{"id":"6add2d62-9f6e-47e4-acf1-9ea8dd50cef0","arxiv_id":"2505.14717","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Aneumo is a large open dataset pairing 10,660 deformed aneurysm geometries with 85,280 CFD flow fields and masks, benchmarked with a DeepONet-SwinT model.","lead":"The authors built Aneumo, a dataset of 10,660 synthetic aneurysm shapes derived from 427 real patient geometries, each paired with CFD-simulated blood flow fields and segmentation masks. It matters because large open datasets like this could let machine learning models predict aneurysm hemodynamics in real time, which currently takes hours of simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic aneurysm shapes from polygon offset (App. A.1) are validated only subjectively; volume change rates up to 3.5 (Sec. 3.4) suggest non-physiological geometries, threatening the dataset's clinical transferability.","rationale":"The paper's central contribution is a large-scale dataset that combines realistic aneurysm geometry with CFD fields for machine learning. The single most load-bearing assumption is that the 10,660 synthetic geometries, created by simple polygon offset of aneurysm-free vessels, accurately emulate the morphological evolution of real intracranial aneurysms. If this assumption fails, the entire dataset's clinical relevance and the benchmark's evidence of generalization are undermined. The paper's only support is a subjective statement of neurosurgeon confirmation, with no quantitative validation, and its own data analysis reveals volume change rates up to 3.5, which the authors acknowledge as anomalous. This is not merely a disagreement with external consensus; it is a gap between the claimed fidelity and the described generation pipeline. The flow-rate count inconsistency (Sec. 3.3 lists 0.0030 and 0.003 as separate entries, while App. A.3 lists seven distinct rates) is a concrete numerical error but does not threaten the dataset's validity as severely as geometric non-representativeness. The reader's verdict of CONDITIONAL is appropriate because the concern is addressable through a quantitative morphological validation or a more rigorous expert study; if such evidence is provided, the dataset claim would be substantially strengthened. No change to the verdict is needed from this stress-test pass, as the concern reinforces the reader's existing conditionality rather than demanding rejection.","tokens_in":27962,"tokens_out":7441,"duration_ms":70781,"concrete_test":"Compute standard morphological indices (aspect ratio, size ratio, dome height, neck width, irregularity/undulation index) on all 10,660 synthetic aneurysm sacs and compare their joint distribution against the 427 real AneuX aneurysms (or an external cohort such as ADAM or CHUV). If a substantial fraction (e.g., >5%) of synthetic shapes falls outside the real distribution, or if the two distributions are statistically distinguishable with large effect size (e.g., Bhattacharyya distance >0.5), the deformation process does not represent natural aneurysm evolution. Alternatively, run a blinded forced-choice rating by 5+ neurosurgeons on real versus synthetic shapes and report inter-rater agreement (e.g., Cohen's kappa); chance-level or low agreement would confirm the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Aneumo is a high-fidelity resource for aneurysm ML rests on the 10,660 synthetic shapes being representative of real aneurysm morphology. Appendix A.1 describes generating each synthetic aneurysm by selecting a region on the aneurysm-free vessel and applying Geomagic Wrap's polygon offset with a random distance d in [0.5, 1.0], producing a smooth radial bulge. This method does not reproduce the complex, asymmetric morphologies (blebs, daughter sacs, irregular walls) that characterize clinical aneurysms. The only authenticity support is the statement that neurosurgeons confirmed the shapes (Sec. 3.1); no inter-rater reliability, blinding, or quantitative morphological comparison is reported. Section 3.4 shows volume change rates reaching 3.5, i.e., the deformed model is 3.5 times the volume of the healthy reference vessel; the authors label such cases 'anomalies.' If a non-negligible fraction of synthetic geometries lies outside the physiological range, then the 85,280 CFD fields are computed on biased geometry, and models trained on them will not transfer to clinical data. The benchmark's test set contains only 10 real geometries from AneuX, the same source used to create the synthetic shapes, so it is too small and potentially confounded to validate geometric realism. The paper's own Limitations section (Sec. 6) concedes only the steady-state simplification and does not address deformation authenticity, leaving this assumption unexamined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Aneumo, a dataset of 10,660 synthetic intracranial aneurysm geometries derived from 427 real AneuX geometries, together with 85,280 CFD solutions (velocity and pressure fields) generated under a nominal set of eight steady-state mass-flow conditions, segmentation masks, and supporting mesh and VTK/NumPy outputs. The authors also present a benchmark comparing a standard DeepONet with a proposed DeepONet-SwinT architecture for predicting hemodynamic fields, reporting improved accuracy and efficiency on a 20-case test set drawn from AneuX.","tokens_in":28305,"tokens_out":4912,"duration_ms":49937,"significance":"If the dataset counts and the independence claims are verified, Aneumo would be a valuable community resource: it is among the largest open collections pairing aneurysm geometry with field-level CFD hemodynamic data, and it provides multimodal outputs (masks, point clouds, 3D models, flow fields) plus an operator-learning benchmark. The paper also demonstrates a concrete baseline and a scale-up analysis. However, the scientific value depends on two load-bearing points that currently need correction: the exact number of distinct flow conditions and the actual independence of the test set from the training-generation source.","major_comments":[{"comment":"The number of distinct mass-flow conditions is inconsistent. Section 3.3 lists eight values including both 0.0030 and 0.003 kg/s, which are numerically the same value, while Appendix A.3 lists seven distinct values: 0.0010, 0.0015, 0.0020, 0.0025, 0.0030, 0.00375, and 0.0040 kg/s. Since the headline count 85,280 equals 10,660 × 8, the presence of a duplicate means the actual number of unique CFD cases would be 10,660 × 7 = 74,620 unless simulations were intentionally repeated under nominally identical conditions. This discrepancy affects the abstract, the contribution list, Table 1, Section 3.3, and Appendix A.3. The authors should state the exact number of distinct flow conditions, recompute all dataset counts, and correct the statement in Appendix A.3 that 85,280 simulations were completed for each three-dimensional model (that number is the total across all models, not per model).","section":"Section 3.3 and Appendix A.3"},{"comment":"The test set is not independent of the training-generation pipeline. The paper claims the test set is \"completely independent\" because it is drawn from AneuX, but the training geometries in Aneumo are themselves generated from AneuX real geometries (Section 3.1 and Appendix A.1). The 10 test geometries may therefore be the same underlying patient geometries from which synthetic training shapes were deformed, which can inflate the reported generalization performance. The authors need to demonstrate, with a concrete mapping or hash-based geometry check, that no test geometry or its deformed descendants appear in the training/validation sets, or use an external dataset for the test set.","section":"Section 4 (benchmark test set)"},{"comment":"The physiological plausibility of the synthetic aneurysm shapes is not established. The deformation method is a single polygon-offset operation with distance d randomly sampled from [0.5, 1.0], which produces a smooth radial bulge; this does not reproduce the asymmetric, blebbed, or daughter-sac morphologies typical of clinical aneurysms. The only stated validation is that neurosurgeons confirmed the shapes (Section 3.1), with no inter-rater reliability, blinding, or quantitative morphological comparison to real aneurysms. Section 3.4 further reports volume change rates reaching 3.5 and labels these as anomalies, but no threshold, count, or exclusion rule is given. The authors should add a quantitative morphological analysis (e.g., size, aspect ratio, surface irregularity distributions compared with a clinical cohort) and either exclude or separately analyze the out-of-distribution shapes so that downstream users can judge clinical transferability.","section":"Appendix A.1 and Section 3.4"},{"comment":"The geometry-generation protocol is not described reproducibly. Section 3 states that each baseline model undergoes \"at least 20 randomized non-rigid stretching transformations,\" but Appendix A.1 describes only one polygon-offset deformation with a random distance. To produce 10,660 shapes from 427 baselines, an average of about 25 outputs per baseline is needed, yet the text does not specify how many deformation trials are performed per geometry, whether the 20 transformations are a lower bound, or how the final count 10,660 was reached. This must be clarified in the supplementary material, along with the exact parameters of the transition band and the post-deformation smoothing steps.","section":"Section 3 and Appendix A.1"}],"minor_comments":[{"comment":"The phrase \"withth high accuracy\" contains a typo; it should read \"with high accuracy.\"","section":"Section 3.2"},{"comment":"The dataset name is misspelled as \"Anuemo\" in the sentence \"A total of 85,280 CFD simulations were completed for each 3D model in the Anuemo dataset.\"","section":"Section 3.3"},{"comment":"The text says \"105 iterations\" where the intended meaning is clearly 10^5 iterations; please use proper scientific notation.","section":"Appendix A.3"},{"comment":"The text says \"as might be illustrated in a Figure 8, not provided here,\" but Figure 8 is in fact provided below; this internal contradiction should be removed.","section":"Appendix A.3 and Figure 8"},{"comment":"References [24] and [25] are duplicates of the same work and should be merged or replaced with distinct citations.","section":"References"},{"comment":"The scaling experiments in Figures 14 and 15 show 1000 epochs for DeepONet-SwinT, while the main text and earlier figures report 5000 epochs; the training durations should be stated consistently.","section":"Appendix B.3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Peter, the short version: Aneumo is a real resource—10,660 synthetic aneurysm geometries with 85,280 steady-state CFD velocity/pressure fields plus segmentation masks, all open. The ML-for-hemodynamics community has nothing at this scale. The contribution is the data, not the DeepONet-SwinT benchmark, which is an incremental architecture combination. Treat the dataset as promising but not yet fully trustworthy until a few fixes land.\n\nWhat is genuinely new: no prior aneurysm dataset combines this many geometries with field-level flow fields and masks. Aneurisk and CMHA have CFD but only dozens of cases or global metrics; AneuX has 750 models but no flow data. The grid-independence check (0.15 mm, <0.01% difference from 0.10 mm) and the reported residual curves are the right kind of evidence for a CFD dataset. Shipping STL, mesh, VTK, and NPY files is the right move.\n\nWhere it gets soft. First, the deformation pipeline: the synthetic shapes are made by removing the aneurysm and applying Geomagic's polygon offset, which produces smooth radial bulges. That is not how real aneurysms look—they get blebs, daughter sacs, irregular walls. The only support is \"confirmed by neurosurgeons,\" with no inter-rater reliability, no quantitative morphological comparison to real cases, and the paper itself reports volume change rates up to 3.5. The stress-test note is on point. The paper calls these \"anomalies\" but does not say how many or whether excluding them changes the dataset. This is the biggest risk to clinical transferability. It does not kill the dataset for surrogate-model benchmarking, but it undermines the \"authentic evolution\" claim.\n\nSecond, the flow conditions are sloppy. Section 3.3 lists eight mass flow rates but includes 0.003 and 0.0030 as separate entries; Appendix A.3 lists seven. If the actual data were generated with seven conditions, the 85,280 count is wrong. That is a straightforward check and fix.\n\nThird, the benchmark test set is 10 real AneuX geometries—the same source family as the 427 base geometries used to synthesize the training data. That is not fully independent, and with only 10 test geometries the error bars are huge. Fine for a first benchmark, but \"strictly evaluate generalization\" is overstatement.\n\nBottom line: a useful, citable dataset whose core claims are plausible and mostly reproducible, but the paper as written overstates geometric authenticity and independence. The fixes are tractable. I would send it to peer review, ask for a cleaned-up flow-condition list, a quantitative morphological validation of synthetic shapes against real aneurysm datasets, and a test set that either excludes AneuX-derived real cases or clearly reports the overlap.","headline":"A genuinely useful large open CFD dataset for aneurysm ML, but the synthetic geometry realism is under-validated and the flow-condition counts are internally inconsistent.","tokens_in":28852,"tokens_out":2837,"would_cite":true,"duration_ms":29486,"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":"The paper presents Aneumo, an open dataset of 85,280 CFD-simulated pressure and velocity fields from 10,660 aneurysm geometries, and reports that a DeepONet-SwinT model predicts those fields on unseen real geometries more accurately than…","keywords":["intracranial aneurysm","hemodynamics","computational fluid dynamics dataset","operator learning","DeepONet","Swin Transformer","multimodal medical imaging","surrogate models"],"falsifier":"Take Aneumo's synthetic shapes and the 427 parent geometries and run a simple supervised classifier on standard aneurysm morphometrics (volume change rate, aspect ratio, sphericity, neck-to-dome ratio): near-perfect separation would demonstrate distributional mismatch in synthetic shapes. A more direct test is to retrain the DeepONet-SwinT benchmark on the subset of synthetic shapes with volume change rate below 1 and evaluate on the held-out real geometry test set; if the reported accuracy comes mostly from easy, low-deformation cases, the errors on realistic extremes would reveal that the current benchmark overstates generalization.","tokens_in":27793,"feed_emoji":"🩸","tokens_out":8926,"duration_ms":90345,"temperature":0.7,"pith_summary":"This paper tries to establish that a new open dataset, Aneumo, closes a structural gap in aneurysm research: enough paired geometry-hemodynamics data to train machine-learning surrogates that can replace hours-long CFD simulations. Starting from 427 real intracranial aneurysm geometries, the authors remove the aneurysm, apply randomized polygon-offset deformations to synthesize 10,660 shapes, and simulate blood flow in each under eight steady mass-flow conditions, producing 85,280 pressure and velocity field samples together with segmentation masks, 3D surface meshes, and point-cloud representations. On a benchmark built from these data, the paper reports that a DeepONet model augmented with a Swin Transformer geometric encoder predicts pressure and velocity fields on held-out real geometries with lower normalized error and tighter error spread than the plain DeepONet baseline. If these claims hold, the dataset would make fast, geometry-based hemodynamic prediction practical, turning a per-case cost of roughly 45 minutes to 2 hours of supercomputer time into about 0.01 seconds of GPU inference, and providing a common resource for studying how morphology, flow, and rupture risk connect.","feed_headline":"85,280 simulated blood-flow fields for aneurysm AI","feed_subtitle":"Open dataset pairs 10,660 aneurysm geometries with simulated blood-flow fields and benchmarks for fast prediction.","key_machinery":"The argument runs on three coupled mechanisms. First, geometry expansion: each real aneurysm is stripped to a healthy vessel, and a polygon-offset command displaces the wall along local normals by a randomly sampled distance (0.5-1.0 in model units), with transition bands for smooth blending, yielding 10,660 synthetic shapes that are then voxelized into NIfTI masks. Second, flow-field generation: the shapes are meshed with unstructured polyhedral cells and ten prismatic boundary layers, and steady incompressible Navier-Stokes equations are solved in OpenFOAM with the icoFoam solver and PISO pressure-velocity coupling at eight mass-flow rates between 0.001 and 0.004 kg/s, with residuals driven to convergence and a mesh-sensitivity check at 0.15 mm. Third, the machine-learning surrogate: the benchmark object is DeepONet-SwinT, which keeps DeepONet's branch-trunk operator decomposition but replaces the purely MLP geometry encoding with a Swin Transformer that reads the 3D mask, adds a boundary-condition branch and a mass-flow scaler, and outputs pressure, the three velocity components, and pressure difference.","core_discovery":"The central claim is that Aneumo is the first large-scale, high-fidelity, multimodal aneurysm dataset: one public resource that combines 10,660 3D aneurysm models, 10,660 binary segmentation masks, dense point clouds, CFD meshes, and 85,280 simulated velocity/pressure fields spanning eight physiologically motivated inflow rates. The dataset is derived from 427 real aneurysm models by a deaneurysm-and-deform pipeline meant to mimic aneurysm evolution, with each parent shape subjected to at least 20 randomized non-rigid stretch transformations. The paper further claims that a benchmark on this dataset shows an operator-learning architecture, DeepONet-SwinT, learns the geometry-to-hemodynamics mapping well enough to generalize to completely unseen real aneurysm geometries, with lower MNAE and L2 errors, faster convergence, and a more concentrated error distribution than the standard DeepONet, at an inference cost of about 0.01 seconds per case. The authors present these results as evidence that data-driven hemodynamic surrogates are viable for aneurysm research and, eventually, clinical risk assessment.","pith_inferences":["The authors did not separate the effect of deformation extremes: a natural follow-up is to retrain surrogates with the anomalous shapes (volume change rate above, say, 2.0) removed and compare test error on real geometries; this would show whether the tail helps or hurts transfer.","Because only steady-state flow fields are provided, any clinical claim about rupture risk would need extra validation against transient simulations, since pulsatile wall shear stress and oscillatory indices are absent from the dataset and the paper itself lists steady-state as a limitation.","The paired geometry-flow format invites generative tasks the paper does not demonstrate, such as learning a conditional model of aneurysm growth from hemodynamic fields or using flow predictions to guide shape editing; testing these tasks on Aneumo would extend its value beyond surrogate prediction.","A cheap sanity check of the synthetic shapes is a one-class morphometry comparison: if a simple classifier separates the 10,660 synthetic shapes from the 427 parent shapes using size- and shape-normalized statistics with near-perfect accuracy, the deformation distribution is a weak model of natural aneurysm evolution, and clinical transfer should be scrutinized."],"forward_implications":["If Aneumo is as usable as claimed, training surrogate models on it turns a per-case CFD pipeline that takes 45 minutes to 2 hours on 24 CPU cores into GPU inference of about 0.01 seconds per case.","The multimodal format (masks, STL meshes, VTK fields, NPY arrays) lets the same dataset support segmentation, 3D reconstruction, point-cloud learning, and flow-field prediction tasks without re-formatting.","The benchmark's geometry-disjoint splits and independent real-geometry test set provide a common protocol for comparing operator-learning and SciML methods on aneurysm hemodynamics.","The scaling experiment, in which increasing training data from 1,280 to 12,800 cases lowers validation error for both models, suggests that continued expansion of the dataset would further improve surrogate accuracy.","The eight flow-rate sweep can be used to train models that interpolate across physiological inflow conditions, which the paper's validation-set-diversity experiments show matters more than point density for generalization."],"supporting_citations":[{"why":"Supplies the 427 real aneurysm geometries that anchor the deformation pipeline; all 10,660 synthetic shapes derive from it.","marker":"[40]"},{"why":"Defines the DeepONet operator-learning architecture used as the benchmark baseline.","marker":"[47]"},{"why":"Provides the Swin Transformer shifted-window encoder that the proposed DeepONet-SwinT uses to read 3D geometric inputs.","marker":"[46]"},{"why":"Provides the OpenFOAM solver (with icoFoam and PISO) used for all 85,280 CFD simulations.","marker":"[66]"},{"why":"Existing aneurysm dataset that includes CFD fields but only 23 mask annotations; used to establish the scale gap Aneumo fills.","marker":"[2]"},{"why":"Recent multimodal aneurysm dataset with 105 CFD cases; another baseline for the claim that Aneumo is an order of magnitude larger.","marker":"[54]"},{"why":"PISO algorithm referenced as the pressure-velocity coupling method inside the CFD pipeline.","marker":"[37]"}],"fun_headline_variants":["85K flows, 10K shapes, 0.01s inference: Aneumo","Multimodal aneurysm AI dataset with 85K flow fields","Aneumo: 85K simulated flows, 10K aneurysm shapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that aneurysm shapes produced by randomly offsetting the vessel wall by 0.5-1.0 units resemble real aneurysms closely enough that models trained on them transfer to clinical data; the paper's only check is neurosurgeon review, and its own Section 3.4 reports volume-change-rate outliers up to 3.5, indicating some generated shapes are physiologically extreme.","fun_headline_variants_meta":{"raw":{"variants":["85K flows, 10K shapes, 0.01s inference: Aneumo","Multimodal aneurysm AI dataset with 85K flow fields","Aneumo: 85K simulated flows, 10K aneurysm shapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001939,"raw_usage":{"total_tokens":7627,"prompt_tokens":1025,"completion_tokens":6602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":6536}},"tokens_in":641,"tokens_out":6602,"duration_ms":46855,"temperature":1.0,"reasoning_tokens":6536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:24:22.389395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take Aneumo's synthetic shapes and the 427 parent geometries and run a simple supervised classifier on standard aneurysm morphometrics (volume change rate, aspect ratio, sphericity, neck-to-dome ratio): near-perfect separation would demonstrate distributional mismatch in synthetic shapes. A more direct test is to retrain the DeepONet-SwinT benchmark on the subset of synthetic shapes with volume change rate below 1 and evaluate on the held-out real geometry test set; if the reported accuracy comes mostly from easy, low-deformation cases, the errors on realistic extremes would reveal that the current benchmark overstates generalization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OpenFOAM solver (with icoFoam and PISO) used for all 85,280 CFD simulations."},{"cited_title":"Juchler, S","cited_arxiv_id":null,"evidence_quote":"Supplies the 427 real aneurysm geometries that anchor the deformation pipeline; all 10,660 synthetic shapes derive from it."},{"cited_title":"AneuriskWeb project website, http://ecm2.mathcs.emory.edu/aneuriskweb","cited_arxiv_id":null,"evidence_quote":"Existing aneurysm dataset that includes CFD fields but only 23 mask annotations; used to establish the scale gap Aneumo fills."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recent multimodal aneurysm dataset with 105 CFD cases; another baseline for the claim that Aneumo is an order of magnitude larger."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PISO algorithm referenced as the pressure-velocity coupling method inside the CFD pipeline."}],"review_version":1}