{"id":"ddcaa01c-b18e-42a9-9583-291e7fe5798f","arxiv_id":"2506.20696","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"IMC-PINN-FE estimates left-ventricular stiffness and active tension from images, then runs a physics-constrained neural FE simulation of the cardiac cycle about 75x faster than standard FE while matching imaged volumes.","lead":"The authors built a physics-informed neural network that estimates heart-muscle stiffness and contraction strength from MRI or ultrasound, then runs a fast patient-specific heart simulation that follows the images. It aims to turn hours of heart modeling into minutes, which could make personalized cardiac modeling practical in clinics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed Dice gain is evaluated against the same EFDL tracker that supplies the POD motion modes, so the headline motion-fidelity result is not an independent validation of matching imaged motion.","rationale":"The reader's weakest assumption is the load-free reference state; that is a genuine and acknowledged limitation, but it mainly affects the interpretation of stiffness and tension values. The more load-bearing issue for the paper's headline is the motion-fidelity evaluation. The pipeline uses EFDL to track motion, builds its POD basis from those tracked meshes, and then evaluates Dice against meshes propagated by the same EFDL. This is a circular benchmark: a reduced-order model whose basis spans the tracker output will naturally score high against that tracker. EFDL's own agreement with manual segmentations is only 0.829 on adult MRI and 0.802 on fetal echo, so the image-derived ground truth is not an independent gold standard. The authors do acknowledge a general dependency on tracking accuracy, but they do not flag that the central quantitative comparison is self-referential. The test above would settle the concern with existing data and code. The method may well be sound, and the speed claim is not affected by this issue; the reported 0.927 Dice and the motion-consistency improvement, however, are not yet independently supported. I therefore keep the reader's conditional verdict rather than changing it.","tokens_in":20708,"tokens_out":6748,"duration_ms":75695,"concrete_test":"Recompute the Section 3.3.2 Dice evaluation on the same six cases using independent ground-truth references: ACDC myocardium masks for adult MRI and manual LazySnap segmentations for fetal echo, reconstructed into 3D meshes frame-by-frame without using EFDL. Apply the open-source IMC-PINN-FE and heartFEM outputs to this reference and compare Dice. If the IMC-PINN-FE advantage over FE vanishes or falls to the EFDL tracking error level, roughly 0.80 to 0.83 against manual masks, then the claimed 0.927 motion fidelity and the central motion-consistency improvement are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3.2 states that the ground truth for imaged cardiac motion is defined as 4D LV mesh reconstructions propagated across the cardiac cycle using EFDL, chosen for its highest accuracy. Section 2.2.2 and Section 2.2.3 build the POD motion modes from the same single-subject tracked sequence, and the volume targets also come from the tracked meshes. Section 3.3.2 then measures IMC-PINN-FE Dice against that EFDL-propagated ground truth. The 0.927 versus 0.849 comparison therefore shows how well the PINN-FE solver reproduces the output of the tracker that generated its reduced-order basis, not how well it matches the images themselves. Table 1 shows EFDL itself has DSC 0.829 on adult MRI and 0.802 on fetal echo against manual segmentations, so the propagated meshes are not a gold standard. The abstract claim that IMC-PINN-FE matches imaged displacements more accurately is consequently unvalidated on real clinical images. This does not imply the method is incorrect, but the central motion-fidelity claim currently rests on a self-referential benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IMC-PINN-FE, a three-stage pipeline for patient-specific left-ventricular biomechanics: (i) deep-learning motion tracking (a new unsupervised method, EFDL, or an existing co-attention STN) that produces 3D+t surface meshes; (ii) a PINN-based estimator that back-computes myocardial stiffness (C_stiffness) and maximal active tension (T_max) from end-diastolic and peak-systolic states; and (iii) a volume-constrained PINN-FE solver that runs a full cardiac-cycle FE simulation at about 75x speedup relative to a traditional volume-constrained FE model, using POD motion modes extracted from a single subject's tracked motion. The authors report parameter estimates within a few percent relative error on FE-generated validation cases, Dice similarity of 0.927 versus 0.849 for the PINN-FE solver compared with traditional FE, and P-V loops close to those of the FE reference.","tokens_in":20941,"tokens_out":4932,"duration_ms":53068,"significance":"If the central claims hold, this would be a meaningful step toward near-real-time, patient-specific cardiac biomechanical modeling: the pipeline is applicable to both MRI and echocardiography, does not require a large multi-subject training set, releases code, and demonstrates a large runtime reduction. The FE-based parameter-estimation validation in Table 2 is a correct self-consistency check and is a useful proof of concept. However, the headline motion-fidelity claim (Dice improvement and 'matches imaged displacements') is not independently validated because the same EFDL tracking that builds the POD basis and volume targets is also used as the ground truth for evaluation. In addition, the physics loss as written enforces zero total potential energy rather than equilibrium, which is a more fundamental concern. The load-free reference-state assumption is acknowledged but not tested for sensitivity, which affects all downstream parameter and P-V estimates. These issues are fixable in revision but are load-bearing for the paper's main claims.","major_comments":[{"comment":"The Dice improvement reported in Section 3.3.2 is evaluated against a ground truth that is itself produced by the same EFDL tracking that supplies the POD motion modes (Section 2.2.2) and the volume targets (Section 2.3). Section 2.3.2 states that the ground truth for imaged motion is '4D LV mesh reconstructions propagated across the cardiac cycle using EFDL', and Table 1 shows EFDL itself has DSC 0.829 on adult MRI and 0.802 on fetal echo against manual segmentations. Thus the abstract claim that IMC-PINN-FE 'matches imaged displacements more accurately' is not supported by an independent reference; the comparison primarily shows how well the solver reproduces the tracker that generated its reduced-order basis. I recommend adding an evaluation against manual segmentations (at least at ED and ES) or against an independent tracker (e.g., the co-attention STN), and reporting Dice on those references.","section":"§3.3.2 / §2.3.2 / §2.2.2 / §2.3"},{"comment":"The 'governing equation' loss is defined as the squared total potential energy Π = ∫Ωmyo W dV - ∫Ωendo P·u dA. The equilibrium condition for the hyperelastic problem is the stationarity of Π with respect to the displacement field, δΠ = 0, not Π = 0. Enforcing Π^2 → 0 is a scalar energy-balance constraint, not the weak form of force equilibrium, and it can be satisfied by displacement-pressure pairs that violate local equilibrium. This calls into question the 'physics-informed' aspect of the method. The authors should replace this loss with a proper residual of the weak form (or train the network to minimize Π over admissible displacements), and then re-run the validation experiments.","section":"§2.2.3, Eq. (6) and §2.3.1, Eq. (18)"},{"comment":"The load-free reference state is assumed to be the cardiac state at 1/3 into the diastolic duration. All displacements, Green-Lagrange strains, C_stiffness and T_max estimates, and P-V loops are computed relative to this geometry. The paper acknowledges the assumption is 'imperfect', but no sensitivity analysis is provided. If the true unloaded state differs, the estimated parameters and the entire P-V loop inherit a systematic bias. I request a sensitivity study that varies the reference frame (e.g., using frames at 1/4, 1/3, and 1/2 of diastole, or an estimated unloaded state) and reports the resulting changes in C_stiffness, T_max, and the P-V loop.","section":"§2.2, step 2"},{"comment":"The FE-based validation in Table 2 is a self-consistency check: the estimator recovers parameters prescribed in noise-free simulations of the same constitutive model. This demonstrates identifiability and internal consistency, but it does not validate the estimated parameters on real clinical data. For the two real-image cases, agreement is reported only as 'within' a literature range, which is a weak and vague criterion given the wide reported ranges. The abstract and discussion should be tempered accordingly; as it stands, the phrase 'rapidly estimates myocardial stiffness and active tension' overstates the level of clinical validation actually provided.","section":"§2.2.5 / Table 2"}],"minor_comments":[{"comment":"The text says 'the first 10 dominant modes are retained' but Eq. (5) and the surrounding sentence say 'we use the first 20 bases'. Please state the exact number used in the experiments and ensure consistency throughout.","section":"§2.2.2"},{"comment":"The deformation gradient is written as F = ∂x/∂X = I + ∇u·u. This appears to be a typo; with the displacement u, the standard expression is F = I + ∇u (where ∇u is the displacement gradient). If the dotted form is actually implemented, the constitutive model would be incorrect.","section":"§2.2.3"},{"comment":"There are occasional typos, including 'EDFL' for 'EFDL' in Section 2.1.1 and 'enfocing' in Section 3.3.2. A careful proofread is needed.","section":"§3.3.2"},{"comment":"For the image-tracking cases, reporting only 'within' a literature interval is not informative. I suggest reporting the numerical value and the width of the interval, or a percentage deviation from the interval midpoint, so the reader can judge the closeness.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the code release is a plus, but the circular motion-fidelity evaluation and the questionable physics loss need to be addressed before this is acceptable. The physics loss issue is potentially the most serious: if Π=0 is enforced instead of δΠ=0, the method may not actually be solving the equilibrium equations, which would affect both the parameter estimates and the claimed physiological consistency. The FE-based parameter validation is a good start but is only a self-consistency test; the real-data validation is thin. I would advise the editor that the revision should be evaluated with particular attention to the new physics-loss formulation and to an independent motion-consistency benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core machinery is worth engaging with: single-subject POD of intra-subject dynamics for a reduced-order basis, explicit image-volume constraints, and PINN-based back-estimation of stiffness and active tension. That combination is not in Buoso or Motiwale, and the EFDL tracker is a sensible hybrid. Code is available, the 75x speedup over traditional FE is real, and the FE-based parameter estimation recovers prescribed values with <6% error—a correct self-consistency check.\n\nNow the soft spots, in proportion. The stress-test note holds up exactly. Section 2.3.2 defines the ground truth for motion as EFDL-propagated meshes, while Section 2.2.2 builds the POD modes and volume targets from the same EFDL tracking. So the Dice improvement (0.849 to 0.927) shows the PINN-FE reproducing the tracker that generated its reduced-order basis, not matching the images independently. The authors acknowledge tracking dependence in the limitations section, but the abstract's claim of \"matching imaged displacements more accurately\" is not actually validated on real clinical images. That is a load-bearing issue for the paper's headline, not a fatal one—the method could still work, but the comparative claim needs independent ground truth (e.g., manual segmentations, speckle tracking, or a second registration method).\n\nOther soft spots are minor by comparison. The load-free state is assumed to be the geometry at one-third of diastole; the paper says this is imperfect and consistent with prior work, but all stiffness, tension, and strain values inherit whatever bias that assumption carries. The real-clinical validation for parameter estimation is only 2 cases, and the \"ground truth\" there is literature values the paper itself calls \"not ideal.\" The P-V loop comparison is on 2 representative cases. The lack of a Windkessel model and the absence of explicit isovolumetric phase demonstration are acknowledged and are not deal-breakers.\n\nWho gets value from this? Anyone working on cardiac digital twins or fast patient-specific FE, especially in data-limited settings. The paper is a legitimate methodological advance with a clear engineering payoff, and it deserves a serious referee. The main revision demands are: revalidate motion fidelity against an independent reference, add uncertainty quantification, and show more than 2 real clinical cases for the full pipeline. As a reviewer I'd lean major revision rather than reject.","headline":"A genuinely useful PINN-FE package with a real speedup, but the headline motion-fidelity gain is measured against the same tracker that built the model, so don't take the 0.85-to-0.93 Dice at face value.","tokens_in":21464,"tokens_out":2093,"would_cite":true,"duration_ms":22783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed neural network recovers myocardial stiffness and active tension from clinical images in about one minute, then runs whole-cycle heart simulations 75 times faster than conventional FE while matching imaged motion closely.","keywords":["physics-informed neural network","cardiac biomechanics","left ventricular modeling","finite element simulation","image motion tracking","myocardial stiffness estimation","active tension","proper orthogonal decomposition"],"falsifier":"The decisive test is to apply the pipeline to a heart whose true unloaded geometry is measured directly, for example by imaging an explanted heart at zero cavity pressure, alongside its in vivo imaging; then compare $C_{\\text{stiffness}}$ and $T_{\\max}$ recovered under the one-third-diastole reference assumption with values obtained by fitting the same constitutive model from the measured unloaded state. A systematic offset between the two would identify the reference-state assumption, which the paper itself flags as imperfect, as the dominant error source in the parameter estimates.","tokens_in":20517,"feed_emoji":"🫀","tokens_out":14339,"duration_ms":129578,"temperature":0.7,"pith_summary":"This paper proposes IMC-PINN-FE, a physics-informed neural network framework intended to make patient-specific modeling of the beating left ventricle fast enough for clinical use. Its central claim is that myocardial stiffness and peak active tension can be back-computed from image-derived motion plus a few pressure measurements in roughly one minute, where iterative finite element fitting takes hours. The paper further claims that the same solver reproduces the full cardiac cycle about 75 times faster than conventional finite element analysis while matching imaged heart motion better (average Dice 0.927 versus 0.849) and preserving realistic pressure-volume loops. If correct, this would let clinicians obtain a patient's myocardial material properties and simulated heart function in near real time from routine MRI or echocardiography, and would make cardiac digital twins feasible at the point of care.","feed_headline":"Heart muscle properties estimated in seconds, not hours","feed_subtitle":"Back-computes stiffness and active tension from MRI or echo in a minute, then runs the heart cycle 75x faster.","key_machinery":"The central objects are the subject-specific POD motion modes: the first ten Proper Orthogonal Decomposition basis functions of a single patient's tracked displacement fields across the cardiac cycle, which encode deformation compactly and stabilize PINN training. On top of these modes, the framework stacks an energy-balance loss, namely the difference between myocardial strain energy (exponential passive plus calcium-driven active terms) and the work done by cavity pressure on the endocardium, minimized by the two parameter-estimator networks and by the volume-constrained whole-cycle solver, which also enforces imaged volumes, end-diastolic and peak-systolic pressures, and monotone pressure-volume slopes.","core_discovery":"IMC-PINN-FE establishes that one pipeline can solve both the inverse problem and the forward problem of left-ventricular biomechanics: motion extracted from 4D images, reduced-order motion encoding, and physics constraints are combined so that material parameters are first estimated, then the whole cardiac cycle is simulated. Motion is tracked by a proposed unsupervised Fourier-regularized registration network (EFDL) or a co-attention network, and is compressed into Proper Orthogonal Decomposition modes taken from a single subject's own cardiac cycle rather than from a population. Two small neural networks then estimate the exponential transversely-isotropic stiffness coefficient $C_{\\text{stiffness}}$ and the maximum active tension $T_{\\max}$ by minimizing the myocardial potential-energy balance at end-diastole and peak systole, recovering values within a few percent of finite element ground truths. The final volume-constrained PINN-FE solver computes pressure and displacement across the cycle from imaged volumes, and deliberately allows the displacements to deviate from the tracked motion so that the solution satisfies physics and image data simultaneously; this is why it claims to beat both conventional FE and a prior PINN-FE that lacks motion consistency.","pith_inferences":["With parameter estimation this cheap, the practical ceiling on accuracy in this pipeline shifts from the mechanics solver to the quality of image-derived geometry and motion tracking; better tracking and explicit uncertainty quantification would likely improve every downstream quantity.","The single-subject POD basis is a natural scaffolding for cheap parameter sweeps: because the motion modes are already patient-specific, one could simulate many stiffness and tension variants of the same heart to generate training data for other deep learning models.","The imperfect load-free reference state, flagged in the paper, implies that a learned predictor of the true unloaded geometry would probably sharpen all pressure-volume and parameter estimates; this is a stated direction of the authors' future work.","Because the volume-constrained solver consumes imaged volumes directly, plugging in a Windkessel or other lumped circulation model (which the paper notes is absent) would extend the framework to settings with partial image coverage or to predictive hemodynamics."],"forward_implications":["Myocardial stiffness and peak active tension can be recovered per patient in about one minute (63 ± 22 seconds per case), replacing iterative inverse finite element fitting that takes several hours per iteration.","Whole-cardiac-cycle simulation drops from roughly 7.5 hours to under 6 minutes on the same hardware, a near-75-fold speedup that makes near-real-time use plausible.","Simulated LV shapes match image-derived motion markedly better (average Dice 0.927 versus 0.849 for conventional FE) while predicted pressure-volume loops stay nearly identical to FE, supporting the claim that image consistency and physical consistency are compatible.","The motion basis comes from a single subject's own cardiac cycle, so no large multi-subject training dataset is needed for motion encoding and patient specificity is improved.","Because the solver uses the same governing equations as traditional FE, the framework can also serve as a fast forward model for building large simulation databases for further deep learning training."],"supporting_citations":[{"why":"The prior parametric PINN-FE whose shape-mode encoding and potential-energy loss this work extends, and the main comparison baseline for the solver.","marker":"[7]"},{"why":"Supplies the potential-energy governing equation and the constitutive framework on which the physics loss is built.","marker":"[44]"},{"why":"Provides the exponential transversely-isotropic passive strain energy and the calcium-activation active stress model.","marker":"[20]"},{"why":"The co-attention spatial transformer network used as the pre-trained motion-tracking option and as a registration baseline.","marker":"[1]"},{"why":"The Fourier-regularized B-spline registration approach that the proposed EFDL tracker hybridizes with a network.","marker":"[10, 49]"},{"why":"SimpleElastix, the pairwise B-spline registration engine that generates EFDL's raw displacement fields.","marker":"[32]"},{"why":"The empirical formula converting end-diastolic volume into end-diastolic pressure for MRI-based cases.","marker":"[9]"},{"why":"The porcine echocardiography dataset, including hypertrophy cases, and the E/E' formula used to estimate end-diastolic pressure.","marker":"[52]"}],"fun_headline_variants":["Heart properties from one scan in seconds","Physics-informed AI speeds heart modeling 75x","Image-consistent heart simulation at 75x speed","Patient-specific heart model from MRI in seconds","From images to heart biomechanics in seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the heart geometry at one third into the diastolic phase is the true zero-pressure, unloaded reference state; every displacement, strain, stiffness, and active-tension value in the pipeline is measured against this presumed state, which the paper acknowledges is imperfect.","fun_headline_variants_meta":{"raw":{"variants":["Heart properties from one scan in seconds","Physics-informed AI speeds heart modeling 75x","Image-consistent heart simulation at 75x speed","Patient-specific heart model from MRI in seconds","From images to heart biomechanics in seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2114,"prompt_tokens":1065,"completion_tokens":1049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":681,"completion_tokens_details":{"reasoning_tokens":981}},"tokens_in":681,"tokens_out":1049,"duration_ms":12163,"temperature":1.0,"reasoning_tokens":981,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:51:15.064992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive test is to apply the pipeline to a heart whose true unloaded geometry is measured directly, for example by imaging an explanted heart at zero cavity pressure, alongside its in vivo imaging; then compare $C_{\\text{stiffness}}$ and $T_{\\max}$ recovered under the one-third-diastole reference assumption with values obtained by fitting the same constitutive model from the measured unloaded state. A systematic offset between the two would identify the reference-state assumption, which the paper itself flags as imperfect, as the dominant error source in the parameter estimates.","supporting_citations":[{"cited_title":"Personalising left-ventricular biophysical models of the heart using parametric physics-informed neural networks","cited_arxiv_id":null,"evidence_quote":"The prior parametric PINN-FE whose shape-mode encoding and potential-energy loss this work extends, and the main comparison baseline for the solver."},{"cited_title":"Co-attention spatial transformernetworkforunsupervisedmotiontrackingandcardiacstrainanalysisin3dechocardiography","cited_arxiv_id":null,"evidence_quote":"The co-attention spatial transformer network used as the pre-trained motion-tracking option and as a registration baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SimpleElastix, the pairwise B-spline registration engine that generates EFDL's raw displacement fields."},{"cited_title":"Assessment of systolic and diastolic ventricular properties via pressure-volume analysis: a guide for clinical, translational, and basic researchers","cited_arxiv_id":null,"evidence_quote":"The empirical formula converting end-diastolic volume into end-diastolic pressure for MRI-based cases."},{"cited_title":"Morphological, functional, and biomechanical progression of lv remodelling in a porcine model of hfpef","cited_arxiv_id":null,"evidence_quote":"The porcine echocardiography dataset, including hypertrophy cases, and the E/E' formula used to estimate end-diastolic pressure."}],"review_version":1}