{"id":"b60e0d1a-7757-4ed3-8ccc-185d4722d4f6","arxiv_id":"2411.17386","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"vesselFM, trained on three heterogeneous 3D blood vessel data sources, achieves state-of-the-art zero-, one-, and few-shot vessel segmentation across four imaging modalities.","lead":"The authors trained a single 3D model to segment blood vessels using real, randomized, and AI-generated training images. Across MRI, CT, OCT, and electron microscopy data it outperformed four existing medical segmentation models in zero- and few-shot tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot margins on BvEM and OCTA rest on 3 and 2 test volumes from single acquisitions; a full-volume BvEM evaluation is needed to confirm the large-margin claim.","rationale":"The paper's strongest empirical evidence is the zero-shot column of Table 2, where vesselFM exceeds the next-best baseline by 56.6 Dice points on BvEM and by 10.9 on OCTA. These two datasets have the smallest and least independent test sets: BvEM is one volume from which three 500^3 subvolumes are carved after reserving 130 slices for fine-tuning and a 120-slice buffer; OCTA has only two test volumes. The main table has no error bars, and the paired t-tests in Suppl K are underpowered and inappropriate for n=2-3, especially when the 'samples' are spatially correlated subvolumes of a single acquisition. If those subvolumes are not representative of the full volume, the reported large margins on vEM and OCTA would not reflect universal zero-shot transfer. I therefore regard this as the most load-bearing concern, rather than the synthetic-prior issue flagged by the reader: the ablation in Table 6 shows that adding Ddrand improves MSD8 (liver CT) from 5.60 to 29.84 Dice, providing indirect evidence that the corrosion-cast geometry, after heavy augmentation, generalizes beyond brain-like domains. The proposed test, full-volume BvEM evaluation of vesselFM and the strongest baseline, directly settles whether the headline margin is an artifact of subvolume selection. The reader's conditional verdict already requires stronger statistical reporting, so this concern does not change the verdict; it sharpens the required evidence.","tokens_in":20851,"tokens_out":17506,"duration_ms":169041,"concrete_test":"Download the released vesselFM checkpoint and run zero-shot inference with a sliding window over the entire BvEM volume, excluding the 130 fine-tuning/validation slices and the 120-slice buffer, and compute aggregate Dice and clDice against the full ground truth. Apply the same full-volume inference to tUbeNet (the next-best vessel-specific baseline) and, if feasible, MedSAM-2. If vesselFM's margin over the next-best baseline on the full volume falls below 20 Dice points, the Table 2 BvEM result is not robust; if it remains above 20, the statistical concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2's most dramatic zero-shot numbers are BvEM 67.49 Dice (next-best 10.92) and OCTA 46.94 (next-best 36.01). BvEM consists of a single volume: after reserving the first 130 slices for fine-tuning/validation and a 120-slice buffer, the evaluation uses only three 500^3 subvolumes. OCTA uses two test volumes. These are not independent samples, Table 2 reports no variance, and the paired t-tests in Suppl K with n=2-3 cannot support reliable p-values. If the three chosen subvolumes are easier than the volume as a whole, the 56-point margin overstates vesselFM's zero-shot vEM transfer. Since the central claim 'outperforms ... by a large margin' is quantified by exactly these entries, the evaluation's statistical fragility is the most load-bearing weakness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes vesselFM, a 3D blood vessel segmentation foundation model trained on three heterogeneous data sources: a curated real dataset Dreal (23 datasets, >115k patches), a domain-randomized synthetic dataset Ddrand (500k pairs), and a flow-matching-generated dataset Dflow (10k pairs). The authors claim zero-shot, one-shot, and few-shot generalization to four unseen evaluation datasets (SMILE-UHURA, BvEM, OCTA, MSD8), outperforming strong baselines (tUbeNet, VISTA3D, SAM-Med3D, MedSAM-2) by large margins, and they provide extensive ablations of data sources, the generative model, and the segmentation architecture.","tokens_in":21004,"tokens_out":6828,"duration_ms":68015,"significance":"The paper makes a potentially valuable empirical contribution: it assembles the largest curated real dataset for 3D vessel segmentation, introduces a domain-randomization pipeline tailored to vascular geometry, extends flow matching to mask- and class-conditioned medical image generation, and shows that the combination improves zero-shot transfer. The open-source code and checkpoints, the systematic ablations, the inclusion of clDice as a topology-aware metric, and the comparison to specialist models are genuine strengths. If the reported generalization holds under more rigorous evaluation, the work would be a useful resource for the community. The main uncertainty is statistical: the most dramatic margins (BvEM zero-shot 67.49 vs. next-best 10.92, OCTA zero-shot 46.94 vs. next-best 36.01) rest on very few test volumes from single acquisitions, and Table 2 reports no variance or confidence intervals.","major_comments":[{"comment":"The headline claim of 'outperforming all baselines by a large margin' is quantified on very small, non-independent test sets. According to Suppl. A, the BvEM test set consists of three 500^3 subvolumes extracted from a single volume, and the OCTA test set consists of two volumes. Table 2 reports only point estimates, with no per-volume results, standard deviations, or confidence intervals. The paired t-tests in Suppl. K with n=2 (OCTA) and n=3 (BvEM) are not reliable for population inference, and the BvEM subvolumes are not independent samples. This is load-bearing because the zero-shot BvEM and OCTA entries drive the 'large margin' claim. Please report per-subject/per-volume scores, provide confidence intervals (e.g., bootstrap over volumes), and either perform a full-volume BvEM evaluation or explicitly qualify the claim as limited to these small test sets.","section":"Section 4.1, Table 2; Suppl. A; Suppl. K"},{"comment":"The ablation claim that Dflow is 'consistently beneficial' is not fully supported by the detailed results. Table 6 shows that on OCTA zero-shot, Dreal+Ddrand achieves Dice 47.02 while Dreal+Ddrand+Dflow achieves 46.94, and on MSD8 zero-shot the corresponding scores are 29.84 and 29.69. The text in Section 4.2 and Fig. 9 relies on an average across datasets, but two of four datasets show small degradations. To support the claim, please report per-dataset variability and either temper the wording or apply a statistical test that accounts for dataset-level variation. This matters because the contribution of Dflow is a central methodological component.","section":"Section 4.2, Table 6, Fig. 9"},{"comment":"The synthetic vascular prior is asserted to 'accurately preserve both general angioarchitectural and morphological properties characteristic of 3D blood vessels' without validation against an independent geometric benchmark. Since BvEM and OCTA are both mouse-brain datasets and the Ddrand masks originate from mouse-brain corrosion casts, the zero-shot transfer to these modalities could be partly attributable to a species/organ match rather than to universal vascular priors. A concrete test would be to evaluate on a non-mouse-brain vEM or OCTA dataset, or to include a quantitative topological/morphological comparison between Msyn and the target-domain annotations. The large margins on SMILE-UHURA and MSD8 partially mitigate this concern, but the most dramatic zero-shot results (BvEM, OCTA) are exactly the ones affected.","section":"Section 3.2; Section 4.1"}],"minor_comments":[{"comment":"The table lists BvEM with 21,858 patches, which may be confusing because the text states that the first four datasets are exclusively evaluation datasets and are excluded from Dreal. Please clarify that the patch count for evaluation datasets is informational only and does not indicate inclusion in Dreal.","section":"Table 1"},{"comment":"The claim that the chosen UNet 'surpasses' transformer-based networks is technically true but the margin over SwinUNETR-V2 (Dice 74.66 vs. 74.54, clDice 75.27 vs. 74.80) is very small. Please report significance or avoid wording that implies a meaningful superiority over this near-tie.","section":"Section 4.2, Table 5"},{"comment":"The statistical analysis reports only p-values without test statistics, degrees of freedom, or effect sizes. For n=2 or n=3, please provide the actual paired differences and a description of the non-independence of BvEM subvolumes; the current presentation is too terse to be interpretable.","section":"Suppl. K"},{"comment":"The authors acknowledge that vesselFM's tendency to segment unlabeled tubular structures may 'artificially deflate' its reported Dice. This is an honest and important caveat, but it also underscores the need for per-volume results and possibly a secondary evaluation that accounts for over-segmentation (e.g., relaxed ground truth or human review).","section":"Suppl. I"},{"comment":"There are minor typographical issues, including 'V oronoi' (Section 3.2), 'HR-Kindney' (Suppl. A), and inconsistent use of 'vascular' vs. 'blood vessel' terminology. These do not affect the technical content.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with a large amount of engineering and experimentation. The central concern is not the method itself but the statistical support for the strongest claims: the most impressive margins come from 2-3 test volumes with no uncertainty quantification. I believe this can be fixed within the manuscript's scope by adding per-volume results, confidence intervals, a more careful statistical analysis, and a tempered wording of the Dflow ablation claim. I do not see grounds for rejection. The fit with the journal is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on vesselFM. The paper delivers a genuinely useful artifact: a universal 3D blood vessel segmentation model, trained on a curated real dataset (Dreal, 23 datasets, over 115k patches), a domain-randomization pipeline (Ddrand), and mask-conditioned flow-matching samples (Dflow). The zero-shot results on SMILE-UHURA (74.66 vs next-best 48.32 on ten test volumes) are striking. The ablations are thorough and tell a credible causal story: Dreal alone gets 65.45 Dice, adding Ddrand lifts it to 69.38, and adding Dflow reaches 74.66. Each data source earns its keep. The open-source promise and the detailed supplementary material also count in the paper's favor.\n\nThe soft spots are real but mostly addressable. Table 2 reports single numbers with no variance; the BvEM zero-shot margin (67.49 vs 10.92) rests on three 500^3 subvolumes from one acquisition, and OCTA on two test volumes. The paired t-tests in Suppl. K with n=2-3 cannot carry the weight the paper puts on them. A referee should ask for per-volume results and at least a full-volume BvEM evaluation before trusting that 56-point gap. The omission of TriSAM, a vessel-specific baseline built for exactly the BvEM setting, is conspicuous; the paper cites it but never compares against it. That needs to be fixed.\n\nA subtler point: the synthetic geometry prior (corrosion casts) feeds both Ddrand and the masks used to sample Dflow, so the model's notion of 'vesselness' is anchored in that prior. This is self-distillation, not circularity, and the authors are honest about the model also segmenting axons and colon in vEM and CT. They argue this deflates their scores, which is plausible.\n\nThe central claim—that a foundation model trained on these three heterogeneous sources generalizes zero-shot across modalities—survives the statistical critique, thanks partly to the larger SMILE-UHURA and MSD8 evaluations. But the most dramatic numbers are the least robust. The paper deserves a serious referee, provided the authors are pushed to report error bars, per-volume results, and the TriSAM comparison.","headline":"Solid empirical contribution with a genuinely useful open model; the headline zero-shot margins on two datasets rest on very few test volumes and need a TriSAM comparison.","tokens_in":21569,"tokens_out":2745,"would_cite":true,"duration_ms":28330,"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 trained on three heterogeneous data sources can segment 3D blood vessels in unseen imaging modalities, outperforming specialist and general-purpose baselines without fine-tuning.","keywords":["3D blood vessel segmentation","foundation model","zero-shot generalization","domain randomization","flow matching","synthetic medical images","medical image segmentation","tubular structure segmentation"],"falsifier":"Apply vesselFM zero-shot to a 3D vascular dataset whose contrast mechanism is unlike anything in Dreal, Ddrand, or Dflow — for example, human retinal OCTA or coronary X-ray angiography — and compare its Dice against the four baselines from Table 2; the paper's universality claim predicts vesselFM should remain well above the next-best baseline without fine-tuning, while a collapse to near-baseline performance would falsify the claim that the synthetic vascular prior covers genuinely new domains.","tokens_in":20634,"feed_emoji":"🩸","tokens_out":8214,"duration_ms":76384,"temperature":0.7,"pith_summary":"vesselFM is a foundation model built specifically for 3D blood vessel segmentation. The paper's central claim is that training a single segmentation network on three deliberately heterogeneous data sources — a large curated collection of real vascular images, a domain-randomized synthetic corpus, and images sampled from a flow-matching generative model — produces a model that segments vessels in imaging domains it has never seen, without any fine-tuning. The authors report that this model outperforms four state-of-the-art segmentation foundation models across four (pre-)clinically relevant modalities in zero-, one-, and few-shot settings, with zero-shot Dice of 74.66 on ultra-high-field MRA and 67.49 on volume electron microscopy versus next-best scores of 48.32 and 10.92. If true, the implication is that a single open-sourced model could replace dataset-specific annotation and training pipelines for vascular imaging, including for new modalities where annotated data does not yet exist. The paper's method rests on the assumption that corrosion-cast vascular geometry is a bias-free prior for vascular morphology across modalities.","feed_headline":"Zero-shot vessel segmentation jumps to 74.7 Dice on unseen MRAs","feed_subtitle":"Trained on real, randomized, and flow-generated images, the model beats specialist and general-purpose baselines in four modalities.","key_machinery":"The load-bearing machinery is the three-source training distribution, not a new network design. The segmentation model is a standard 3D U-Net (DynUNet-style, 31.4M parameters). What carries the argument is (1) Dreal, more than 115,000 curated $128^{3}$ real image-mask patches spanning 23 datasets and 17 sources; (2) Ddrand, a domain-randomization pipeline in which foreground masks are generated by applying random spatial and artifact transformations to 1,137 corrosion-cast vascular patches and merged into Perlin-noise, Voronoi, or plain backgrounds with a wide range of intensity transformations; and (3) Dflow, synthetic images drawn from a mask- and class-conditioned flow-matching generative model, which learns a velocity field mapping Gaussian noise to data through an ordinary differential equation and is conditioned on the same synthetic masks from Msyn. Together these sources are meant to cover the full domain of 3D vascular images and to broaden Dreal's distribution in a data-driven way.","core_discovery":"On its own terms, the paper's discovery is that the failure of existing foundation models to segment blood vessels is a data-distribution problem, not a network-capacity problem. By replacing a single homogeneous training set with three heterogeneous sources — real annotated volumes, domain-randomized synthetic volumes, and volumes sampled from a mask- and class-conditioned flow-matching generator — the same 3D U-Net becomes a universal vessel segmenter. The paper reports zero-shot Dice of 46.94 on OCTA, 67.49 on BvEM, 74.66 on SMILE-UHURA, and 29.69 on MSD8, each above every baseline; on BvEM the next-best baseline scores 10.92. One- and few-shot fine-tuning with one or three $128^{3}$ patches improves these scores further, and ablations show each data source contributes: adding Ddrand and Dflow to Dreal raises SMILE-UHURA zero-shot Dice from 65.45 to 74.66. The authors conclude that vesselFM provides an out-of-the-box solution for 3D blood vessel segmentation across (pre-)clinically relevant modalities.","pith_inferences":["The same three-source recipe — real curated data, a geometric prior from corrosion casts, and a generative model — may transfer to other tubular structures such as axons, bile ducts, or lymphatics, since the paper observes the model already segments axons and colon; this is an extension, not a paper claim.","The reported universality is bounded by the evaluation domains; a stronger test would be to apply vesselFM zero-shot to modalities with very different contrast, such as human retinal OCTA or X-ray angiography, which the paper does not report.","The synthetic-mask prior could be validated more directly by comparing vessel-radius and tortuosity statistics of Msyn against ground truth in the four evaluation datasets; if the statistics mismatch, the zero-shot margin would be expected to shrink.","The 70/20/10 sampling weights and 128^3 patch size are engineering choices whose sensitivity is untested; an independent replication varying these would reveal how robust the zero-shot margin is."],"forward_implications":["One open-sourced model can replace dataset-specific training for 3D vessel segmentation in unseen modalities, including modalities whose contrast mechanism (OCTA, vEM) is absent from training.","Bootstrapping annotations becomes practical: fine-tuning with a single 128^3 patch improves vesselFM's Dice on BvEM from 67.49 to 78.27 and on OCTA from 46.94 to 72.10, so clinicians can adapt the model to a new protocol with minimal labeling.","The two synthetic sources are not interchangeable with real data: ablations show Dreal alone reaches 65.45 Dice on SMILE-UHURA, and adding Ddrand and Dflow adds 9.21 Dice, implying that synthetic data generation is a necessary part of the recipe, not a convenience.","General-purpose medical segmentation foundation models (SAM-Med3D, MedSAM-2, VISTA3D) are not reliable zero-shot vessel segmenters; a task-specific foundation model with a tubular inductive bias is needed.","Because the generative model outperforms a diffusion baseline (Med-DDPM) by 4.32 Dice when used as a data source, flow matching is claimed as the better synthetic-image engine for this task."],"supporting_citations":[{"why":"Supplies the 1,137 corrosion-cast vascular patches that seed all synthetic mask geometry in Ddrand and Dflow.","marker":"[54]"},{"why":"Provides the conditional flow matching objective and straight-line ODE used to train the generative model F.","marker":"[32]"},{"why":"Establishes the domain randomization strategy for medical image segmentation that Ddrand adapts to 3D vessels.","marker":"[4]"},{"why":"Provides the mask-conditioned diffusion baseline that F extends with flow matching and class conditioning, and that is compared in ablations.","marker":"[14]"},{"why":"Defines the nnU-Net/DynUNet 3D U-Net architecture used as vesselFM's segmentation backbone.","marker":"[26]"},{"why":"SMILE-UHURA is the unseen MRA evaluation dataset where vesselFM reports 74.66 zero-shot Dice.","marker":"[9]"},{"why":"BvEM is the unseen vEM evaluation dataset where vesselFM reports 67.49 zero-shot Dice versus 10.92 for the next-best baseline.","marker":"[51]"},{"why":"VISTA3D is the CT-specific foundation-model baseline that vesselFM outperforms even on MSD8 CT data VISTA3D was trained on.","marker":"[23]"},{"why":"SAM-Med3D is the general-purpose medical segmentation foundation-model baseline that fails zero-shot on vessels.","marker":"[53]"},{"why":"MedSAM-2 is the SAM-2-based medical segmentation baseline that fails zero-shot on vessels.","marker":"[61]"}],"fun_headline_variants":["VesselFM: zero-shot vessel segmentation across 4 modalities","Universal vessel segmentation: zero-shot beats all baselines","One model, zero training: vessel segmentation for any modality","VesselFM beats all baselines in zero-shot vessel segmentation","Three data sources, one model: segment any vessel zero-shot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 1,137 corrosion-cast vascular patches used to generate synthetic masks capture the general shape, scale, and connectivity properties of blood vessels well enough that a model trained on those masks transfers to modalities with entirely different contrast mechanisms, such as OCTA flow signal and electron-microscopy ultrastructure.","fun_headline_variants_meta":{"raw":{"variants":["VesselFM: zero-shot vessel segmentation across 4 modalities","Universal vessel segmentation: zero-shot beats all baselines","One model, zero training: vessel segmentation for any modality","VesselFM beats all baselines in zero-shot vessel segmentation","Three data sources, one model: segment any vessel zero-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001,"raw_usage":{"total_tokens":4260,"prompt_tokens":1001,"completion_tokens":3259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":3176}},"tokens_in":617,"tokens_out":3259,"duration_ms":21429,"temperature":1.0,"reasoning_tokens":3176,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:09:24.418428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply vesselFM zero-shot to a 3D vascular dataset whose contrast mechanism is unlike anything in Dreal, Ddrand, or Dflow — for example, human retinal OCTA or coronary X-ray angiography — and compare its Dice against the four baselines from Table 2; the paper's universality claim predicts vesselFM should remain well above the next-best baseline without fine-tuning, while a collapse to near-baseline performance would falsify the claim that the synthetic vascular prior covers genuinely new domains.","supporting_citations":[{"cited_title":"Simulation-Based Segmentation of Blood Vessels in Cerebral 3D OCTA Images","cited_arxiv_id":null,"evidence_quote":"Supplies the 1,137 corrosion-cast vascular patches that seed all synthetic mask geometry in Ddrand and Dflow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the conditional flow matching objective and straight-line ODE used to train the generative model F."},{"cited_title":"SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining","cited_arxiv_id":null,"evidence_quote":"Establishes the domain randomization strategy for medical image segmentation that Ddrand adapts to 3D vessels."},{"cited_title":"Conditional Diffusion Models for Semantic 3D Brain MRI Synthesis","cited_arxiv_id":null,"evidence_quote":"Provides the mask-conditioned diffusion baseline that F extends with flow matching and class conditioning, and that is compared in ablations."},{"cited_title":"nnU-Net: a self-configuring method for deep learning-based biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"Defines the nnU-Net/DynUNet 3D U-Net architecture used as vesselFM's segmentation backbone."},{"cited_title":"SMILE-UHURA Challenge -- Small Vessel Segmentation at Mesoscopic Scale from Ultra-High Resolution 7T Magnetic Resonance Angiograms","cited_arxiv_id":"2411.09593","evidence_quote":"SMILE-UHURA is the unseen MRA evaluation dataset where vesselFM reports 74.66 zero-shot Dice."},{"cited_title":"TriSAM: Tri-Plane SAM for zero-shot cortical blood vessel segmentation in VEM images","cited_arxiv_id":"2401.13961","evidence_quote":"BvEM is the unseen vEM evaluation dataset where vesselFM reports 67.49 zero-shot Dice versus 10.92 for the next-best baseline."},{"cited_title":"Medical SAM 2: Seg- ment medical images as video via Segment Anything Model","cited_arxiv_id":null,"evidence_quote":"MedSAM-2 is the SAM-2-based medical segmentation baseline that fails zero-shot on vessels."}],"review_version":1}