.24 vs 86.97 = 1.27 (or 1.03 vs the adapter row). Small, but worth correcting. Code and data are promised but not available yet.\n\nWho gets value: anyone working on SAM-based medical segmentation or topology-aware post-processing. The paper deserves peer review, but I would send it back for a major revision that re-runs or clearly re-implements the expert baselines on held-out splits, adds at least one non-fundus external test, and fixes the ablation text. If the authors release code and data, the internal comparison becomes independently checkable, which would make this a solid contribution.\n\nVerdict: conditional, leaning positive on the method, skeptical on the expert-model claims.","headline":"A useful SAM-based vessel segmenter whose internal comparison is solid, but the 'surpass 17 expert models' claim relies on scores imported from heterogeneous prior papers and is not established.","tokens_in":20030,"tokens_out":2548,"would_cite":false,"duration_ms":29580,"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":"A single model segments vessels across four imaging modalities, with 34.6% better connectivity than SAM-based rivals.","keywords":["vessel segmentation","structure-agnostic","Segment Anything Model","micro vessel enhancement","morphological correction","topology preservation","ClDice","multi-modality"],"falsifier":"Run all 17 expert models on the exact training, validation, and test splits used for OVS-Net and recompute Dice, IoU, ClDice, and $\\beta_0$; the claim of surpassing expert models fails if a majority of the re-run experts match or exceed OVS-Net's scores.","tokens_in":19060,"feed_emoji":"🩺","tokens_out":2518,"duration_ms":26117,"temperature":0.7,"pith_summary":"This paper tries to show that one vessel-segmentation model can work across four imaging modalities instead of needing a separate specialist per anatomy. It argues that the key obstacles are small, low-contrast vessels and broken connectivity, and that both can be addressed by augmenting the Segment Anything Model with a micro-vessel branch and a post-processing network that repairs disconnected segments. If the claim holds, a single model could serve coronary, retinal, OCTA, and pelvic vessel analysis while producing fewer broken vessels.","feed_headline":"One model beats specialists at vessel segmentation on 17 datasets","feed_subtitle":"A SAM-based network with a micro-vessel branch and topology repair lifts Dice and cuts broken vessels by 34.6%.","key_machinery":"The central mechanism is a hybrid encoder built on SAM: the frozen ViT image encoder acts as the macro-vessel extraction branch, while a ConvNeXt-based micro-vessel enhancement branch with CBAM attention and a feature pyramid captures fine vascular details; cross-attention fuses the two branches. The mask decoder then merges low-level ConvNeXt features into SAM's decoder features to restore boundary detail. A separate U-Net post-processing network, trained with MSE and ClDice losses, takes the initial mask and repairs broken vessel segments while preserving natural discontinuities.","core_discovery":"The paper claims that OVS-Net, a structure-agnostic adaptation of SAM, outperforms both fine-tuned SAM variants and 17 task-specific expert models on multi-modality vessel segmentation, while improving vascular connectivity by 34.6% over the best SAM-based baseline. The evidence is a compiled benchmark of 17 datasets spanning X-ray coronary artery, fundus retina, OCTA retina, and X-ray pelvic iliac artery images, evaluated with Dice, IoU, ClDice, and the normalized Betti number $\\beta_0$. The authors present OVS-Net's gains as coming from its hybrid encoder and a morphology-repair post-processor, and report that it also holds up on two held-out external datasets.","pith_inferences":["The same post-processing idea may transfer to other thin, branching structures such as neurons, roads, or river networks, where fragmentation is the dominant failure mode.","If the comparability of quoted expert-model scores is set aside, the most robust evidence in the paper is the internal SAM-based comparison, where all baselines were retrained under the same protocol.","A natural next test is whether the model still generalizes to 3D vessel data or to modalities absent from the 17 datasets, such as MRI angiography, since the current claim is limited to 2D images."],"forward_implications":["A single OVS-Net could replace multiple anatomy-specific vessel segmenters for clinical deployment, cutting the need for retraining per modality.","Connectivity gains of the reported size mean the output is more likely to support hemodynamic and topological analyses that overlap-based metrics alone do not capture.","The design suggests that freezing a large ViT encoder and training only adapters plus a small CNN branch can yield a generalizable medical segmentation model.","The morphology-repair post-processor could be attached to other base segmenters, not just OVS-Net, to improve their vessel connectivity.","Reporting $\\beta_0$ alongside Dice could become a standard practice for evaluating tubular-structure segmentation quality."],"supporting_citations":[{"why":"Supplies the SAM ViT encoder and mask decoder that OVS-Net adapts as its backbone.","marker":"[8]"},{"why":"HQ-SAM provides the idea of fusing early and final ViT features into the mask decoder to recover mask detail.","marker":"[28]"},{"why":"MedSAM is the primary medical SAM baseline against which OVS-Net is compared in the internal validation.","marker":"[21]"},{"why":"Introduces ClDice, the topology-preserving loss and connectivity metric used for training and evaluation.","marker":"[12]"},{"why":"Reports the expert-model connectivity scores on DRIVE, CHASE, and DCA1 used in Table VII.","marker":"[31]"},{"why":"Masked Autoencoders is the pre-training method for the ViT encoder that OVS-Net keeps frozen.","marker":"[36]"},{"why":"ConvNeXt is the convolutional backbone of the micro-vessel enhancement module.","marker":"[37]"}],"fun_headline_variants":["One model beats 17 expert models in vessel segmentation","Structure-agnostic vessel segmentation wins on 17 datasets","Vessel segmentation: 34.6% connectivity boost with OVS-Net","Universal vessel segmentation: OVS-Net outdoes specialists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that Dice and IoU numbers quoted from earlier expert-model papers are directly comparable to its own scores, even though those models were trained on different data splits, preprocessing, and sometimes datasets.","fun_headline_variants_meta":{"raw":{"variants":["One model beats 17 expert models in vessel segmentation","Structure-agnostic vessel segmentation wins on 17 datasets","Vessel segmentation: 34.6% connectivity boost with OVS-Net","Universal vessel segmentation: OVS-Net outdoes specialists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00094,"raw_usage":{"total_tokens":3998,"prompt_tokens":903,"completion_tokens":3095,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":3024}},"tokens_in":519,"tokens_out":3095,"duration_ms":23554,"temperature":1.0,"reasoning_tokens":3024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:52:56.504888+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run all 17 expert models on the exact training, validation, and test splits used for OVS-Net and recompute Dice, IoU, ClDice, and $\\beta_0$; the claim of surpassing expert models fails if a majority of the re-run experts match or exceed OVS-Net's scores.","supporting_citations":[{"cited_title":"Segment Anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the SAM ViT encoder and mask decoder that OVS-Net adapts as its backbone."},{"cited_title":"Segment Anything in High Quality,","cited_arxiv_id":null,"evidence_quote":"HQ-SAM provides the idea of fusing early and final ViT features into the mask decoder to recover mask detail."},{"cited_title":"Segment anything in medical images,","cited_arxiv_id":null,"evidence_quote":"MedSAM is the primary medical SAM baseline against which OVS-Net is compared in the internal validation."},{"cited_title":"clDice – A Novel Topology- Preserving Loss Function for Tubular Structure Segmentation,","cited_arxiv_id":null,"evidence_quote":"Introduces ClDice, the topology-preserving loss and connectivity metric used for training and evaluation."},{"cited_title":"Deep Closing: Enhancing Topological Connectivity in Medical Tubular Segmentation,","cited_arxiv_id":null,"evidence_quote":"Reports the expert-model connectivity scores on DRIVE, CHASE, and DCA1 used in Table VII."},{"cited_title":"Masked Autoencoders Are Scalable Vision Learners,","cited_arxiv_id":null,"evidence_quote":"Masked Autoencoders is the pre-training method for the ViT encoder that OVS-Net keeps frozen."},{"cited_title":"A ConvNet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"ConvNeXt is the convolutional backbone of the micro-vessel enhancement module."}],"review_version":1}