{"id":"0a7bd851-026c-44a3-a7eb-0c29ccf76e21","arxiv_id":"2508.03326","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"PINNs reconstruct smooth velocity and pressure fields from synthetic 4D-flow MRI data for anemic to polycythemic blood, and combining the PINN velocity field with the vWERP estimator gives the most accurate pressure drops.","lead":"This paper tests physics-informed neural networks (PINNs) for reconstructing blood velocity, pressure, and viscosity from synthetic low-resolution 4D-flow MRI data. It covers five hematocrit levels in a realistic aortic model and benchmarks pressure-drop estimates against the vWERP method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pressure-drop claim is contradicted by Table 5: max-relative errors reach 69.7% (PINN) and 36.6% (PINN+vWERP), and vWERP-only errors are never tabulated, so 'clinically admissible' and 'outperform vWERP' are not supported.","rationale":"The reader's CONDITIONAL verdict is reasonable, but the more direct problem is internal reporting rather than the noiseless synthetic data. A synthetic benchmark can still be valuable, yet the paper's own Table 5 contradicts the abstract's 'below 5%' and 'clinically admissible accuracy' phrasing when that language is read as applying to pressure drops. The missing vWERP-only table means the core comparison is not auditable from the paper. A fair test is to compute time-resolved errors and tabulate all methods; this does not invalidate the methodology itself but moves the required revision from 'add noisy experiments' to 'correct the claims and metrics and provide the missing comparison.' Thus the reader's CONDITIONAL verdict stands unchanged, with the rationale sharpened: the central claim needs better quantitative support before it can be accepted, and the internal evidence gap is the most load-bearing issue.","tokens_in":28797,"tokens_out":5090,"duration_ms":59629,"concrete_test":"Recompute the pressure-drop benchmark with a time-resolved relative L2 error, e.g. ||\\delta p_est - \\delta p_ref||_{L2([0,T])} / ||\\delta p_ref||_{L2([0,T])}, for all four outlets and all five hematocrit levels, and tabulate the vWERP-only errors alongside the PINN-based methods. If the time-resolved errors of PINN+vWERP are not consistently below vWERP's, the 'outperform' claim fails; if they exceed 5% (or the stated clinical threshold), the 'clinically admissible' claim fails. This requires only the existing code and data and settles whether the headline is an artifact of the peak-only metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim in the Abstract and Section 5.3 is that pressure-drop estimates are obtained 'with a clinically admissible accuracy' and that PINN-based methods outperform vWERP. The paper's own quantitative table undermines this. Section 5.3 defines e_dp as the relative error in the absolute maximum of the pressure drop. Table 5 shows direct PINN errors from 27.30% to 69.68% and PINN+vWERP errors from 7.97% to 36.61% across hematocrit levels and outlets. These numbers are not below 5%, and no clinical threshold is given that would make 8-37% errors 'clinically admissible.' The abstract's phrase 'pressure drops with relative errors below the 5%' is therefore either inconsistent with Table 5 or refers only to the whole pressure field, not to the pressure-drop biomarker; either way the headline claim overstates the evidence. In addition, Table 5 reports only the three PINN-based strategies. vWERP-only errors appear only qualitatively in Figure 12, so the quantitative claim 'outperform vWERP' cannot be checked from the paper. The chosen metric, peak absolute error rather than a time-resolved error, also favors the super-sampled methods and may hide larger errors at other cardiac phases. The Reader's noiseless synthetic-data concern compounds this, but the missing comparison and metric choice are internal to the paper and thus more immediately load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-informed neural network (PINN) framework for estimating velocity, pressure, and apparent viscosity fields from synthetic 4D-flow MRI data in a realistic aortic geometry, using a hematocrit-dependent power-law rheology model. Methodological contributions include a voxel-averaging observation operator, curriculum training, adaptive loss balancing, and residual-based collocation refinement. The authors evaluate five hematocrit levels, report pressure-field reconstruction errors below 1.66%, and compare pressure-drop estimates from direct PINN evaluation, the vWERP estimator alone, and a hybrid PINN+vWERP strategy that super-samples the PINN velocity field. The central quantitative claims are that pressure fields are accurate, that pressure-drop estimates are clinically admissible, and that PINN-based approaches outperform vWERP.","tokens_in":29126,"tokens_out":4007,"duration_ms":49257,"significance":"If the claims are appropriately scoped, the paper contributes a useful demonstration of PINN state estimation with a realistic 4D-flow measurement operator and systematic variation of hematocrit-driven rheology. The pressure-field errors (0.93-1.66%, Fig. 10) are well quantified and support the core methodological claim of accurate whole-field pressure reconstruction. The hybrid PINN+vWERP idea is sensible and the qualitative improvement in time resolution is plausible. However, the paper's headline claims about pressure-drop accuracy and superiority over vWERP are not supported by the tabulated numbers, and the noiseless synthetic-data setup limits the clinical interpretation. These issues are fixable by reframing claims and adding missing quantitative comparisons.","major_comments":[{"comment":"The abstract's phrase \"pressure drops with relative errors below the 5% in the whole pressure field\" is ambiguous, and the Introduction states \"particularly accurate pressure drop estimations, relative errors as low as 1%\" (end of Section 1). Table 5 reports e_dp values between 27.30% and 69.68% for the direct PINN and between 7.97% and 36.61% for PINN+vWERP across outlets and hematocrit levels. Unless the intended claim concerns only the whole pressure field (Fig. 10, errors 0.93-1.66%), the abstract and Introduction overstate the accuracy of pressure-drop estimates. Please rewrite these claims to distinguish whole-field pressure errors from the pressure-drop biomarker, and provide an explicit clinical threshold if \"clinically admissible accuracy\" is retained.","section":"Abstract; Section 5.3; Table 5"},{"comment":"The claim that PINN-based methods \"outperform vWERP in terms of both accuracy and time resolution\" is not quantitatively verifiable from the paper. Table 5 reports errors only for the three PINN-based strategies, and no numerical e_dp values are given for vWERP alone. Figure 12 shows only one hematocrit level (32.5%) and no error metric. Please add vWERP1st and vWERP2nd columns to Table 5 (or a companion table) and specify the comparison metric. In addition, e_dp is defined on the maximum absolute pressure drop over the cycle; a time-resolved error measure (e.g., L2 relative error over the cardiac phase per outlet) would strengthen the claim that the super-sampled methods are more accurate throughout the cycle, not only at the peak.","section":"Abstract; Section 5.3; Table 5; Figure 12"},{"comment":"The synthetic 4D-flow data are generated without noise or artifacts (Section 3.3: \"without taking into account noise or the more complex artifacts typically associated with MR imaging\"), and the same power-law Navier-Stokes model generates the data and acts as the physics prior in the PINN loss. The experiments are therefore a consistency check rather than a clinical validation. The abstract's \"clinically admissible accuracy\" overreaches this evidence. Please add a limitations paragraph and reframe the clinical claims as proof-of-concept, or include experiments with realistic noise, aliasing, and displacement artifacts.","section":"Section 3.3; Abstract"},{"comment":"The pressure inference relies on the ground-truth global mean pressure pmean as a data term (Eq. 3.1, Table 3), and all pressure-error metrics are computed after shifting the estimated pressure to match the reference mean at each time instant (Section 5.2, Fig. 8 caption). This should be stated explicitly as a limitation: the PINN does not estimate the pressure constant from 4D-flow velocity data alone. The paper should discuss how pmean would be obtained in a clinical setting and assess the sensitivity of the reported pressure-field errors to uncertainty in pmean. Relative pressure drops are less affected, but the presentation of pressure-field accuracy needs this caveat.","section":"Eq. (3.1); Table 3; Section 5.2; Figure 8"}],"minor_comments":[{"comment":"The text says \"For the 35% hematocrit, the difference was 4.4%\" but the simulated hematocrit levels are 20%, 32.5%, 45%, 57.5%, and 70%; this should read 32.5%.","section":"Section 5.1, Figure 6 discussion"},{"comment":"The caption of Table 5 says \"Errors in absolute maxima\" while the text defines e_dp using max|δp|; please clarify that the maximum is taken over time and that the error is relative to the reference maximum, and state whether all outlets follow the same convention.","section":"Table 5 caption and Eq. (5.1)"},{"comment":"The figure caption mentions that pressure is corrected to the same average as the reference at each timestep, but the main text also applies this correction for the error metrics in Section 5.2; making this explicit in the text near Eq. (5.1) would avoid confusion about what the reported pressure errors represent.","section":"Figure 8 caption"},{"comment":"The statement that all code is written in PyTorch could be complemented by a data/code availability statement; this is not required for acceptance but would improve reproducibility.","section":"Section 3.2, implementation details"}],"recommendation":"major_revision","confidential_remarks":"The comparative claim against vWERP is the weakest part of the paper: without a numerical vWERP-only baseline, the central selling point cannot be verified. The noiseless, same-model synthetic setup is acceptable for a proof-of-concept, but the clinical language in the abstract and conclusions should be toned down. The methodological core -- a realistic voxel-averaging operator, curriculum training, and adaptive weighting -- is sound and likely of interest to the math.NA and hemodynamics communities, so major revision rather than rejection is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something useful: it builds a careful synthetic 4D-flow MRI benchmark for PINN-based velocity and pressure reconstruction, with a time-aware voxel-averaging observation operator and hematocrit-dependent power-law rheology across five hematocrit levels. The engineering is thoughtful — curriculum training, adaptive loss balancing, and RAR-G are all deployed sensibly, and the pressure field reconstruction errors (0.93–1.66% relative) are well quantified and genuinely good. The comparison against vWERP is a reasonable and timely benchmark, and the PINN+vWERP hybrid is a sensible idea that may well help in practice.\n\nThe soft spot is the abstract. It claims pressure drops are estimated with relative errors below 5%, and that the results outperform vWERP with clinically admissible accuracy. The paper's own Table 5 shows peak pressure-drop errors of 27–70% for the direct PINN and 8–37% for PINN+vWERP across outlets and hematocrit levels. Those numbers are not below 5% by any reading. The \"below 5%\" figure appears to refer to the whole pressure field, not the pressure-drop biomarker, but the abstract does not say that. Also, vWERP-only errors are never tabulated, so the quantitative claim of outperforming vWERP is not actually checkable from the paper. That is a significant gap between framing and evidence.\n\nOther limitations are real but more honestly disclosed. The synthetic data are noiseless and artifact-free, the physics loss uses the same constitutive model that generated the data, and the network receives the ground-truth global mean pressure. These are stated in the text, so the reader can weigh them. They do mean the results are a consistency check rather than a clinical validation, which is fine as long as the claims are scaled accordingly.\n\nI would send this to peer review. The methods are thorough, the synthetic benchmark is well executed, and the pressure field results are worth reporting. But the authors need to revise the abstract, include the vWERP-only numbers, and either change the metric or stop claiming clinically admissible accuracy for pressure drops. As it stands, the paper overstates its own headline result, and that should be fixed before publication.","headline":"Solid synthetic benchmark for PINN-based hemodynamic state estimation, but the abstract's pressure-drop accuracy claim is contradicted by the paper's own Table 5.","tokens_in":29634,"tokens_out":1529,"would_cite":false,"duration_ms":19795,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M32","76D05","92C35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a physics-informed neural network can reconstruct aortic velocity, pressure, and hematocrit-dependent viscosity from low-resolution 4D-flow MRI-like data, and that combining it with the vWERP estimator yields the…","keywords":["physics-informed neural networks","4D-flow MRI","hemodynamics","pressure drop","hematocrit","non-Newtonian rheology","aorta","state estimation"],"falsifier":"A direct test would run the same pipeline on real 4D-flow MRI acquisitions of an aorta with a known pressure-drop reference (for example, from catheterization or a phantom with a known stenosis). If, under real noise and artifacts, the combined PINN+vWERP estimate's peak pressure-drop error exceeds the roughly 8-11% range seen in the synthetic experiments, the claimed clinically admissible accuracy would be falsified for clinical data.","tokens_in":1533,"feed_emoji":"","tokens_out":1745,"duration_ms":88427,"temperature":0.7,"pith_summary":"The paper asks whether a physics-informed neural network can turn low-resolution 4D-flow MRI-like velocity measurements into full, physically coherent velocity, pressure, and viscosity fields in a realistic aorta, and whether those fields yield reliable pressure-drop biomarkers. It reports that it can: across five hematocrit levels from anemic to polycythemic, the network reconstructs pressure with coefficient of determination above 0.99 and relative field errors mostly near 1%, reproduces flow splitting through the aortic branches, and recovers shear-thinning viscosity patterns. It further claims that using the network's smooth, time-supersampled velocity as input to the standard vWERP pressure estimator beats vWERP applied directly to the raw images, bringing maximum pressure-drop errors down to single digits in the best cases. The payoff would be a non-invasive route to clinical biomarkers currently obtained by catheterization, without assuming Newtonian blood behavior.","feed_headline":"PINN + vWERP nails aortic pressure drops","feed_subtitle":"Low-resolution 4D-flow data plus physics yields pressure errors down to 8%, a step toward non-invasive diagnosis.","key_machinery":"The load-bearing object is the neural field together with a measurement operator that mimics 4D-flow MRI: instead of point samples, the observation loss compares the network to local spatiotemporal averages over 2-mm voxels and 42.6-ms cardiac phases, computed by quasi-Monte Carlo quadrature with scrambled low-discrepancy points. This operator connects the low-resolution data to the high-resolution field, and a boundary-volume loss makes the network's local averages see zero velocity outside the lumen. The physics are enforced pointwise by automatic differentiation of the power-law Navier-Stokes residual at roughly 20 million collocation points, with adaptive inverse-Dirichlet loss weighting and a two-stage curriculum that first fits the data alone and then activates the physics. For pressure drops, the vWERP estimator, a virtual work-energy identity that uses solenoidal Stokes test functions to convert the weak Navier-Stokes equations into outlet pressure differences, is fed either the raw images or the PINN's smooth velocity; the second route is the paper's winning strategy.","core_discovery":"On its own terms, the work establishes a reconstruction pipeline: a fully connected neural field maps space-time points to velocity and pressure, and is trained by minimizing a sum of four losses: voxel-averaged velocity observations through a realistic observer operator, one scalar mean-pressure observation, Navier-Stokes residuals with a power-law viscosity whose consistency index and power-law index depend on hematocrit, and a no-slip wall condition enforced in a boundary volume. The paper's central quantitative claims are that this yields pressure fields with relative error between 0.93% and 1.66% over the whole space-time domain across all five hematocrit levels, that the velocity reconstructions reproduce the bifurcation of flow through the aortic branches with errors of a few percent, and that the best pressure-drop estimates come from post-processing the PINN velocity with vWERP, outperforming both direct PINN pressure reads and raw-image vWERP. In the descending aorta, the best peak pressure-drop error drops to about 8% with the combined method, while direct PINN errors range roughly from 27% to 70% depending on outlet and hematocrit. The authors conclude that the best strategy is to reconstruct first a high-fidelity, Navier-Stokes-consistent velocity field and then apply a work-energy estimator to recover the pressure drop.","pith_inferences":["This is an editorial inference, not the paper's claim: because the same power-law model generates the synthetic data and supplies the physics loss, the reported accuracy is a self-consistency check, and real MRI noise, aliasing, and flow-displacement artifacts would likely degrade it in ways the paper does not quantify.","The hybrid strategy suggests a general principle: for derivative-based or integral biomarkers, smooth physics-constrained velocity reconstruction can act as a denoiser and time-supersampler upstream of classical estimators, and testing this on wall shear stress and oscillatory shear index is a natural next step.","Replacing the power-law model with a different constitutive law, or adding vessel-wall compliance, would test whether the gains persist when the data-generation model and the network's physics model are not identical.","Because the measurement operator already handles spatial and temporal voxel averaging, adding realistic noise models directly into the synthetic data generation would give a low-cost read on how much of the claimed accuracy survives clinical conditions."],"forward_implications":["Pressure-drop biomarkers for aortic coarctation, valve stenosis, and congenital heart disease could be computed non-invasively from low-quality 4D-flow acquisitions, avoiding risky catheterization.","The PINN+vWERP pipeline recovers systolic pressure-drop peaks that raw low-time-resolution images miss, because the network supplies about 1000 samples per cardiac cycle instead of the 22 image phases.","Estimates remain accurate across a wide hematocrit range, suggesting the method applies to anemic, normal, and polycythemic patients without switching rheology models.","The same reconstructed fields locate high wall-shear-stress regions near the inflow jet impact, a region relevant for identifying disease-prone areas of the aortic wall."],"supporting_citations":[{"why":"Supplies the network architecture and the parameter-estimation PINN setup that this work extends with a more realistic observation operator and non-Newtonian rheology.","marker":"[57]"},{"why":"Provides the state-of-the-art relative-pressure estimation framework and the comparison baseline that the paper's pressure-drop results are measured against.","marker":"[6]"},{"why":"Introduces the virtual work-energy relative pressure (vWERP) estimator that the paper feeds with both raw images and PINN-reconstructed velocity.","marker":"[87]"},{"why":"Supplies the work-energy equation formulation from which the vWERP pressure-drop identity is derived.","marker":"[36]"},{"why":"Establishes the original physics-informed neural network framework that defines how governing equations enter the training loss.","marker":"[41]"},{"why":"Provides the empirical hematocrit-dependent power-law parameters used to define the five rheological cases from anemic to polycythemic.","marker":"[70]"},{"why":"Supports the claim that viscosity-model choice materially changes biomarker estimates such as wall shear stress in aortic simulations.","marker":"[14]"},{"why":"Defines the 2023 consensus acquisition parameters that justify treating the synthetic 2-mm, 42.6-ms voxel data as low-quality 4D-flow MRI.","marker":"[62]"}],"fun_headline_variants":["PINN velocity + vWERP beats raw MRI pressure drops","Physics-informed nets estimate aortic pressure to under 2% error","Hematocrit-aware PINNs reconstruct aortic flow and pressure from MRI","AI + work-energy law yields accurate aortic pressure from sparse MRI"],"cache_read_input_tokens":31744,"weakest_assumption_plain":"The load-bearing premise is that the synthetic 4D-flow MRI measurements are representative of real clinical data, meaning they include no noise, aliasing, flow-displacement artifacts, or other image degradation beyond local spatiotemporal averaging.","fun_headline_variants_meta":{"raw":{"variants":["PINN velocity + vWERP beats raw MRI pressure drops","Physics-informed nets estimate aortic pressure to under 2% error","Hematocrit-aware PINNs reconstruct aortic flow and pressure from MRI","AI + work-energy law yields accurate aortic pressure from sparse MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2409,"prompt_tokens":1094,"completion_tokens":1315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":1254}},"tokens_in":710,"tokens_out":1315,"duration_ms":13055,"temperature":1.0,"reasoning_tokens":1254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:30:02.116094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would run the same pipeline on real 4D-flow MRI acquisitions of an aorta with a known pressure-drop reference (for example, from catheterization or a phantom with a known stenosis). If, under real noise and artifacts, the combined PINN+vWERP estimate's peak pressure-drop error exceeds the roughly 8-11% range seen in the synthetic experiments, the claimed clinically admissible accuracy would be falsified for clinical data.","supporting_citations":[{"cited_title":"Garay, J","cited_arxiv_id":null,"evidence_quote":"Supplies the network architecture and the parameter-estimation PINN setup that this work extends with a more realistic observation operator and non-Newtonian rheology."},{"cited_title":"Bertoglio, R","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art relative-pressure estimation framework and the comparison baseline that the paper's pressure-drop results are measured against."},{"cited_title":"Donati, C","cited_arxiv_id":null,"evidence_quote":"Supplies the work-energy equation formulation from which the vWERP pressure-drop identity is derived."},{"cited_title":"Walburn, D","cited_arxiv_id":null,"evidence_quote":"Provides the empirical hematocrit-dependent power-law parameters used to define the five rheological cases from anemic to polycythemic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 2023 consensus acquisition parameters that justify treating the synthetic 2-mm, 42.6-ms voxel data as low-quality 4D-flow MRI."}],"review_version":1}