{"id":"803e7271-61b8-472a-b6d3-dc2d9681cc42","arxiv_id":"2507.13404","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A volume-guided conditional diffusion model predicts aortic centerlines, which drive contour extraction and NURBS fitting to produce CFD-ready aorta meshes from CT/MRI volumes.","lead":"AortaDiff builds 3D aorta surface models from CT or MRI scans by predicting the aortic centerline with a diffusion model, then using that centerline to guide contour extraction and NURBS surface fitting. The authors show the pipeline works with only 18 training volumes and that the resulting meshes can be used in blood-flow simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'minimal dependency on large labeled datasets' claim rests on ScribblePrompt, a large pretrained segmentation model, rather than on the volume-guided CDM; the true surface generator is an external model.","rationale":"The reader identified exactly the same load-bearing assumption: the pipeline ultimately depends on ScribblePrompt, a large pretrained model, and if it fails on noisy or unseen slices the end-to-end claim collapses. I agree and would keep the verdict at CONDITIONAL, because the issue is addressable with an ablation and does not by itself falsify the reported numbers on the 12 test cases. My stress-test adds specificity: the minimal-dependency claim is not merely vulnerable to ScribblePrompt failure; it is structurally false in its current wording, since the final mesh is produced by ScribblePrompt's contours, not by the diffusion model. The CDM's role is reduced to supplying 16 prompts per branch, and the paper's own limitations section admits the centerlines can be off-center, leaning on segmentation robustness. A concrete ablation replacing ScribblePrompt would settle whether AortaDiff contributes any data-efficient surface-generation capability beyond prompt generation and NURBS fitting. Secondary concerns (centerline point count selected on the test set, the fairness of baseline comparisons, and the absence of released code/data) reinforce the conditional verdict but are less central than the ScribblePrompt dependency.","tokens_in":25624,"tokens_out":5317,"duration_ms":63532,"concrete_test":"Ablate the ScribblePrompt stage while keeping the CDM centerlines fixed: replace ScribblePrompt with (a) a non-learning segmentation method (e.g., Otsu thresholding or region growing) on the same orthogonal slices, and (b) SAM/MedSAM with the same centerline point prompts. Recompute the final mesh CD/HD/EMD from Table 4 for the VMR and AVT test cases. If the mesh metrics degrade to the level of the supervised baselines (LoGB-Net, TotalSeg) or worse, then the end-to-end and minimal-dependency claims are not supported by the proposed method itself. Additionally, report the size of ScribblePrompt's pretraining corpus to quantify the total labeled-data burden of the full pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of AortaDiff is an end-to-end workflow with minimal dependency on large labeled datasets. However, Section 3.3 shows the surface is not generated by the diffusion model at all: the CDM only outputs 16 coarse centerline points per branch, and each point is then used as a prompt to ScribblePrompt, a SAM-based model explicitly described as trained on large-scale medical datasets with simulated scribbles. Section 4.3 credits ScribblePrompt's 'strong prior knowledge gained through large-scale self-supervised pretraining' for the segmentation quality that determines the final mesh geometry (Table 4). Thus the reported mesh fidelity is largely inherited from an external large-data model, not from the proposed data-efficient CDM. The comparison against supervised baselines such as LoGB-Net and TotalSeg is therefore not a fair test of data efficiency: AortaDiff enjoys a powerful pretrained segmentation prior while those baselines are trained (or fine-tuned) from comparatively limited aortic data. The paper's own limitation statement admits the CDM-generated centerlines are 'not perfectly centered' and that 'the robustness of the segmentation module mitigates the major impact on performance,' confirming that ScribblePrompt, not the CDM, is load-bearing. If ScribblePrompt were replaced by a non-learning method or by SAM/MedSAM, the headline mesh-quality advantage and the minimal-dependency claim would likely collapse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"AortaDiff proposes a three-stage pipeline for constructing 3D aortic surface meshes from CT/MRI volumes: a volume-guided conditional diffusion model (CDM) generates ordered 16-point centerlines for the main aorta and supra-aortic branches; each centerline point is used as a prompt for ScribblePrompt to segment a vessel contour on an orthogonal slice; and the extracted contours are fitted with a NURBS surface to produce a smooth, CFD-compatible mesh. The method is trained on 18 human aorta cases from VMR and tested on four VMR plus eight AVT cases, with quantitative comparisons against point-cloud diffusion baselines for centerlines, segmentation baselines for contours, and marching-cubes/surface-nets pipelines for meshes. The paper also reports qualitative OpenFOAM CFD results comparing AortaDiff meshes with ground-truth meshes. The central claims are full automation, minimal dependency on large labeled datasets, and high geometric fidelity leading to CFD-ready surfaces.","tokens_in":1730,"tokens_out":1753,"duration_ms":69623,"significance":"If the claims hold, the practical contribution is substantial: an automatic volume-to-mesh workflow for the aorta and its branches that avoids manual contouring and uses only tens of aorta-specific training samples for the centerline generator. The modular design is a strength, and the paper ships a large amount of quantitative evaluation across centerline, contour, and mesh stages, including a parameter study and visualization of the diffusion process. The NURBS output is genuinely useful for CFD and for parametric editing. However, the significance is currently undercut by attribution problems: the part that actually produces the vessel surface is ScribblePrompt, an externally pretrained large-scale segmentation model, not the proposed diffusion model. The experimental comparisons therefore do not isolate the contribution of AortaDiff, and the 'minimal dependency on large labeled datasets' claim is overstated as written. The paper is best viewed as presenting a promising workflow whose contribution needs sharper framing and additional ablations.","major_comments":[{"comment":"The claim of 'minimal dependency on large labeled datasets' is not supported by the experimental design. The proposed CDM generates only 16 coarse centerline points per branch; the actual vessel surface is determined by ScribblePrompt segmentations on orthogonal slices (Section 3.3) and by NURBS fitting. Section 4.3 explicitly credits ScribblePrompt's 'strong prior knowledge gained through large-scale self-supervised pretraining' for the segmentation quality, and Section 6 admits that the CDM-generated centerlines are 'not perfectly centered' and that 'the robustness of the segmentation module mitigates the major impact on performance.' The comparisons against LoGB-Net and TotalSeg in Tables 2-4 are therefore not a clean test of data efficiency: AortaDiff inherits a powerful pretrained segmentation prior while the supervised baselines are trained from comparatively limited aortic data. The authors should either restrict the minimal-dependency claim to the centerline-generation module or add an end-to-end ablation that replaces ScribblePrompt with SAM, MedSAM, or a non-learning contouring method and reports the resulting mesh metrics.","section":"3.3, 4.3, 6"},{"comment":"The centerline baseline comparison may be unfair. The authors state that DPM, PVD, and DiT-3D were 'adapted' into conditional models 'by following AortaDiff's conditioning strategy,' but no hyperparameter tuning, training duration, or convergence evidence is provided for these baselines. The reported gaps are extremely large (e.g., overall CD 0.33 mm for AortaDiff versus 2.39 mm for DPM on VMR), which raises the concern that the baselines were not trained to their potential, rather than that AortaDiff is intrinsically superior. Please report the baseline hyperparameters, training iterations, learning-rate schedules, and per-seed variance, or use published conditional variants with their own recommended settings.","section":"4.2, Table 1"},{"comment":"The number of centerline points (16) was selected through a parameter study conducted on all test cases from both VMR and AVT. This is test-set-based model selection and makes the reported metrics optimistic: the method and the evaluation set are no longer independent. The centerline point count should be chosen on a held-out validation split or justified through an independent criterion, and the final test metrics should be computed only after fixing that hyperparameter.","section":"Appendix Section 4, Table A1"},{"comment":"The CFD validation is exclusively qualitative. The paper claims 'a high degree of consistency' and 'comparable' velocity, pressure, and WSS fields between AortaDiff and ground-truth meshes, but no quantitative hemodynamic metrics are reported. Since 'CFD-compatible meshes' is a core contribution, the authors should provide numerical comparisons such as pressure-drop error, WSS error, or flow-rate distribution differences across branches. If such metrics are not available, the CFD claim should be softened to 'successful simulation execution' rather than geometric and hemodynamic equivalence.","section":"4.5, Figure 8"}],"minor_comments":[{"comment":"The phrase 'minimal dependency on large labeled datasets' should be qualified to make clear that the dependency applies to aorta-specific annotations; as written, it conflicts with the use of ScribblePrompt, a model trained on large-scale medical imaging data.","section":"Abstract and Section 1"},{"comment":"ScribblePrompt is described as 'a SAM-based model,' but ScribblePrompt is not a variant of SAM in the original literature; please verify the description and cite the underlying architecture correctly.","section":"3.3"},{"comment":"The choice of 16 centerline points is introduced in the main text as 'empirical' and only justified later in Appendix Section 4; a forward reference to the parameter study would help the reader.","section":"3.1"},{"comment":"The definitions of chamfer distance, Hausdorff distance, and Earth mover's distance are given by references only; a one-line formula or a precise reference to the exact definitions used would improve reproducibility.","section":"4.2"},{"comment":"The 'Ad-Hoc Expert Feedback' section is anecdotal and not a substitute for the quantitative evaluation presented elsewhere; if retained, it should be clearly labeled as an informal case report rather than a validation result.","section":"5"},{"comment":"Reference [15] contains a malformed DOI ('arXiv.11474' instead of a full arXiv identifier); please correct it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core diffusion component is quite thin relative to the claimed contribution: the proposed model outputs only 16 points per branch, and the surface quality is effectively determined by ScribblePrompt and NURBS. The authors should be asked to reframe the 'minimal dependency' and 'end-to-end surface generation' claims, and to provide an ablation that isolates the CDM's contribution. If the authors can do that, the paper could be acceptable; without it, the headline claims overstate what is actually demonstrated. The test-set-based selection of the centerline point count and the lack of quantitative CFD metrics are also important to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the engineering is real and the reported gains are large, but the headline claim about minimal dependency on labeled data is misleading. The segmentation stage carries the mesh quality, and that stage is a large pretrained model.\n\nWhat's genuinely new here is the decomposition: a volume-conditioned diffusion model that generates coarse centerline points (16 per branch) via point-feature lookup from a ViT encoder, and the automated use of those points as prompts to segment orthogonal slices, then NURBS fitting into a CFD-ready surface. That's a sensible way to sidestep full point cloud generation with scarce data, and the numbers back it up: centerline CD drops to 0.33 mm on VMR and 0.50 mm on AVT, far below the adapted baselines, and the final meshes beat marching-cubes/surface-nets baselines by roughly 2x in CD. The CFD results are qualitative, but the flow fields look plausible and the mesh is watertight.\n\nThe main soft spot, which the stress-test note correctly identifies, is the framing. The paper claims 'minimal dependency on large labeled datasets' and an end-to-end pipeline. But the contours that determine the final surface come from ScribblePrompt, a SAM-based model explicitly trained on large-scale medical data. The CDM only produces 16 coarse points per branch. The authors admit this themselves in the limitations: the centerlines are 'not perfectly centered' and the segmentation module 'mitigates the major impact.' If ScribblePrompt were replaced with a non-pretrained segmenter, the mesh advantage would likely collapse. So the contribution is a strong integration of an external segmentation prior, not a data-efficient generative model for surfaces.\n\nOther issues, in decreasing order: (1) The baseline centerline models were adapted by the authors with no evidence of fair tuning. (2) The 16-point centerline resolution was picked after a parameter study on the full test set, so the reported metrics carry selection bias. (3) The CFD validation is purely qualitative; a couple of quantitative hemodynamic metrics (e.g., flow rate or WSS error) would help. (4) No code or data released. (5) The overlap with the authors' own arXiv papers [14,15] is not clearly delineated, so the incremental novelty is hard to assess.\n\nNone of this sinks the core pipeline. The paper deserves a serious referee, but a referee should demand a fair baseline comparison, an ablation where the segmentation module is varied, and clearer positioning relative to the authors' prior work. I'd send it back for major revision.","headline":"The pipeline works and the numbers are strong, but the 'minimal dependency on large labeled datasets' claim leans on ScribblePrompt's pretrained priors, not on the diffusion model.","tokens_in":26410,"tokens_out":3142,"would_cite":false,"duration_ms":32109,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a volume-conditioned diffusion model can generate aortic centerlines from CT/MRI volumes, and that using those centerlines as automatic prompts for a pretrained segmenter plus NURBS fitting yields CFD-compatible…","keywords":["conditional diffusion model","aorta mesh generation","centerline generation","NURBS surface fitting","computational fluid dynamics","CT/MRI segmentation","multi-branch vessel modeling","volume-guided generation"],"falsifier":"Run the identical AortaDiff pipeline on the AVT volumes but replace ScribblePrompt with a non-trainable intensity-threshold or region-growing contour extractor driven by the same centerline prompts, then recompute final mesh Chamfer distance and Hausdorff distance against the ground-truth meshes; if the errors remain near 0.39 mm CD, the diffusion model is carrying the pipeline, whereas a large jump would show that the claimed end-to-end accuracy is inherited from the external segmenter.","tokens_in":25450,"feed_emoji":"🫀","tokens_out":4975,"duration_ms":54605,"temperature":0.7,"pith_summary":"AortaDiff claims to take a raw CT or MRI volume and output a smooth, watertight, simulation-ready 3D aorta mesh without manual contouring and with only tens of labeled training cases. The trick is to avoid generating the full mesh or point cloud, and instead first generate the aortic centerline with a conditional diffusion model, then use each centerline point as an automatic prompt for a pretrained segmentation model to extract cross-sectional contours, and finally fit a NURBS surface through those contours. If the claim holds, cardiovascular researchers could replace hours of manual mesh construction with an automatic pipeline that also handles aneurysms and coarctation, producing meshes whose simulated blood flow fields closely match those from manually built ground-truth meshes.","feed_headline":"From CT scan to CFD-ready aorta mesh without manual tracing","feed_subtitle":"AortaDiff trains on 18 labeled aortas and beats segmentation baselines on 12 test cases.","key_machinery":"The central object is the centerline-as-a-1D-image: each aortic centerline is represented as 16 ordered points stored in a three-channel 1D image. A vision transformer extracts a feature volume from the input CT/MRI volume, and at each diffusion timestep the model looks up feature vectors at the current centerline point locations, concatenates them with the noisy 1D centerline latent, and uses a 1D U-Net to predict the added noise. This volume-guided conditional denoising lets the model generate patient-specific centerlines without large point-cloud training sets. The second load-bearing mechanism is prompt-based contour extraction: each generated centerline point defines an orthogonal slicing plane through the volume, and the point itself is used as a prompt for ScribblePrompt, a segmentation model pretrained on large-scale medical images, yielding lumen contours that are resampled, aligned with ICP, and fitted with a cubic NURBS surface.","core_discovery":"On the paper's own terms, the discovery is that a compact 1D representation—the centerline encoded as a 16-point 1D image with x, y, and z as channels—is enough of a target for a volume-conditioned diffusion model to capture the aorta's shape and branching topology, and that downstream contour extraction and NURBS fitting convert that sparse signal into a high-fidelity surface. Training on 18 human aorta volumes from the VMR dataset, the model generates centerlines with Chamfer distance 0.33 mm on VMR and 0.50 mm on the held-out AVT dataset, and final NURBS meshes with 0.28 mm and 0.39 mm Chamfer distance respectively, outperforming point-cloud diffusion baselines and supervised segmentation followed by marching cubes or surface nets.","pith_inferences":["The reported mesh accuracy is likely inherited in large part from ScribblePrompt's pretrained segmentation ability rather than from the diffusion model alone; a fair decomposition would swap in different segmenters on the same generated centerlines to quantify how much each stage contributes.","The same centerline-to-contour-to-NURBS recipe could extend to other tubular anatomies, such as coronary or cerebral arteries, where annotated mesh data is equally scarce.","The fixed 16-point centerline sampling is a strong inductive bias; an adaptive sampling scheme based on local curvature could handle longer or more tortuous vessels without the contour-interference failures observed at 25 points.","Because the output is a parametric NURBS surface, isogeometric analysis and statistical shape modeling are natural next steps, but the paper only demonstrates steady-state CFD, leaving pulsatile and fluid-structure interaction simulations as open tests."],"forward_implications":["Researchers can obtain CFD-compatible aortic meshes from CT/MRI volumes without manual contouring, since the pipeline is fully automatic from volume to surface.","The training requirement drops to tens of labeled volumes, making mesh generation practical for datasets where full mesh labels are scarce.","The generated NURBS meshes flow directly into CFD solvers, and simulated velocity, pressure, and wall shear stress fields on them match those computed on ground-truth meshes.","Pathological aortas, including aneurysm and coarctation cases, can be constructed as reliably as healthy ones, supporting in silico experiments for surgical planning.","The modular design allows centerline generation, segmentation, and surface fitting to be upgraded independently as better components become available."],"supporting_citations":[{"why":"ScribblePrompt provides the actual lumen segmentation on cross-sectional slices, making it load-bearing for contour extraction.","marker":"[66]"},{"why":"Denoising diffusion probabilistic models supply the generative framework that the centerline CDM builds on.","marker":"[25]"},{"why":"The vascular model repository provides the training volumes and ground-truth aorta meshes used in the main evaluation.","marker":"[64]"},{"why":"The AVT dataset provides the held-out CT test cases used to demonstrate cross-modality generalization.","marker":"[51]"},{"why":"The NURBS book supplies the curve and surface fitting method used to turn extracted contours into the final CFD-ready mesh.","marker":"[50]"},{"why":"The vision transformer encoder extracts the volumetric feature representation that conditions the diffusion model on the input image.","marker":"[13]"}],"fun_headline_variants":["Diffusion model turns CT/MRI into CFD-ready aorta meshes","AortaDiff: 18 labeled cases yield accurate aorta surfaces","Centerline-guided diffusion auto-generates aorta meshes from CT","AortaDiff: sparse centerline trick enables high-fidelity aorta meshes","From CT volume to CFD-grade aorta mesh with limited data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that ScribblePrompt, a segmentation model pretrained on large amounts of medical image data, will correctly outline the aorta on every cross-sectional slice when given only the diffusion-generated centerline point as a prompt; if that segmentation fails, the final mesh quality collapses regardless of how well the centerline was generated.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model turns CT/MRI into CFD-ready aorta meshes","AortaDiff: 18 labeled cases yield accurate aorta surfaces","Centerline-guided diffusion auto-generates aorta meshes from CT","AortaDiff: sparse centerline trick enables high-fidelity aorta meshes","From CT volume to CFD-grade aorta mesh with limited data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1403,"prompt_tokens":984,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":330}},"tokens_in":600,"tokens_out":419,"duration_ms":5398,"temperature":1.0,"reasoning_tokens":330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:41:00.306633+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical AortaDiff pipeline on the AVT volumes but replace ScribblePrompt with a non-trainable intensity-threshold or region-growing contour extractor driven by the same centerline prompts, then recompute final mesh Chamfer distance and Hausdorff distance against the ground-truth meshes; if the errors remain near 0.39 mm CD, the diffusion model is carrying the pipeline, whereas a large jump would show that the claimed end-to-end accuracy is inherited from the external segmenter.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ScribblePrompt provides the actual lumen segmentation on cross-sectional slices, making it load-bearing for contour extraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Denoising diffusion probabilistic models supply the generative framework that the centerline CDM builds on."},{"cited_title":"Montalt-Tordera, E","cited_arxiv_id":null,"evidence_quote":"The AVT dataset provides the held-out CT test cases used to demonstrate cross-modality generalization."},{"cited_title":"Piegl and W","cited_arxiv_id":null,"evidence_quote":"The NURBS book supplies the curve and surface fitting method used to turn extracted contours into the final CFD-ready mesh."},{"cited_title":"Dosovitskiy, L","cited_arxiv_id":null,"evidence_quote":"The vision transformer encoder extracts the volumetric feature representation that conditions the diffusion model on the input image."}],"review_version":1}