{"id":"e2207d7a-e21d-43b1-a5f4-33987caf662d","arxiv_id":"2508.19862","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A conditional mesh-to-mesh GAN with local KNN and global graph branches predicts future thoracic aortic aneurysm shape and diameter more accurately than baseline mesh networks on a private longitudinal dataset.","lead":"This paper trains a mesh-based neural network to predict how a patient's aortic aneurysm will change shape over time from a single baseline CT scan. It combines local surface detail with whole-vessel shape and reports lower prediction errors than comparison methods on a new 590-scan dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated non-rigid registration (Sec. 4.1) underpins the L1 objective and all reported vertex-based and diameter metrics; without a registration fidelity check, the claimed performance gains may be registration artifacts rather than anatomical growth.","rationale":"I agree with the reader's weakest_assumption and consider it the single most load-bearing concern. The method's objective function (L1 on registered vertices) and the primary clinical metric (MIS diameter) both operate in the registered space. If the non-rigid registration is inaccurate, the reported numbers do not reflect anatomical growth. The paper's own text in Sec. 5.2.3 explicitly ties the choice of L1 loss to the assumption of one-to-one correspondence, making the dependency explicit. The missing GCNConv+KCN row and lack of statistical significance tests are secondary: they could be fixed by adding one table row or a paired test without changing the interpretation, whereas a registration flaw invalidates the core measurements. Therefore, I recommend keeping the conditional verdict and requiring the proposed registration fidelity check before acceptance.","tokens_in":11329,"tokens_out":6191,"duration_ms":71245,"concrete_test":"Select a random subsample (e.g., 30 patients) from TAAMesh. For each patient's baseline and follow-up CT, manually place 10 anatomical landmarks (e.g., sinotubular junction, innominate artery origin, left subclavian artery origin) on the original meshes. After the identical IRTK non-rigid registration pipeline, measure the Euclidean residual between corresponding landmarks in template space. Report mean and 95th percentile. Additionally, compute the MIS diameter on the original mesh and on the registered mesh for the same scan; if the mean absolute difference is comparable to the reported MCMeshGAN vs. best-baseline MIS error difference (2.887 vs 3.385 mm, Table 1), then the clinical endpoint is dominated by registration distortion. If residuals are <0.5 mm and diameter changes <0.5 mm, the registration concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MCMeshGAN outperforms baselines in geometric accuracy and MIS diameter rests on the preprocessing step in Sec. 4.1 that non-rigidly registers every CT-derived mesh to a common template to obtain 10,000 vertices with one-to-one correspondence. After this step, the L1 loss (Sec. 3.5) and per-vertex MAE are computed in registered (template) coordinates, and the MIS diameter is computed on the registered meshes. The paper provides no validation of this registration: no landmark error, no comparison of MIS diameter or centerline length before vs. after non-rigid registration, no residual evaluation. If the registration warps the aneurysm sac (e.g., shrinking or expanding local regions), then (i) the 'anatomical growth' encoded in the ground-truth registered meshes is partly a registration artifact, and (ii) the diameter measurements used as the clinically meaningful endpoint may be biased. The paper explicitly relies on the registration to justify why L1 loss outperforms CD loss: Sec. 5.2.3 states 'L1 loss ... tends to establish a one-to-one correspondence between vertices.' Thus the entire architecture and evaluation are coupled to this unvalidated assumption. Although all baseline methods are evaluated with the same registration, the extrapolation to real patient anatomy and 'clinically important diameter estimation' is not justified. This is the weakest load-bearing assumption because if it fails, the quantitative conclusions in Tables 1-2 are meaningless, regardless of statistical considerations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MCMeshGAN, a multimodal conditional mesh-to-mesh GAN for predicting thoracic aortic aneurysm growth from 3D surface meshes and clinical metadata (age, sex, time interval). The method combines a global GCN branch and a local KNN-based convolutional branch, conditioned on ordinal binary encodings of patient attributes. The authors also present TAAMesh, a private longitudinal dataset of 590 CT-derived meshes from 208 patients. On a held-out test set, MCMeshGAN reports the lowest MAE, CD, HD, and MIS diameter error among five adapted baselines, and ablations study the contributions of backbone, conditions, and losses. Source code is publicly released.","tokens_in":11746,"tokens_out":5029,"duration_ms":51307,"significance":"If the results hold, this is a useful step toward 3D mesh-to-mesh generative modeling of aneurysm progression, an area that is underexplored compared with 2D image or feature-based approaches. The public release of code and the introduction of a longitudinal TAA mesh dataset are concrete contributions. The dual-branch architecture that balances global structure and local detail is reasonable and well motivated. However, the strength of the empirical claims is currently conditional: the evaluation depends on an unvalidated non-rigid registration, one key baseline result is missing, and no statistical significance testing is provided. With these issues addressed, the paper would be a solid contribution to the medical image computing community.","major_comments":[{"comment":"The entire evaluation pipeline relies on IRTK non-rigid surface registration to establish one-to-one vertex correspondence across patients and time points. All reported metrics (MAE, CD, HD, MIS error) and the L1 loss are computed on registered meshes. The manuscript provides no validation of this registration: no landmark or centerline error, no comparison of MIS diameter or anatomy before vs. after non-rigid registration, and no residual analysis. If the registration locally warps the aneurysm sac, the reported accuracy numbers, as well as the argument that L1 loss outperforms CD loss because of one-to-one correspondence (Sec. 5.2.3), could reflect registration artifacts rather than biological growth. The authors should add a registration-fidelity validation and/or a sensitivity analysis showing that the main conclusions are robust to the registration choice.","section":"Sec. 4.1 and Sec. 5.2.3"},{"comment":"In Table 1, the row 'GCNConv [8] GCN+Condition+KCN' is blank (all entries are '-'), while the text states that 'the KCN branch significantly enhances all GCN frameworks, particularly GCNConv and GCN2Conv' and that 'GCNConv obtains the best improvements from KCN.' Since GCNConv is the GCN backbone used in MCMeshGAN, this result is essential support for the central design choice. Please report the result, or explicitly explain why it is missing (e.g., training failure) and adjust the claim accordingly.","section":"Table 1, Sec. 5.1"},{"comment":"The central claim that MCMeshGAN 'consistently outperforms' baselines is based on point estimates without statistical inference. The reported standard deviations are large relative to the differences (e.g., MAE 1.285±0.993 vs. 1.605±3.516 for GCN2Conv+KCN; CD 1.831±1.251 vs. 2.228±2.457), so the superiority is not established. No paired tests, confidence intervals, or effect sizes are provided. MIS error is reported only as a mean, without standard deviation or CI. The authors should perform per-patient paired statistical tests (e.g., Wilcoxon signed-rank) or bootstrap CIs for all metrics, including MIS error.","section":"Sec. 5.1, Table 1"},{"comment":"The evaluation is conducted on a single private dataset with no external validation. The abstract and conclusion refer to 'clinically deployable prediction' and 'generalization,' but the limitations section correctly notes that the model was only tested on this private dataset. Given the single-center data and moderate sample size, the clinical and generalization claims are overstated. At a minimum, the wording should be tempered; ideally, an additional validation experiment (nested cross-validation, leave-one-center-out, or evaluation on a public AAA dataset) would strengthen the support for generalization.","section":"Sec. 4.1 and Conclusion"}],"minor_comments":[{"comment":"The ordinal binary encoding of the time interval is not fully specified: it is unclear how negative intervals beyond -50 or positive intervals beyond +50 are handled, and what happens exactly at the boundary index 50. Please clarify the encoding with a precise formula.","section":"Sec. 3.2, Fig. 2"},{"comment":"The table header appears twice (the row 'Methods' is repeated). Please merge into a single header and use clear grouping for the '+Condition' and '+Condition+KCN' blocks.","section":"Table 1"},{"comment":"The subscripts 'FN ei' and 'FCenter' are unclear; the notation for the neighbor and center convolution functions should be defined consistently.","section":"Sec. 3.3, Eq. (4)"},{"comment":"Typo: 'essenstial' should be 'essential'.","section":"Sec. 5.1"},{"comment":"The text references 'Section IV-A' but should refer to Section 4.1.","section":"Sec. 5.2.3"},{"comment":"The claim that age has a greater impact than sex because of its larger dynamic range is speculative. Consider reporting feature importance or performing an interaction analysis to support this conclusion.","section":"Sec. 5.2.1"},{"comment":"The KNN neighborhood size K is set to 8 without sensitivity analysis. A small grid search over K would improve confidence in the architecture's robustness.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper has solid technical components and a new dataset, but the load-bearing issues (unvalidated registration, missing GCNConv+KCN result, and lack of statistical testing) need to be addressed before publication. The public code and dataset are commendable. I would welcome a revised version that provides the missing validation and tempers overclaims regarding clinical deployment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look: it is the first mesh-to-mesh conditional GAN I know of for aneurysm growth, and the dual-branch KCN+GCN design is a sensible way to handle the local-vs-global tension that plain GCNs have. The authors also built and released a new longitudinal TAA mesh dataset (590 records, 208 patients) and published code. The held-out test comparison against multiple baselines, with ablations for the condition branch and losses, is mostly clean: MCMeshGAN wins on MAE, CD, HD and MIS diameter by a decent margin, and adding KCN to other backbones generally helps. That is a real empirical contribution.\n\nBut the stress-test concern lands. The entire evaluation pipeline routes through the IRTK non-rigid surface registration (Sec. 4.1) that puts every mesh into a common 10,000-vertex correspondence. The L1 loss, the per-vertex MAE, and the MIS diameter are all computed on registered meshes. There is no validation of that registration: no landmark or residual error, no comparison of diameter or centerline length before vs. after warping. If the registration shrinks or expands local sac regions, then the 'ground truth' growth is partly a registration artifact, and the diameter numbers—the clinical endpoint—could be biased. The authors themselves lean on the registration to explain why L1 beats CD (Sec. 5.2.3). Because all baselines share the same registration, the relative ranking is less exposed, but the absolute accuracy and the clinical interpretation are not. This is a load-bearing assumption and needs a dedicated validation section (e.g., registration fidelity metrics, or measuring MIS on unregistered vs. registered meshes).\n\nOther soft spots are smaller: the GCNConv+KCN row in Table 1 is missing, which is conspicuous because MCMeshGAN is exactly that combination; there are no significance tests; MIS error has no standard deviation; and the dataset is single-center and private, so external validation remains future work (the authors say so).\n\nOverall, I would send this to peer review. The architecture and dataset merit referee time. But I would ask for registration validation and the missing baseline before acceptance.","headline":"A genuinely new application and dataset; the headline numbers all run through an unvalidated non-rigid registration step, so the clinical claims need a registration sanity check.","tokens_in":12196,"tokens_out":2677,"would_cite":true,"duration_ms":26261,"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":"MCMeshGAN claims that a mesh-to-mesh GAN conditioned on age, sex, and target time interval predicts future thoracic aortic aneurysm geometry more accurately than five graph and point-cloud baselines.","keywords":["aneurysm growth prediction","mesh-to-mesh translation","conditional GAN","graph convolutional network","KNN convolution","thoracic aortic aneurysm","longitudinal CT","personalized medicine"],"falsifier":"Run the trained model on an independent longitudinal thoracic aortic aneurysm dataset and compare predicted meshes against ground truth using a correspondence-free metric such as surface Dice or clinically reported MIS diameter; if the reported accuracy drops to baseline levels, the claimed gains are likely artifacts of the registration. Alternatively, validate the non-rigid registration by checking whether corresponding vertices track identifiable anatomical landmarks.","tokens_in":11295,"feed_emoji":"🫀","tokens_out":5872,"duration_ms":57837,"temperature":0.7,"pith_summary":"The paper tries to establish that a GAN operating directly on 3D surface meshes can predict how a thoracic aortic aneurysm will deform over months to years. The proposed MCMeshGAN combines a local K-nearest-neighbour convolution branch with a global graph-convolution branch, and conditions the generated mesh on patient age, sex, and the requested time interval. The authors built a longitudinal dataset of 590 CT-derived meshes from 208 patients and report that MCMeshGAN beats five adapted baselines on per-vertex error, Chamfer distance, Hausdorff distance, and the clinically used maximum-inscribed-sphere diameter. If correct, the approach would let a clinician start from one baseline scan and generate an estimated future aneurysm surface at any chosen time interval, supporting personalized surveillance scheduling.","feed_headline":"Mesh GAN predicts aneurysm growth in 3D","feed_subtitle":"Age, sex, and time interval condition a mesh-to-mesh model that beats five baselines on geometric and diameter error.","key_machinery":"The mechanism is a three-branch generator: a condition branch that encodes age, sex, and time interval into 100-dimensional ordinal binary vectors; a local KCN branch that finds K nearest neighbours in a CNN feature space and convolves over each vertex and its neighbours; and a global GCN branch that applies normalized graph-Laplacian message passing over four layers. A fusion module combines the features and a discriminator supplies adversarial loss. The KCN branch is the load-bearing novelty: it compensates for GCN over-smoothing by explicitly modelling fine local variation, which is where aneurysm diameter changes are expressed.","core_discovery":"The central claim is that mesh-to-mesh translation with multimodal conditioning works for aneurysm growth: given a baseline thoracic aortic aneurysm mesh, patient age and sex, and a target time interval (positive for future, negative for past), MCMeshGAN generates the patient's aneurysm mesh at that time. On the test set, it reports a mean absolute error of 1.285 ± 0.993 mm, Chamfer distance of 1.831 ± 1.251 mm, Hausdorff distance of 9.815 ± 5.131 mm, and MIS diameter error of 2.887 mm, all lower than the five baselines, including when the local branch is added to them. The paper interprets this as evidence that the global graph branch preserves overall anatomy while the KNN-based local bran","pith_inferences":["The paper only evaluates on its own private dataset, so its own stated limitation is that generalization to other scanners, protocols, and populations is unestablished; an external dataset test is the natural next check.","All geometric metrics are computed on corresponding vertices, so a registration-free metric (e.g., surface Dice or clinically reported diameters) would test whether the gains reflect anatomy rather than the alignment procedure.","The 100-dimensional ordinal time encoding implies a bounded time horizon and discrete interval granularity; a continuous time embedding might extend the model to arbitrary follow-up lengths.","No comparison is made to inter-observer or scan-rescan variability of MIS diameter; whether 2.887 mm error is clinically acceptable depends on how it compares to measurement noise in routine practice."],"forward_implications":["If the reported accuracy holds, a single baseline CT mesh plus age, sex, and a requested interval yields a predicted 3D aneurysm surface, enabling individualized surveillance intervals instead of uniform annual imaging.","The time-interval conditioning supports both prospective prediction (positive intervals) and retrospective reconstruction (negative intervals), so the same model can be trained and evaluated on all longitudinal pairs in the dataset.","The ablation results indicate that the KCN branch improves every graph-convolution baseline, suggesting the local-detail module is a reusable component for mesh generation beyond aneurysm prediction.","The lowest MIS diameter error (2.887 mm) is the result most tied to clinical decisions, since maximum diameter thresholds guide surgical referral."],"supporting_citations":[{"why":"Defines the graph-convolution message-passing formula that the global branch is built from and that the GCNConv baseline uses.","marker":"[8]"},{"why":"Supplies the automated CT segmentation model used to convert scans into aortic masks, the first step of the mesh pipeline.","marker":"[28]"},{"why":"Baseline GCN variant with residual and identity mappings, used to test whether countering over-smoothing alone matches the KCN branch.","marker":"[22]"},{"why":"Baseline spectral graph convolution, the strongest non-KCN performer, used to isolate the contribution of local detail.","marker":"[23]"},{"why":"Baseline MLP/GCN hybrid used to show condition-branch gains are not GCN-specific.","marker":"[24]"},{"why":"Hierarchical point-cloud baseline used as a non-graph deep learning comparison.","marker":"[29]"},{"why":"Graph-learning library used to implement all GCN baselines under the same condition branch for fair comparison.","marker":"[30]"},{"why":"Prior deep-learning aneurysm-growth predictor on 2D patches, the approach this work extends to direct 3D mesh prediction.","marker":"[26]"}],"fun_headline_variants":["Mesh GAN predicts aneurysm growth in 3D from clinical data","Personalized 3D aneurysm growth predicted by mesh-to-mesh GAN","Aneurysm growth: multimodal GAN outperforms five baselines","MCMeshGAN: predicting aortic aneurysm growth with age and sex","New GAN generates 3D aneurysm shapes for future timepoints"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central premise is that the non-rigid surface registration used in preprocessing establishes a true one-to-one correspondence between vertices across scans; the paper does not validate that correspondence, so if registration misaligns vertices, the reported per-vertex errors reflect registration artifacts rather than anatomical growth.","fun_headline_variants_meta":{"raw":{"variants":["Mesh GAN predicts aneurysm growth in 3D from clinical data","Personalized 3D aneurysm growth predicted by mesh-to-mesh GAN","Aneurysm growth: multimodal GAN outperforms five baselines","MCMeshGAN: predicting aortic aneurysm growth with age and sex","New GAN generates 3D aneurysm shapes for future timepoints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1757,"prompt_tokens":794,"completion_tokens":963,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":870}},"tokens_in":538,"tokens_out":963,"duration_ms":10830,"temperature":1.0,"reasoning_tokens":870,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:22:36.234877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on an independent longitudinal thoracic aortic aneurysm dataset and compare predicted meshes against ground truth using a correspondence-free metric such as surface Dice or clinically reported MIS diameter; if the reported accuracy drops to baseline levels, the claimed gains are likely artifacts of the registration. Alternatively, validate the non-rigid registration by checking whether corresponding vertices track identifiable anatomical landmarks.","supporting_citations":[{"cited_title":"Semi-supervised classification with graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Defines the graph-convolution message-passing formula that the global branch is built from and that the GCNConv baseline uses."},{"cited_title":"TotalSegmentator: robust segmentation of 104 anatomic structures in CT images,","cited_arxiv_id":null,"evidence_quote":"Supplies the automated CT segmentation model used to convert scans into aortic masks, the first step of the mesh pipeline."},{"cited_title":"Simple and deep graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Baseline GCN variant with residual and identity mappings, used to test whether countering over-smoothing alone matches the KCN branch."},{"cited_title":"Simple spectral graph convolution,","cited_arxiv_id":null,"evidence_quote":"Baseline spectral graph convolution, the strongest non-KCN performer, used to isolate the contribution of local detail."},{"cited_title":"Graph neural networks are inherently good generalizers: Insights by bridging GNNs and MLPs,","cited_arxiv_id":null,"evidence_quote":"Baseline MLP/GCN hybrid used to show condition-branch gains are not GCN-specific."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space,","cited_arxiv_id":null,"evidence_quote":"Hierarchical point-cloud baseline used as a non-graph deep learning comparison."},{"cited_title":"Deep learning on multiphysical features and hemodynamic modeling for abdominal aortic aneurysm growth prediction,","cited_arxiv_id":null,"evidence_quote":"Prior deep-learning aneurysm-growth predictor on 2D patches, the approach this work extends to direct 3D mesh prediction."}],"review_version":1}