{"id":"b34435c2-28f3-4589-a354-63d11bea9464","arxiv_id":"2505.10407","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AneuG generates realistic intracranial aneurysm meshes and parent vessels using a two-stage VAE with graph harmonic deformation encoding and differentiable conditioning on morphological markers.","lead":"A new two-stage VAE generates synthetic intracranial aneurysm meshes with parent vessels, and can be conditioned on clinical shape measurements. It aims to provide realistic training data for blood-flow prediction, where real aneurysm imaging data is scarce.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Conditioning accuracy is circular: MMC is used both as the training loss and as the evaluation metric, so Table 2 does not validate clinically meaningful morphological controllability.","rationale":"After reading the paper, the central claim is twofold: AneuG is the first generative model that jointly models the aneurysm complex and parent vessels, and it can generate shapes with specified clinically significant morphological markers. The first part is supported by the architectural design and by comparisons in Table 1, though with a small dataset and no error bars. The second part is where the paper's strongest practical contribution lies, and it is exactly where the evaluation is weakest. The MMC is used inside the training loop (Lcond) and then reused as the CA metric. This creates a closed loop: the model is trained to satisfy a differentiable proxy, and the evaluation reports how well it satisfies that same proxy. This is not an internal inconsistency of the method; the method may still work. But the claim concerns clinically significant parameters, so the proxy needs an external anchor. The paper does not provide one, and the unusual LI definition further indicates that the MMC's marker definitions may not match clinical conventions. The reader identified the same weakest assumption. I agree with the reader's conditional verdict: the method is promising and the code is promised, but the central controllability claim is not yet demonstrated. The concrete test above would settle whether the concern lands: if independent measurements match requested values, the circularity is benign; if not, the paper's headline contribution is overstated. Thus the reader's CONDITIONAL verdict remains appropriate, and no adjustment is needed.","tokens_in":6666,"tokens_out":5924,"duration_ms":55409,"concrete_test":"Generate a fixed set of 50 AneuG meshes with prescribed AR, NW, LI, and V values. Measure the four markers on these meshes with an independent implementation that does not share the GHD registration or MMC code, for example 3D Slicer's aneurysm analysis module or a second software package, and have a neuroradiologist (blinded to the requested values) measure the same markers on renderings. Compute the relative L2 error between requested and independently measured values. If the independent error is materially larger than the CA reported in Table 2, or if absolute deviations exceed clinically meaningful thresholds, the conditioning claim is not established and the paper should be revised to either validate MMC against clinical measurements or soften the claim. As a complementary sanity check, run the MMC on the real AneuX meshes and compare to any available clinical marker values.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that AneuG generates IA meshes with specified clinically relevant morphological markers. The evidence for this claim rests on the conditioning accuracy (CA) rows in Table 2. However, CA is computed with the same differentiable Morphological Marker Calculator (MMC) that is optimized during stage-I training via Lcond (Sections 2.1 and 2.3). The MMC measures neck width, aspect ratio, lobulation index, and dome volume on meshes that share the canonical GHD connectivity, and the paper provides no validation that these differentiable computations reproduce clinical or radiological measurement protocols. Consequently, the reported CA values (relative L2 errors between about 1.8% and 12.9%) only show that the generator can match the differentiable proxy it was trained against; they do not show that an independent observer would recover the requested markers. The concern is reinforced by the unusual definition of lobulation index as dome surface area divided by dome volume, which is not a fraction and differs from common clinical lobulation descriptors. A secondary issue is that all evaluations appear to be on the same 116-case cohort without an explicit held-out split, which could inflate both fidelity metrics and CA, but the metric circularity is the load-bearing problem because it targets the controllability claim directly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AneuG, a two-stage VAE-based generative model for 3D intracranial aneurysm (IA) meshes. Stage I encodes the aneurysm complex into Graph Harmonic Deformation (GHD) tokens and trains a conditional VAE with a differentiable Morphological Marker Calculator (MMC) as a conditioning loss plus a Morphing Energy Alignment (MEA) constraint; Stage II generates parent-vessel centrelines from Fourier coefficients conditioned on the generated aneurysm complex. The authors compare unconditional sample quality against PCA and a diffusion baseline using FPD, KPD, TMD, and Chamfer distances, and report conditioning accuracy for aspect ratio, neck width, lobulation index, and dome volume in Table 2. They also provide qualitative demonstrations of conditional morphing and preliminary CFD results.","tokens_in":6952,"tokens_out":4332,"duration_ms":40366,"significance":"If the central claims held, AneuG would be a valuable contribution: source code is promised, GHD tokenization is a sensible low-dimensional representation for meshes with fixed connectivity, and the two-stage design that ties parent vessels to the aneurysm complex addresses a real gap in existing IA shape models. The unconditional results in Table 1 are suggestive, and the application to controlled flow-simulation cohorts is exactly what the community needs. However, the headline controllability claim in Table 2 is currently validated only against the same differentiable marker calculator used in the training objective, and the experimental section does not describe a held-out split or error bars. These omissions make the contribution defensible but not fully established; the architecture and the unconditional-fidelity evidence are the strongest parts.","major_comments":[{"comment":"The conditioning-accuracy metric is circular. The differentiable Morphological Marker Calculator (MMC) computes NW, AR, LI, and V and is minimized as Lcond during stage-I training; the same MMC is then used to compute the CA rows in Table 2. Because the evaluation metric is the same differentiable proxy as the training objective, the reported CA values (1.80-12.91%) demonstrate that the generator matches the calculator, not that a clinician or an independent measurement protocol would recover the requested marker values. This directly targets the paper's central controllability claim. The authors should validate MMC against clinical or radiological measurements, or against manual annotations on a held-out set, and report CA with an independent measurement protocol.","section":"Section 2.3 and Table 2"},{"comment":"No train/validation/test split is described, and all reported numbers appear to be computed on the same 116-case cohort used for training. Without a held-out split or cross-validation, the reconstruction and fidelity metrics can be optimistically biased, and the conditioning numbers cannot be interpreted as generalization performance. Please report a fixed split (or repeated cross-validation) and per-seed standard deviations, especially since the dataset is small and the VAE is trained on the same cases whose distributions are evaluated in Tables 1 and 2.","section":"Section 3, Tables 1 and 2"},{"comment":"The lobulation index (LI) is defined as dome surface area divided by dome volume. This quantity has units of inverse length and is not a fraction, nor is it a recognized clinical descriptor of lobulation; the paper does not show that it correlates with the presence of a daughter sac or with established lobulation features used in the clinical literature. Since Table 2 and the qualitative 'daughter sac' demonstration in Fig. 2f rely on this marker, the authors should either adopt a standard lobulation measure (e.g., number or depth of blebs) and validate it, or clearly reposition LI as a purely geometric summary and temper the clinical claims made through it.","section":"Section 2.3"},{"comment":"The claim that AneuG 'simultaneously models the aneurysm complex and its parent vessels' is not quantitatively evaluated. Table 1 reports only metrics on the aneurysm-complex shape (FPD/KPD/TMD/CDv/CDn), and there is no evaluation of generated centreline fidelity, parent-vessel radius profiles, or the realism of the merged two-vessel mesh. As written, the second-stage contribution rests on a single qualitative figure. Please add quantitative vessel-level metrics or explicitly restrict the validation claim to the aneurysm complex.","section":"Section 2.4 and Fig. 2"}],"minor_comments":[{"comment":"There is a typo: 'the the largest publicly available IA dataset' should read 'the largest publicly available IA dataset'.","section":"Section 3, Dataset"},{"comment":"The notation Rk(vk) is garbled in the text, and the roles of the rotation matrix, the sine basis functions, and the branch index k should be spelled out more clearly.","section":"Equation (2)"},{"comment":"The GHD mode truncation n and the number of Fourier modes are free parameters that affect the reported reconstruction numbers; please state their values in the main text or appendix.","section":"Section 2.2"},{"comment":"The check-mark columns for MEA and MMC are ambiguous across the different condition groups; label each configuration explicitly, for example 'MEA on/off' and 'MMC on/off' under each marker group.","section":"Table 2"},{"comment":"Some reference entries use 'et al.' without full author lists, and reference [19] appears unused in the text; please check citation usage and complete the bibliographic details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a machine-learning-for-clinical-shape-generation venue. The main risk is the circularity of the conditioning evaluation: if the authors can validate MMC against an independent measurement or add human-annotated marker evaluation, the paper could become acceptable. I would not reject based on disagreement with existing approaches; the concern is methodological and fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution — a two-stage VAE that generates aneurysm complexes via GHD tokens and then conditions parent vessels on them, with differentiable control over clinical morphology markers. The combination is new as far as I can tell, and the code is public. But the key evidence for the headline controllability claim is weaker than the prose suggests, because the accuracy metric is computed with the same differentiable calculator (MMC) that appears in the training loss. That is a self-consistency check, not an independent validation.\n\nWhat is good: GHD encoding is a smart fit for the small-data regime; the MEA loss specifically targets posterior collapse in conditional training; and stage II's tangent constraint at the vessel-aneurysm junction is a sensible way to avoid unrealistic merges. The flow simulation demonstration is a nice application, and the qualitative trend (higher AR -> larger low-WSS regions) matches existing hemodynamics literature. Those are genuine strengths.\n\nThe soft spots are proportionate but real. The circularity is load-bearing: Table 2's CA numbers (1.8–12.9% relative L2) only show the generator can hit targets as measured by MMC. If MMC's neck width, dome volume, or the odd \"lobulation index\" (surface area divided by volume, which is not a fraction) diverge from how clinicians measure these markers, the control shown may not transfer to real anatomy. No evidence is provided that MMC reproduces clinical measurement protocols. A separate validation against manual or radiological measurements would fix this. Also, everything appears to be evaluated on the same 116-case cohort, with no explicit held-out split and no error bars, so the fidelity numbers (FPD, KPD, CD) are hard to interpret. The comparison set is thin: PCA and a generic 3D diffusion model, but no domain-specific IA generation baseline, though that may be because none exists.\n\nDespite these gaps, the method itself is coherent, the math is standard, and the code release lets others verify the pipeline. The circularity does not sink the paper; it means the most novel claim is not yet proven as strongly as stated. I would send it to peer review and ask for (1) independent validation of MMC against clinical measurements, (2) a train/test split with error bars, and (3) clarification of the lobulation index definition. If those are addressed, it is a solid contribution to the IA simulation and generative shape modelling community.","headline":"A genuinely novel two-stage IA mesh generator with morphological conditioning, but the controllability evidence is partly circular because the same differentiable marker calculator is used as both loss and test metric.","tokens_in":7396,"tokens_out":2573,"would_cite":true,"duration_ms":25623,"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":"AneuG is claimed to be the first deep generative model that jointly models intracranial aneurysm pouches and their parent vessels while allowing conditioning on clinically significant morphological markers.","keywords":["intracranial aneurysm","mesh generation","variational autoencoder","conditional generation","graph harmonic deformation","morphological markers","parent vessel modeling","hemodynamics"],"falsifier":"Take a set of AneuG-generated meshes conditioned on specific marker values, have a neuroradiologist or an independent software package measure neck width, aspect ratio, lobulation index, and dome volume using the standard clinical protocol, and compare those measurements with the requested values; a systematic discrepancy would show that the reported conditioning accuracy reflects the calculator's own conventions rather than clinical morphology.","tokens_in":6524,"feed_emoji":"🧠","tokens_out":5360,"duration_ms":50972,"temperature":0.7,"pith_summary":"The paper introduces AneuG, a two-stage variational-autoencoder generator for intracranial aneurysm meshes. Its central claim is that, for the first time, one model can generate the aneurysm pouch and its parent vessels together, conditioned on clinically relevant morphological values such as neck width, aspect ratio, lobulation index, and dome volume. This matters because large real-world aneurysm datasets are scarce, and controlled synthetic shape cohorts would allow training of fast blood-flow predictors and systematic studies of how shape features drive hemodynamics. The reported experiments, on 116 middle-cerebral-artery aneurysms, show the model beating a PCA statistical-shape baseline and a diffusion-based generator in fidelity and reconstruction error.","feed_headline":"AneuG shapes brain-aneurysm meshes to chosen clinical markers","feed_subtitle":"A two-stage VAE jointly models aneurysm pouch and parent vessels, enabling controlled synthetic cohorts for blood-flow studies.","key_machinery":"The load-bearing object is the Graph Harmonic Deformation (GHD) token: scalar coefficients applied to eigenvectors of the canonical mesh's cotangent Laplacian, which warp the canonical mesh into a target aneurysm complex. GHD gives a low-dimensional encoding that preserves node correspondence, so the same mesh can be differentiated through the Morphological Marker Calculator to compute neck width, aspect ratio, lobulation index, and dome volume. The Morphing Energy Alignment term then matches the distribution of rigidity and Laplacian-smoothness energies between real and generated shapes, countering posterior collapse under conditional training. Stage two uses Fourier-mode coefficients as centerline tokens and a tangent-regularization loss so generated vessels connect smoothly to the aneurysm complex.","core_discovery":"On its own terms, the paper establishes that a two-stage VAE can generate anatomically plausible intracranial aneurysm meshes whose clinically measured shape markers can be set by the user. Stage one encodes the aneurysm complex as Graph Harmonic Deformation tokens and learns a conditional latent space; a differentiable Morphological Marker Calculator feeds marker errors back into training, while Morphing Energy Alignment keeps generated shapes within the population's energy statistics. Stage two generates parent-vessel centrelines as Fourier-mode beams conditioned on the complex, matches tangent directions at the connection, and propagates the cross-section to form a complete mesh. With only 116 training shapes, AneuG reports lower Fréchet and kernel PointNet++ distances, higher diversity, and lower Chamfer distances than PCA and latent-diffusion baselines, and its conditioning accuracy improves when both marker conditioning and energy alignment are active. The authors also demonstrate controlled morphing of a real shape's aspect ratio and lobulation index, and show in flow simulations that higher aspect ratio enlarges low-wall-shear-stress regions while a daughter sac produces more complex flow.","pith_inferences":["A natural next test, not run in the paper, is to measure markers on AneuG outputs with an independent clinical tool or neuroradiologist; because the same differentiable calculator is used for both conditioning and evaluation, only such an external check can confirm that the controls are clinically meaningful.","The GHD representation is template-based, so the same two-stage recipe may transfer to other vessel territories or to organs where a canonical mesh and morphing energies can be defined; the paper does not claim this.","The comparison to a diffusion baseline is informative but narrow: the reported gap may reflect data efficiency of VAE plus GHD on 116 samples rather than a general advantage over diffusion generators, and a larger-scale comparison would clarify that.","With conditioning applied to a real shape's latent code, the model effectively offers counterfactual aneurysm generation, which could be used to probe rupture-risk markers such as daughter sacs while holding other geometry fixed."],"forward_implications":["If AneuG's central claim holds, researchers can generate large synthetic aneurysm cohorts with prescribed clinical marker statistics, removing a major data bottleneck for training real-time blood-flow prediction networks.","Because the latent code can be fixed while the marker condition is varied, single-marker effects on hemodynamics become directly testable; the paper's own simulations show this for aspect ratio and lobulation.","Joint generation of parent vessels produces complete CFD-ready meshes, avoiding the previous practice of merging pouches with idealized or unrelated healthy vessels.","The method's reported success with 116 training samples suggests that rare aneurysm locations, where large datasets do not exist, are a feasible target for the same approach.","Both proposed training components do measurable work: conditioning accuracy improves with the differentiable marker calculator, and conditional fidelity improves with morphing-energy alignment."],"supporting_citations":[{"why":"Supplies the Graph Harmonic Deform mesh-morphing method that produces the GHD tokens used in stage one.","marker":"[16]"},{"why":"The PCA statistical shape model serves as the biomedical shape-generation baseline AneuG is compared against.","marker":"[14]"},{"why":"The Michelangelo latent-diffusion generator is the deep-learning baseline for 3D shape generation.","marker":"[15]"},{"why":"Supplies the clinical definitions of neck width and aspect ratio that the Morphological Marker Calculator implements.","marker":"[5]"},{"why":"Provides the Fréchet and kernel PointNet++ distance metrics used to evaluate generation fidelity.","marker":"[13]"},{"why":"Defines Total Mutual Difference, the diversity metric reported for generated shapes.","marker":"[20]"},{"why":"Defines the as-rigid-as-possible rigidity energy used in the Morphing Energy Alignment loss.","marker":"[17]"},{"why":"Defines the Laplacian smoothness energy used as the second morphing-energy statistic in MEA.","marker":"[18]"}],"fun_headline_variants":["AneuG: control aneurysm mesh shapes via clinical markers","Two-stage VAE tailors aneurysm meshes to clinical traits","Generative model shapes aneurysms to clinical specs","AneuG lets you set aneurysm morphology in generated meshes","Synthetic aneurysm meshes with user-set morphology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the differentiable Morphological Marker Calculator's mesh-based measurements of neck width, aspect ratio, lobulation index, and dome volume agree with the clinical measurement protocols these markers come from; conditioning accuracy is computed with the same calculator that enforces the conditioning.","fun_headline_variants_meta":{"raw":{"variants":["AneuG: control aneurysm mesh shapes via clinical markers","Two-stage VAE tailors aneurysm meshes to clinical traits","Generative model shapes aneurysms to clinical specs","AneuG lets you set aneurysm morphology in generated meshes","Synthetic aneurysm meshes with user-set morphology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1303,"prompt_tokens":996,"completion_tokens":307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":612,"tokens_out":307,"duration_ms":3210,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:08:57.281191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of AneuG-generated meshes conditioned on specific marker values, have a neuroradiologist or an independent software package measure neck width, aspect ratio, lobulation index, and dome volume using the standard clinical protocol, and compare those measurements with the requested values; a systematic discrepancy would show that the reported conditioning accuracy reflects the calculator's own conventions rather than clinical morphology.","supporting_citations":[{"cited_title":"Ap- plied Sciences 12(1), 394 (2022)","cited_arxiv_id":null,"evidence_quote":"The PCA statistical shape model serves as the biomedical shape-generation baseline AneuG is compared against."},{"cited_title":"In: Advances in Neural Information Processing Systems 36, pp","cited_arxiv_id":null,"evidence_quote":"The Michelangelo latent-diffusion generator is the deep-learning baseline for 3D shape generation."},{"cited_title":"Surgical Neurology International 12, 506 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the clinical definitions of neck width and aspect ratio that the Morphological Marker Calculator implements."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp","cited_arxiv_id":null,"evidence_quote":"Defines Total Mutual Difference, the diversity metric reported for generated shapes."},{"cited_title":"In: Belyaev, A., Garland, M","cited_arxiv_id":null,"evidence_quote":"Defines the as-rigid-as-possible rigidity energy used in the Morphing Energy Alignment loss."},{"cited_title":"In: Proceedings of the 26th Annual Con- ference on Computer Graphics and Interactive Techniques (SIGGRAPH ’99), pp","cited_arxiv_id":null,"evidence_quote":"Defines the Laplacian smoothness energy used as the second morphing-energy statistic in MEA."}],"review_version":1}