{"id":"89530cb9-2023-4054-a0ed-8cb91c3f4c50","arxiv_id":"2508.19303","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A U-Net trained on finite-element-simulated displacement fields reconstructs aortic wall shear-modulus maps from 2D ultrasound, with promising results on COMSOL, phantom, and a small clinical cohort.","lead":"This paper trains a U-Net neural network on simulated ultrasound displacement data to map vessel-wall deformation into a stiffness map of abdominal aortic aneurysms. If the method holds up in larger patient studies, it could add a non-invasive stiffness biomarker to routine ultrasound screening for aneurysm rupture risk.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Max-displacement normalization (Eqs. 19-22) makes absolute modulus output inversely proportional to the measured displacement amplitude; a ~50% displacement underestimation would explain the factor-of-two phantom modulus overestimation, undermining the quantitative claim.","rationale":"The reader's weakest assumption was the general representativeness of the FE training data for real ultrasound physics. My concern is more specific and identifies an untested amplifier within that gap: the max-displacement normalization in Eqs. (19)-(22) converts any systematic amplitude error in the measured displacement field into an inversely proportional error in the absolute modulus. The phantom results, where absolute modulus is overestimated by up to a factor of two while the modular ratio is roughly preserved, are exactly what this mechanism predicts. This is a stronger and more concrete failure mode than the broad sim-to-real mismatch because it does not depend on unmodeled nonlinearities or radial gradients; it is a consequence of the chosen scaling procedure. The paper itself acknowledges noise sensitivity and phantom offsets, which supports the honesty of the reporting, but it does not identify the normalization as the likely amplifier. My proposed test would settle the mechanism directly. I do not believe this changes the reader's CONDITIONAL verdict: the paper is clearly positioned as a feasibility study, the phantom and COMSOL results demonstrate relative contrast ability, and the absolute scale issue is an acknowledged limitation that could, in principle, be addressed with calibration or a more robust scale estimator. However, the condition should be explicit: the quantitative shear modulus claim is contingent on demonstrating that the displacement amplitude bias is understood and corrected. The simulated NMSE of 0.73% and the high DSC scores are legitimate support for the network's ability to map displacement shape to modulus shape, but they do not validate the absolute scale. The ITR comparison is useful and shows similar bias, so the DL is not uniquely defective, but that does not rescue the quantitative claim. Thus the verdict remains CONDITIONAL, with my concern sharpening the specific condition that needs to be met.","tokens_in":18669,"tokens_out":12926,"duration_ms":131815,"concrete_test":"Run DL inference on one simulated test case after multiplying both displacement components by beta = 0.5, 0.8, 1.0, and 1.2 while keeping the pressure normalization unchanged. If the reconstructed modulus scales as 1/beta, the output is directly controlled by displacement amplitude. Then build an FE forward model of Phantom 1 with the independently measured 17.4 kPa modulus and measured pulse pressure and compare the simulated maximum displacement with the maximum displacement from the registered ultrasound data. If the registered max is about half the FE value, the factor-of-two phantom overestimation is explained by the normalization mechanism, and the central quantitative claim requires a displacement-amplitude calibration or correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The DL model's quantitative scale is set by Eqs. (19)-(22): the input is normalized by nm = max|u|/P, the network predicts G = mu*nm, and the final modulus is mu_dl = G/nm. Because the input is divided by nm, the network output G is invariant to uniform scaling of the measured displacement; the predicted absolute modulus is inversely proportional to nm. Any systematic underestimation of displacement amplitudes by the regularized registration of Eq. (8), with its divergence-penalizing regularization and use of only two selected frames, therefore translates directly into proportional modulus overestimation. The phantom results are consistent with this: for uniform Phantom 1, true modulus is 17.4 kPa but DL predicts 31-34 kPa (factor ~1.9), implying a ~50% displacement underestimation. This is not merely a 'sim-to-real' noise gap; it is a scale-recovery design that places the entire quantitative accuracy on a single, easily biased scalar (the maximum displacement), a mechanism the paper does not discuss. The COMSOL results show a different bias (underestimation at high contrast), so the input amplitude error varies across domains, but the point stands: the 0.73% NMSE is an in-distribution test of the U-Net, not validation of absolute modulus accuracy on real data. The clinical results have no ground truth and only compare against strain derived from the same displacement fields, so they cannot settle the scale.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning-based framework for 2D ultrasound elasticity imaging of abdominal aortic aneurysms (AAAs). The authors generate a large dataset of 36,000 2D plane-strain finite element simulations (30,000 training examples) with randomized vessel geometry, angular shear-modulus variations, and boundary conditions, then train a U-Net to map pressure-normalized displacement fields to shear-modulus distributions. The model is evaluated on (i) a simulated test set, where it achieves NMSE 0.0073 and DSC 0.999; (ii) COMSOL 3D digital phantoms, where it recovers modulus contrast with moderate accuracy; (iii) four physical tissue-mimicking phantoms with independently measured shear moduli, where contrast ratios are roughly preserved but absolute moduli are systematically overestimated by up to a factor of about two; and (iv) four clinical AAA ultrasound exams, where the average predicted modulus is compared against an inverse strain measure. The deep-learning reconstruction is compared with an iterative method (ITR), showing comparable contrast recovery at a fraction of the computation time (0.058 s vs. 158.57 s per example). The central claim is that a model trained only on FE simulations can predict shear modulus from experimental ultrasound elastography displacement estimates.","tokens_in":18955,"tokens_out":6965,"duration_ms":74491,"significance":"If the quantitative accuracy were established, this work would be a meaningful step toward real-time, non-invasive stiffness mapping of AAA walls using standard 2D ultrasound, with potential value for rupture-risk assessment beyond maximum diameter. The paper has clear strengths: a large and transparently described simulated training dataset, multi-domain evaluation (simulation, COMSOL, physical phantoms, and clinical data), inclusion of an established iterative baseline, and explicit acknowledgment of several limitations. The reported speed advantage of the deep-learning method is substantial. However, the absolute quantitative claim is currently not supported by the experimental results, because the normalization scheme makes the final modulus inversely proportional to a single displacement-amplitude scalar, and the phantom data show a factor-of-two bias. The clinical validation is partly circular. These issues are load-bearing for the stated goal of quantitative tissue characterization, but they appear addressable through additional analysis and calibration, so the work merits major revision rather than rejection.","major_comments":[{"comment":"The max-displacement normalization makes the predicted absolute modulus inversely proportional to the measured displacement amplitude. Because the network input is u_hat = u_tilde/nm and the output is rescaled by mu_dl = G/nm, the network itself is invariant to uniform scaling of the measured displacement, and the entire absolute scale is carried by the single scalar nm. Any systematic error in displacement magnitude from the two-frame, divergence-regularized registration of Eq. (8) therefore propagates directly and inversely into the final modulus. The phantom results are consistent with this: Phantom 1 (uniform 17.4 kPa) is predicted as 34.1 and 31.1 kPa for the upper and lower regions, a roughly 1.8-2.0x overestimation implying about 50% underestimation of nm. The paper acknowledges the offset only as \"reasons for these offsets require further study\" without identifying this mechanism. This is load-bearing because the central claim is quantitative modulus estimation from experimental data, not just relative contrast.","section":"Section III-C, Eqs. (19)-(22); Table IV"},{"comment":"The clinical sanity check is partly circular. The average predicted modulus and the pressure-normalized maximum principal strain are both computed from the same measured displacement field, and since the predicted modulus is inversely proportional to the maximum displacement magnitude, an inverse correlation with strain is expected a priori. This comparison therefore does not validate the absolute modulus scale against an independent reference. The authors should state this limitation explicitly and, ideally, supplement the clinical evaluation with a calibration phantom or an independent stiffness measurement, even if only in a subset of patients.","section":"Section V, Fig. 12"},{"comment":"The training distribution lacks radial modulus variations, includes a fixed Poisson ratio (0.45), noiseless displacement fields, and simplified boundary conditions; the authors acknowledge these points in the Discussion. However, the conclusion that the model \"can be used to predict shear modulus from both simulated and experimental ultrasound elastography displacement estimates\" is stronger than the evidence supports. The experimental results show large absolute bias, and the absence of radial gradients in training means the model cannot represent them by construction (the authors note that DL reconstructions lack the radial artifact seen in ITR). The quantitative generalization claim requires either additional evidence (e.g., noise-robustness studies, phantom calibration) or a more careful statement that the model is currently validated for relative, angular contrast mapping.","section":"Section II-A and Discussion"}],"minor_comments":[{"comment":"The coordinates used for boundary conditions appear inconsistent with the stated domain ([-0.12, 0.12] m in x, [0, 0.20] m in y). For example, \"ux(0.20, 0) = 0\" and later references to x = 0.20 and x = 0.50 fall outside the stated x-range. Please clarify the coordinate convention used in the FE model.","section":"Section II-A"},{"comment":"The text first says \"we calculated the total accumulated displacement field... from diastole to systole\" and then immediately notes that \"no accumulation of displacement fields was performed here.\" Please reconcile these statements so the methodology is unambiguous.","section":"Section III-A"},{"comment":"The term \"Modular Ratio\" should be \"Modulus Ratio\" or \"shear-modulus ratio\" to match standard terminology and avoid confusion with modular arithmetic.","section":"Tables III and IV"},{"comment":"The abbreviation \"PV A % wt\" should be defined (presumably polyvinyl alcohol weight percent) in the table caption or text.","section":"Table I"},{"comment":"The quantity defined in Eq. (24) is a relative squared error (sum of squared errors divided by sum of squared ground-truth values), not a normalized mean squared error as commonly defined. Consider renaming it or clarifying the definition.","section":"Eq. (24)"},{"comment":"The abstract states the model achieves an NMSE of 0.73% on simulated data and \"generalizes\" to phantom data; it would be helpful to state explicitly that the 0.73% NMSE is on the simulated test set only, since the experimental results have larger errors and are quantified mainly via modulus ratios.","section":"Abstract"},{"comment":"The training and validation loss curves would be easier to interpret with a logarithmic y-axis, since the NMSE drops to about 0.01 and the early training behavior is compressed.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid feasibility study with careful multi-domain evaluation, but the central quantitative claim is undermined by the scale-recovery mechanism: the max-displacement normalization transfers all absolute calibration to a single measured scalar, and the phantom data show a factor-of-two bias that is not analyzed. If the authors can quantify the displacement-amplitude error, add a calibration step or an independent validation of the absolute scale, and soften or more precisely scope the generalization claim, the paper could be suitable for publication. The clinical sanity check should also be framed with its circularity made explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on ultrasound elastography or sim-to-real deep learning for inverse problems. The genuinely new part is not the U-Net itself—the same group introduced that in [19,31]—but the AAA-specific FE dataset (36k varied geometries and angular modulus patterns) and the multi-domain validation against COMSOL, physical phantoms, and four clinical exams, all compared head-to-head with an iterative method. That is real work, and the comparison is useful. On simulated test data the network gets NMSE 0.0073 and DSC 0.999; on COMSOL it tracks contrast ratios reasonably and beats ITR on NMSE. So the paper does what it claims as a feasibility study.\n\nThe soft spots matter. The biggest is that the network's absolute output scale is set by Eq. (22): the raw output is divided by the max measured displacement magnitude. If the registration underestimates displacement amplitudes by 50%, you get a 2x modulus overestimation, and the phantom results look like exactly that—uniform Phantom 1 (true 17.4 kPa) comes out at 34.1 kPa by DL and 37.6 kPa by ITR. The paper never discusses this mechanism. The authors do say the offsets \"require further study,\" but they do not mention that their own normalization makes modulus output inversely proportional to measured displacement amplitude. Because the COMSOL results show a different bias (underestimation at high contrast), it is not one clean calibration error; it is a scale-recovery issue that deserves at least a sensitivity analysis. That is the load-bearing caveat.\n\nAlso, the training inputs are noiseless 2D plane-strain FE fields with no radial modulus variation and fixed Poisson ratio. The authors acknowledge this in the Discussion, and it is acceptable for a feasibility study, but it plainly limits the quantitative claim. The clinical data have no ground truth; the inverse-strain correlation is sanity support, not validation of absolute modulus. And the lack of released code/data is unfortunate given how easy it would be to test the normalization sensitivity numerically.\n\nWho benefits: vascular ultrasound researchers and people building deep-learning inverse solvers for elastography. It deserves peer review, but a referee should ask for an explicit displacement-amplitude sensitivity study and a clear statement that only contrast ratios, not absolute moduli, are currently supported on real data.","headline":"A solid feasibility study for DL-based AAA stiffness mapping from 2D ultrasound, with real multi-domain validation, but the absolute modulus numbers on real data are off by ~2x and the max-displacement normalization makes the quantitative claim fragile.","tokens_in":716,"tokens_out":684,"would_cite":true,"duration_ms":43016,"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":"A deep learning model, trained only on finite-element simulations, can reconstruct aortic wall shear modulus from 2D ultrasound displacement fields.","keywords":["abdominal aortic aneurysm","ultrasound elastography","deep learning","inverse elasticity problem","shear modulus reconstruction","finite element simulation","U-Net","medical imaging"],"falsifier":"Take a physical phantom or ex-vivo aortic sample with an independently measured shear modulus, from mechanical testing or MR elastography, image it with the same ultrasound protocol, and compare the DL-predicted modulus to the measured value across a range from about 1 to 170 kPa; if the predicted values do not increase monotonically with true stiffness or deviate from truth by more than a clinically acceptable margin in a substantial fraction of cases, the claim that the FE-trained model generalizes would be refuted.","tokens_in":18442,"feed_emoji":"🩺","tokens_out":6178,"duration_ms":60538,"temperature":0.7,"pith_summary":"The paper aims to show that a deep neural network trained exclusively on simulated finite-element displacement fields can replace slow iterative solvers for ultrasound elastography of abdominal aortic aneurysms. The clinical motivation is that maximum aneurysm diameter alone predicts rupture risk poorly, while wall stiffness is a direct mechanical marker that standard 2D ultrasound could measure if the inverse problem of converting displacement to shear modulus could be solved in real time. The authors report that their U-Net reconstructs shear modulus distributions from axial and lateral displacement inputs with 0.73% normalized mean squared error on simulated test data, and that the same model produces plausible stiffness maps on 3D-simulation digital phantoms, tissue-mimicking physical phantoms, and clinical patient exams. Because a prediction takes about 0.058 seconds versus about 159 seconds for the iterative baseline, the paper's contribution is a route to practical, non-invasive, quantitative stiffness imaging on hardware already used for AAA screening.","feed_headline":"AI maps aortic stiffness from 2D ultrasound in under a second","feed_subtitle":"Neural net trained on simulated displacements beats iterative solver speed by ~2,700x on phantom stiffness maps.","key_machinery":"The load-bearing object is a U-Net with seven downsampling and seven upsampling layers and a latent dimension of 512, trained with the normalized mean squared error loss between predicted and ground-truth modulus images. The inputs are the axial and lateral components of the pressure-normalized displacement field, gridded to 128x128 pixels with all pixels outside the vessel wall set to zero, which gives the network an implicit segmentation of the lumen. Pressure normalization, dividing displacements by the measured pulse pressure, is what lets the network output quantitative modulus values in kPa rather than relative stiffness. The FE training set varies aneurysm radius, wall thickness, angular modulus distribution, boundary conditions, and lumen pressure across 36,000 forward models, and the network learns the inverse map from those displacement fields.","core_discovery":"The paper's central claim, stated in its conclusion, is that a deep learning model trained only on finite-element simulated data can predict shear modulus from both simulated and experimental ultrasound elastography displacement estimates. On the 3,000-image simulated test set the model achieves a mean normalized mean squared error of 0.0073 and a mean Dice score of 0.999, indicating quantitatively close and spatially accurate modulus reconstructions. On 3D-simulation digital phantoms it matches the true upper/lower modulus ratio more closely than the iterative method at high contrast, and on physical phantoms the predicted regional modulus ratios track the independently measured values (1.09, 2.48, 5.05, 11.17 versus expected 1.00, 2.78, 5.47, 9.77). For clinical exams, where no ground truth exists, the authors show that the predicted average modulus is inversely related to the measured pressure-normalized maximum principal strain, the expected compliance relationship, for both the deep learning and iterative reconstructions. The claim is that a sufficiently diverse set of simulated training examples teaches the network the displacement-to-modulus map well enough to generalize across simulation, phantom, and clinical domains.","pith_inferences":["Because training data contain no radial modulus variation, the network's clean radial reconstructions may reflect what it never saw; adding radial gradients to the FE training distribution is a direct test of whether the current phantom overestimates, such as Phantom 1 at 34.1 kPa versus its true 17.4 kPa, shrink.","The paper's observed inverse relationship between predicted modulus and pressure-normalized strain suggests a built-in clinical self-check: a deployed system could flag unreliable reconstructions whenever those two quantities move in the same direction rather than inversely.","If region-ratio rather than absolute modulus is the target biomarker, the phantom results imply the model may already be clinically useful even while absolute values are biased; a longitudinal study linking predicted ratios to rupture or growth would settle that question.","Adding simulated ultrasound speckle noise and out-of-plane motion to the training set is the natural next experiment, and the paper's own discussion anticipates that this would improve clinical generalization."],"forward_implications":["A standard 2D ultrasound system could provide near-real-time regional stiffness maps during an AAA screening exam, since inference takes about 0.058 seconds per image.","Stiffness, or the ratio of stiffness between wall regions, could be added to maximum diameter as a routine rupture-risk marker without new imaging hardware.","Training data can be generated entirely by simulation, avoiding the need for large collections of clinically measured modulus ground truth.","The deep learning approach matches or beats the iterative reconstruction baseline in accuracy on the tested phantoms while being thousands of times faster, making iterative-quality reconstruction feasible in clinical time.","Pressure-normalized displacement inputs allow the same trained network to be applied across different patients and transducers, provided pulse pressure is measured."],"supporting_citations":[{"why":"Provides the pressure-normalized strain concept, phantom construction, transducer point-spread function, and clinical imaging protocol that the FE training and normalization build on.","marker":"[24]"},{"why":"Establishes the U-Net displacement-to-modulus reconstruction framework and the image registration algorithm this work extends.","marker":"[31]"},{"why":"Supplies the iterative soft-prior elasticity reconstruction used as the comparison baseline throughout the experiments.","marker":"[18]"},{"why":"Defines the U-Net architecture adapted as the generator for the deep learning model.","marker":"[41]"},{"why":"Inspires the image-to-image generator design used in the network.","marker":"[27]"},{"why":"Documents the clinical AAA ultrasound protocol and expert vessel segmentation used for the clinical data.","marker":"[37]"},{"why":"Provides the open-source mesh generator used to create the finite-element registration meshes.","marker":"[38]"}],"fun_headline_variants":["Deep learning maps aortic stiffness from 2D ultrasound","AI turns ultrasound into aneurysm wall stiffness maps","Simulation-trained deep net predicts AAA stiffness on ultrasound","Neural network estimates AAA wall stiffness from 2D ultrasound","Rapid elasticity imaging of aortic aneurysms via deep learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's success beyond simulation rests on the assumption that the displacement patterns seen in real ultrasound exams are sufficiently close to the noiseless, plane-strain, angular-only-variation finite-element simulations used for training, so that what the network learned in simulation transfers to clinical measurements.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning maps aortic stiffness from 2D ultrasound","AI turns ultrasound into aneurysm wall stiffness maps","Simulation-trained deep net predicts AAA stiffness on ultrasound","Neural network estimates AAA wall stiffness from 2D ultrasound","Rapid elasticity imaging of aortic aneurysms via deep learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1702,"prompt_tokens":1066,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":558}},"tokens_in":682,"tokens_out":636,"duration_ms":7024,"temperature":1.0,"reasoning_tokens":558,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:57:54.250290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a physical phantom or ex-vivo aortic sample with an independently measured shear modulus, from mechanical testing or MR elastography, image it with the same ultrasound protocol, and compare the DL-predicted modulus to the measured value across a range from about 1 to 170 kPa; if the predicted values do not increase monotonically with true stiffness or deviate from truth by more than a clinically acceptable margin in a substantial fraction of cases, the claim that the FE-trained model generalizes would be refuted.","supporting_citations":[{"cited_title":"Detecting regional stiffness changes in aortic aneurysmal geometries using pressure-normalized strain,","cited_arxiv_id":null,"evidence_quote":"Provides the pressure-normalized strain concept, phantom construction, transducer point-spread function, and clinical imaging protocol that the FE training and normalization build on."},{"cited_title":"Ultrasound elastic modulus reconstruction using a deep learning model trained with simulated data,","cited_arxiv_id":null,"evidence_quote":"Establishes the U-Net displacement-to-modulus reconstruction framework and the image registration algorithm this work extends."},{"cited_title":"Investigating the impact of spatial priors on the performance of model-based ivus elastography,","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative soft-prior elasticity reconstruction used as the comparison baseline throughout the experiments."},{"cited_title":"Intermediate pressure-normalized principal wall strain values are associated with increased abdominal aortic aneurysmal growth rates,","cited_arxiv_id":null,"evidence_quote":"Documents the clinical AAA ultrasound protocol and expert vessel segmentation used for the clinical data."},{"cited_title":"Gmsh: a three-dimensional finite element mesh generator with built-in pre- and post-processing facilities,","cited_arxiv_id":null,"evidence_quote":"Provides the open-source mesh generator used to create the finite-element registration meshes."}],"review_version":1}