{"id":"e1f04970-54ed-4011-a057-08d347765c14","arxiv_id":"2501.09046","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A benchmark of six geometric deep learning backends finds transformer-based LaB-GATr most accurate for predicting pressure-derived vFFR fields on patient-specific coronary artery meshes, with pressure drop as the best training target.","lead":"This study compares six geometric deep learning models as fast surrogates for blood-flow simulations that compute virtual FFR in coronary arteries. On synthetic arteries most models worked; on real patient anatomies a transformer model (LaB-GATr) was the most accurate, and predicting pressure drop instead of pressure directly gave the best results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patient-level data leakage is the load-bearing risk: 427 CFD runs come from 105 patients with no described patient-level split, so the claimed patient-specific generalization and LaB-GATr superiority may be inflated by anatomy memorization.","rationale":"I read the paper as an empirical benchmark, not a theoretical proof. Its central claims are: (1) on patient-specific data, LaB-GATr/transformers are uniquely strong; (2) pressure drop is the best training target. For (1) to be true, the evaluation must test generalization to unseen patients. The paper never specifies the patient-specific split. Since 427 simulations come from 105 patients with 3–5 runs each, random simulation-level splitting creates a serious leakage risk. This is the load-bearing flaw because every reported patient metric could be optimistic, and it interacts with the acknowledged small dataset (Section 4.4). I do not think the surface-constancy assumption in Section 2 is the weakest point: wall pressure is the quantity used for FFR, and cross-sectional pressure variation in coronary arteries is small relative to the trans-stenotic drop; in any case, the same assumption underlies the CFD practice the paper benchmarks. The paper is transparent about other limitations: LaB-GATr has fewer parameters than intended (Section 2.2), and the pressure-drop advantage on patient data is not directly tested (Section 4.4). The conclusion 'only transformer-based backends' is also somewhat stronger than Table 7, where DiffusionNet-S achieves accuracy 0.71 and correlation 0.72. These are secondary. A concrete patient-level split check would settle whether the central comparison is valid. I recommend maintaining the conditional verdict: accept only after authors provide the patient-level split and, ideally, code/data. If the split is simulation-random, the patient-specific claims should be substantially revised.","tokens_in":19519,"tokens_out":5775,"duration_ms":61298,"concrete_test":"Rerun the patient-specific experiments with a strict patient-level split: assign all CFD runs from the same reconstructed anatomy to the same fold (e.g., 5-fold cross-validation by geometry), then recompute per-point difference, approximation disparity, correlation, bias, and accuracy for Tables 6–7. If LaB-GATr's margin over DiffusionNet-S/MLP-L shrinks or its 0.79 accuracy drops toward 0.7, the reported patient-specific generalization is substantially leakage-driven. As a minimal check, request the exact assignment of the 427 runs to train/validation/test from the authors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that models trained on 427 patient-specific CFD runs generalize to unseen patient anatomies. The paper never states how these runs were split. Section 2.1.2 reports 105 reconstructed geometries and 3–5 steady CFD runs per geometry with slightly varied boundary conditions, yielding 427 simulations; Section 2.4 describes a train/validation/test split only for the 1,500 synthetic cases. If the patient-specific split was random at the simulation level, the same coronary anatomy appears in both training and test sets. Because all models receive geometry-derived input features (normals, geodesics, centerline radius, spectral descriptors, Table 2), a model can memorize a patient's anatomy and interpolate among its BC variants rather than learning a CFD surrogate for new anatomies. That would inflate every patient-specific metric in Tables 6–7 and could manufacture or exaggerate LaB-GATr's apparent advantage. Note also that DiffusionNet-S reaches 0.71 accuracy and 0.72 correlation, so the 'only transformers' claim is sensitive to which metric is considered. The authors' own limitation about dataset size (Section 4.4) makes leakage more plausible: 427 is not large, and 105 independent anatomies is the effective sample size for cross-patient generalization. This is a correctness risk, not a physics dispute; I do not see the surface-constancy assumption in Section 2 as the main threat, since wall pressure is the clinically relevant FFR quantity and cross-sectional pressure variation in coronaries is usually small.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a benchmark of six geometric deep learning backends (MLP, PointNet++, DiffusionNet, DeltaConv, LaB-VaTr, and LaB-GATr) for predicting pressure-related scalar fields on coronary artery surface meshes, using steady-state CFD solutions as ground truth. Two experiments are reported: a synthetic dataset of 1,500 left-coronary bifurcations, where models are trained to predict pressure, pressure drop, or vFFR directly; and a patient-specific dataset of 427 CFD simulations from 105 anatomies, where models are trained only on the pressure-drop output. The main claims are that pressure drop is the optimal learning variable, that transformer-based architectures (specifically LaB-GATr) are the only models achieving strong performance on patient-specific data, and that GDL surrogates can estimate vFFR in under two minutes per case. An ablation study on input features is included, identifying surface normals, geodesics, vessel radius, inflow, and inlet pressure as important.","tokens_in":19726,"tokens_out":8345,"duration_ms":77314,"significance":"If the central claims hold, the paper provides a useful empirical comparison for an application of practical interest: replacing expensive patient-specific CFD with learned surrogates for vFFR estimation. The benchmark is systematic in its control of input features, model sizes, and evaluation metrics, and the synthetic-dataset portion is large and well-defined. A notable strength is that all models are trained and evaluated on the same held-out CFD fields, making the comparison internally fair. The patient-specific results are, however, the load-bearing part of the paper, and they are currently undermined by an undocumented split and the absence of statistical testing. In addition, the selection of the pressure-drop output is not fully consistent with the reported synthetic-dataset metrics. These issues are fixable within the manuscript's scope, but they need to be addressed before the conclusions can be accepted.","major_comments":[{"comment":"The patient-specific dataset comprises 427 CFD simulations from 105 anatomies, with 3–5 runs per anatomy, but the manuscript never describes how these runs are split into training, validation, and test sets. If the split is performed at the simulation level, the same coronary anatomy appears in both training and test, and the geometry-derived input features listed in Table 2 (normals, geodesics, centerline radius, spectral descriptors) allow the model to memorize an anatomy and interpolate among its boundary-condition variants rather than to learn a surrogate that generalizes to unseen patients. Because the central claim is generalization to patient-specific anatomies, the split must be at the patient level, i.e., all CFD runs of a single geometry should be kept in the same fold. Please state the exact split (number of patients per fold) and, if the initial analysis was simulation-level, re-run the experiments with a patient-level split and report the revised Tables 6 and 7.","section":"Section 2.1.2 and Section 2.4"},{"comment":"The claim that 'transformer-based architectures were the only ones to achieve strong performance' is not supported by the reported metrics alone. DiffusionNet (S) achieves a correlation of 0.72 in Table 7, which is higher than LaB-GATr (S)'s 0.69, and its accuracy of 0.71 is close to LaB-GATr's 0.79; LaB-GATr is better on per-point difference and bias, but the differences are small relative to the reported standard deviations. No paired significance tests, confidence intervals, or per-patient/per-lesion analyses are provided, and pointwise metrics over meshes contain thousands of spatially correlated vertices, so visual inspection of Figure 6 is not a substitute. Please add paired statistical comparisons (e.g., bootstrap or Wilcoxon tests with the patient as the unit of analysis) and revise the 'only' claim in accordance with the results.","section":"Section 3.2, Tables 6 and 7"},{"comment":"The choice of pressure drop as the best output variable is not fully supported by the synthetic-dataset results. In Table 5, LaB-GATr trained to predict vFFR directly attains a per-point difference of 2.28E-04 and an approximation disparity of 5.40E-04 for the S variant, which is an order of magnitude better than the corresponding pressure-drop results in Table 4 (4.39E-03 and 4.97E-03). The paper's own Section 4.4 acknowledges that the superiority of pressure drop may not transfer to patient-specific data. The criterion used to select the output variable should be stated explicitly (e.g., average performance across all backends, training stability, or a defined composite metric) and applied consistently; as written, the selection procedure conflicts with the quantitative results for the best-performing model and weakens the third claimed contribution.","section":"Section 2.4.2 and Tables 4–5"}],"minor_comments":[{"comment":"The assumption that pressure is constant on the vessel cross-section justifies training exclusively on surface meshes. Since the target is wall pressure, this assumption is not fatal for the benchmark, but it would be helpful to add an empirical check on the CFD data (e.g., comparing wall and interior pressures in stenotic regions) to quantify its validity.","section":"Section 2"},{"comment":"The text states that 'All statistical tests were conducted in Python using the SciPy library,' but no statistical test results are reported anywhere in the manuscript. Please either add the significance tests or remove this sentence.","section":"Section 2.5"},{"comment":"The text says 'Figure 6 shows for each model (rows) the reconstructed vFFR field for 6 cases,' but Figure 6 displays only three models. Please make the number of displayed models explicit and consistent between text and figure caption.","section":"Section 3.2 and Figure 6"},{"comment":"The sentence 'DiffusionNet (S-variant) yielded the highest correlation coefficient (0.72), and together with LaB-GATr (S-variant) was the only model to achieve an accuracy above 0.7' should read 'the only models' because two models satisfy this condition.","section":"Section 3.2"},{"comment":"In Eq. (1), the notation r_{i-j} for the relative position is ambiguous; define it explicitly as the difference of coordinate vectors, e.g., r_i - r_j.","section":"Section 2.2.2"},{"comment":"The discussion of why pressure drop is preferable (homogeneity across patients, zero value at the ostium) is reasonable, but it is a post hoc explanation; a quantitative comparison of the target distributions (e.g., variance of pressure drop vs. pressure across the training set) would make the argument more concrete.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The missing patient-level split is the main concern. If the split was in fact patient-level and was simply not described, the paper would be much stronger. Please also check whether the 'only transformers' claim and the pressure-drop selection can be defended after adding proper statistical comparisons and clearly defining the selection criterion. The paper fits the journal's scope and is potentially valuable; the requested changes are local to the experimental protocol and analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read. The paper is a head-to-head benchmark of six geometric deep learning backends for predicting pressure/vFFR fields on coronary artery meshes, with 1,500 synthetic bifurcation cases and 427 patient-specific CFD runs. The synthetic part is well executed: same input features across models, two model sizes, three training targets, and a sensible ablation study. The finding that pressure drop is a better training target than pressure or vFFR directly is clearly supported on the synthetic data, and the reasoning about homogeneity across patients is plausible. The patient-specific results are more suggestive than definitive: LaB-GATr (S) gets the best test loss, per-point difference, and accuracy, but DiffusionNet (S) matches or beats it on correlation (0.72 vs 0.69) and also has accuracy above 0.7. So the abstract's 'only transformers' phrasing is too strong, and the conclusion overstates what the tables show.\n\nThe load-bearing problem is the patient-specific split. The paper never says whether the 427 runs were split by simulation or by patient. Since there are only 105 anatomies with 3-5 runs each, a simulation-level split means the same anatomy appears in training and test, and the model can memorize geometries and interpolate between BC variants. That would inflate every patient-specific metric and could manufacture LaB-GATr's apparent edge. This is not a physics quibble; it's a correctness question. The surface-constancy assumption in Section 2 looks fine for this purpose: wall pressure is the clinically relevant quantity and cross-sectional variation in coronaries is usually small.\n\nOther soft spots: no code or data, no statistical significance tests, and the ablation uses only PointNet++. The authors are candid about LaB-GATr's parameter mismatch and about dataset size, which is good. The citation pattern is fine; key methods from prior work are cited normally.\n\nWho is this for? Anyone working on GDL surrogates for hemodynamics or vFFR will want this as a comparison point. It deserves peer review, but the authors need to describe the exact patient split, report per-patient or at least per-anatomy metrics, and ideally release code/data. As it stands, I'd treat the synthetic results as solid and the patient-specific claim as plausible but unverified.","headline":"Useful benchmark with a solid synthetic study, but the patient-specific split is unspecified and the 'only transformers' claim is overstated.","tokens_in":20372,"tokens_out":2051,"would_cite":true,"duration_ms":20726,"reading_group":"yes","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 transformer-based geometric deep learning models, in particular LaB-GATr, are the only backends that generalize to patient-specific coronary artery meshes for predicting pressure-derived vFFR fields, and that…","keywords":["coronary artery disease","fractional flow reserve","virtual FFR","geometric deep learning","transformer networks","pressure drop","CFD surrogate","surface mesh hemodynamics"],"falsifier":"Run the same six backends on stenotic patient-specific meshes while also computing the CFD pressure difference between the vessel wall and the centerline in the lesion; if that wall-to-centerline difference exceeds the few-Pa error the paper reports, or if wall-only vFFR diverges from volumetric vFFR in the cases LaB-GATr misses, the surface-only premise rather than the architectures is the bottleneck.","tokens_in":19248,"feed_emoji":"🫀","tokens_out":7094,"duration_ms":61245,"temperature":0.7,"pith_summary":"This paper asks which geometric deep learning architecture can serve as a fast surrogate for CFD in computing virtual fractional flow reserve (vFFR) from coronary artery meshes. It benchmarks six backends—an MLP, PointNet++, DiffusionNet, DeltaConv, and two transformers—first on 1,500 synthetic left coronary bifurcations and then on 427 patient-specific CFD simulations. On simple synthetic geometries most backends perform well, especially when trained to predict pressure drop rather than absolute pressure or vFFR directly. On real patient anatomies only the transformer LaB-GATr generalizes strongly, keeping lesion-level vFFR predictions accurate, and the full pipeline runs in under two minutes per case. The paper concludes that transformer-based networks are the viable choice for complex, heterogeneous patient data.","feed_headline":"Transformers beat other AI models at predicting virtual FFR","feed_subtitle":"Only transformers kept virtual FFR accurate in stenotic lesions; inference takes under 30 seconds.","key_machinery":"The enabling setup is surface-only training: under an assumption of laminar, developed flow, pressure is taken as constant across each vessel cross-section, so the wall mesh carries all information needed for pressure, pressure drop, and vFFR fields. Each mesh vertex is assigned 28 input features that combine geometric descriptors (surface normal, geodesic distance to the inlet, centerline radius and tangent, Laplacian eigenvectors, heat and wave kernel signatures) with physical boundary conditions (inflow rate, inlet pressure, CFD node type). The decisive component for patient data is the transformer backend: LaB-GATr compresses the mesh through cross-attention to a coarse token set and interpolates back, giving each output vertex global context that local-convolution models such as DiffusionNet, DeltaConv, and PointNet++ lack.","core_discovery":"On the paper's own terms, the central discovery is that architecture choice determines whether learned hemodynamics transfer from synthetic to patient-specific anatomies. In the synthetic benchmark, several backends reached good accuracy, and pressure drop was consistently the best training target for all of them. On 427 patient-specific CFD runs, however, only LaB-GATr—a geometric algebra transformer for large biomedical meshes—achieved strong per-point accuracy and diagnostic vFFR accuracy in stenotic lesions, with the smallest bias (0.004 in stenotic-region vFFR) and the highest accuracy (0.79). The paper interprets this as evidence that global-context aggregation, which transformers provide, is what is needed for topologically complex and heterogeneous datasets.","pith_inferences":["An implication the authors do not develop is that the surface-only premise could be stress-tested by comparing wall and centerline CFD pressures in the exact lesions where LaB-GATr still fails; that would separate architecture limits from modeling limits.","A testable extension is to retrain LaB-GATr on a larger multi-center cohort and validate against invasive FFR rather than CFD, since the benchmark's ground truth is itself simulated.","The conclusion that global-context transformers are needed on heterogeneous topologies plausibly transfers to other surface hemodynamic fields such as wall shear stress, but that is an extrapolation beyond the pressure-related fields this paper measures.","Because the ablation found a small feature subset matching the full set, the clinical pipeline could be simplified by dropping spectral descriptors, potentially reducing preprocessing time and sensitivity to mesh discretization."],"forward_implications":["Machine-learning vFFR can be produced in under two minutes per case, compared with 30 to 60 minutes of steady-state CFD, if feature precomputation and inference are as fast as reported.","On simple bifurcation geometries, a range of backends can replace CFD for pressure-related fields, so model choice matters less in that regime.","New work in this area should default to pressure drop as the network output, with inlet pressure supplied as an input feature, rather than predicting absolute pressure or vFFR directly.","For patient-specific coronary anatomies, transformer-based backends, and specifically LaB-GATr, are the models to use if the goal is accurate stenosis-level vFFR.","The essential input features are surface normals, geodesics, vessel radius, inflow, and inlet pressure; spectral descriptors add little once these are present."],"supporting_citations":[{"why":"It supplies the 427 patient-specific steady CFD simulations used as ground truth and the automated boundary-condition pipeline.","marker":"[22]"},{"why":"It introduces LaB-GATr and LaB-VaTr, the transformer backends that carry the paper's main patient-specific results.","marker":"[36]"},{"why":"It defines the geometric algebra transformer architecture that LaB-GATr scales to large biomedical meshes.","marker":"[37]"},{"why":"It provides PointNet++, a compared backend and the model used for the input-feature ablation study.","marker":"[40]"},{"why":"It introduces DiffusionNet, the spectral-diffusion backend that performs strongly on synthetic pressure and vFFR tasks.","marker":"[41]"},{"why":"It introduces DeltaConv, the anisotropic convolution backend whose stable training required the pressure-drop target.","marker":"[42]"},{"why":"It supplies prior evidence that pressure drop is an effective learning variable for coronary vFFR estimation, which the paper extends.","marker":"[32]"},{"why":"It supports pressure drop as the preferred learning target through the approximation-disparity metric and steady-state prior work.","marker":"[54]"}],"fun_headline_variants":["Transformers beat other backends in virtual FFR benchmark","Transformers only achieve strong vFFR on patient data","LaB-GATr transformer wins vFFR prediction benchmark","Transformers generalize where other models fail on vFFR","Best vFFR predictions come from transformer networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pressure is constant across each vessel cross-section under laminar, developed flow, which justifies training only on surface meshes; if stenotic or recirculating flow breaks that uniformity, every model loses information that the CFD ground truth contains.","fun_headline_variants_meta":{"raw":{"variants":["Transformers beat other backends in virtual FFR benchmark","Transformers only achieve strong vFFR on patient data","LaB-GATr transformer wins vFFR prediction benchmark","Transformers generalize where other models fail on vFFR","Best vFFR predictions come from transformer networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000402,"raw_usage":{"total_tokens":2115,"prompt_tokens":985,"completion_tokens":1130,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1050}},"tokens_in":601,"tokens_out":1130,"duration_ms":8764,"temperature":1.0,"reasoning_tokens":1050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:19:16.490467+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same six backends on stenotic patient-specific meshes while also computing the CFD pressure difference between the vessel wall and the centerline in the lesion; if that wall-to-centerline difference exceeds the few-Pa error the paper reports, or if wall-only vFFR diverges from volumetric vFFR in the cases LaB-GATr misses, the surface-only premise rather than the architectures is the bottleneck.","supporting_citations":[{"cited_title":"Nannini, S","cited_arxiv_id":null,"evidence_quote":"It supplies the 427 patient-specific steady CFD simulations used as ground truth and the automated boundary-condition pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces LaB-GATr and LaB-VaTr, the transformer backends that carry the paper's main patient-specific results."},{"cited_title":"Brehmer, P","cited_arxiv_id":null,"evidence_quote":"It defines the geometric algebra transformer architecture that LaB-GATr scales to large biomedical meshes."},{"cited_title":"Sharp, S","cited_arxiv_id":null,"evidence_quote":"It introduces DiffusionNet, the spectral-diffusion backend that performs strongly on synthetic pressure and vFFR tasks."},{"cited_title":"Wiersma, A","cited_arxiv_id":null,"evidence_quote":"It introduces DeltaConv, the anisotropic convolution backend whose stable training required the pressure-drop target."},{"cited_title":"Rygiel, P","cited_arxiv_id":null,"evidence_quote":"It supplies prior evidence that pressure drop is an effective learning variable for coronary vFFR estimation, which the paper extends."}],"review_version":1}