{"id":"a019282f-c83a-4657-b504-94418dc68850","arxiv_id":"2507.05783","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A biomechanics-informed registration model extracts strain-derived features that, together with volumetric features from ground-truth segmentations, classify ACDC cardiac diseases with 100% reported test accuracy.","lead":"This paper pairs deep-learning image registration with a physics-based penalty that mimics heart muscle stiffness, then uses the resulting deformation and strain features to classify five cardiac conditions on the ACDC MRI benchmark. The authors report 100% test accuracy, but the classifier relies on expert-drawn segmentations at test time, so the result is not an automatic diagnosis pipeline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 100% test accuracy is not secured: volumetric and label-based features are computed from ground-truth segmentations at test time, so the classifier is oracle-assisted rather than an automatic or demonstrably explainable diagnostic pipeline.","rationale":"The reader's weakest assumption identifies the most load-bearing concern: the classification pipeline uses ground-truth segmentations at test time, so the reported 100% accuracy is not a property of a complete automatic diagnostic system. I agree with this assessment. The paper's own text is explicit that volumetric features come from ground-truth segmentation and that biomechanical features come from predefined anatomical labels, but no automatic segmentation module is integrated or evaluated for the classification task. Because the ground-truth masks encode chamber geometry, volumes, and cavity boundaries, and because those volumetric ratios are already standard disease indicators, the classifier's perfect test performance could be driven by oracle-supplied geometry rather than by the proposed biomechanical registration and strain features. Other concerns, such as the mismatched comparison with nnU-Net and the lack of independent strain validation, are secondary; they affect the strength of the surrounding claims but do not undermine the headline as directly as the test-time oracle does. The proposed concrete test, rerunning classification with automatic segmentations, is feasible and would settle whether the claimed accuracy transfers to a realistic input setting. I find no reason to move the reader's rejection: the paper would need an automatic segmentation module, matched baselines, and clear ablations before its central claim is supported.","tokens_in":11340,"tokens_out":4522,"duration_ms":55934,"concrete_test":"Fix the trained LR classifier and the selected feature set from Table 4. Recompute all test-set features using automatic segmentations, either the multi-frame registration labels from Section 2.3 or a pretrained nnU-Net, instead of ground-truth masks, and report test accuracy. If accuracy drops below 100% or below the best previously published ACDC classifier, the headline result depends on test-time ground truth. As a secondary check, rerun the same comparison using only the mu/kappa biomechanical features, excluding volumetric ratios, to isolate whether the biomechanical features alone carry any discriminative signal without oracle geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.4 states that volumetric features are extracted from the ground-truth segmentation at ED and ES, and that biomechanical features are computed within six anatomical labels shown in Figure 2. Table 4 lists features such as LV/RV volumetric ratios and label-region mu/kappa statistics, all of which require the ground-truth masks at test time. No automatic segmentation module is described for inference; the registration-based label propagation from Section 2.3 is not used to replace ground truth in the classification pipeline. Therefore, the reported 98% training and 100% test accuracy is achieved with oracle access to disease-relevant geometry at test time. The central claim that biomechanics-informed features enable accurate, explainable CVD identification from cine-MRI is not established for raw image input: the classifier may simply be encoding disease-correlated chamber geometry supplied by the ground-truth labels, and the biomechanical features are not independently shown to carry the discriminative signal. This is a validity concern rather than a stylistic one, because substituting automatic segmentations would likely degrade the reported accuracy and would at minimum invalidate the 100% headline as a property of a fully automatic system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a cascaded deep-learning image registration model regularized by a Neo-Hookean strain-energy term and applies it to cine-MRI from the ACDC dataset. The registration is evaluated by propagating ground-truth labels and comparing Dice scores with nnU-Net. The authors then extract features from locally defined shear and bulk moduli within six anatomical labels, together with volumetric features, and use a greedy feature-selection procedure with five classifiers. Logistic regression reportedly achieves 98% training and 100% test accuracy. The paper claims that this constitutes a biomechanics-informed, explainable cardiovascular disease identification pipeline.","tokens_in":11575,"tokens_out":7044,"duration_ms":77259,"significance":"If the claims held, the work would be significant: it would connect physics-informed registration with downstream classification and propose interpretable biomechanical features for cardiac diagnosis. The multi-frame label-propagation strategy and the per-feature ablation analysis are useful elements. However, the current evidence does not support the headline claims. Classification is oracle-assisted by ground-truth segmentations at test time, the so-called biomechanical parameters are algebraic rescalings of the regularization terms, and the registration comparison is against a different task. These are load-bearing validity issues, not presentation problems. The paper also provides no code or detailed training configuration, which further limits reproducibility.","major_comments":[{"comment":"The classification features are computed from ground-truth segmentations at test time: volumetric features are \"extracted from the ground-truth segmentation at ED and ES,\" and biomechanical features are computed within six anatomical labels. No automatic segmentation module is described for inference, and the registration-based multi-frame label propagation of Section 2.3 is not used to produce test-time masks for the classifier. The reported 98% training and 100% test accuracy is therefore obtained with oracle access to disease-relevant chamber geometry, so the abstract's claim of an automatic diagnostic system from cine-MRI is not supported. To make the claim load-bearing, the authors would need to run the full pipeline with automatically produced segmentations and report classification performance on those.","section":"Section 2.4 and Table 4"},{"comment":"Equations (23)-(25) define mu(p) and kappa(p) as the neighborhood average of Phi_dis (resp. Phi_vol) divided by the local value, multiplied by the global constants mu and kappa. Since Phi_dis and Phi_vol are exactly the two terms of the Neo-Hookean regularization in Eq. (21), the extracted \"biomechanical parameters\" are by construction deterministic functions of the model's own objective and of the assumed literature values mu = 2 kPa and kappa/mu = 50. They are not measured tissue properties, and the classification features therefore do not independently establish that the model has learned biomechanical characteristics of the tissue. At minimum, the manuscript should rename these quantities (e.g., \"deformation-energy features\") and temper the physical-interpretation claims.","section":"Section 2.2.3, Eqs. (23)-(25)"},{"comment":"The registration evaluation compares Dice of propagated ground-truth labels (propagation and multi-frame) against the Dice of nnU-Net's direct segmentation output. These are different tasks: label propagation transforms existing expert masks, while nnU-Net predicts segmentations from intensities. The table can only support a claim that propagated labels achieve certain Dice, not that the proposed registration is more accurate than a state-of-the-art segmentation method. A registration baseline (e.g., VoxelMorph or a classical DIR method) evaluated by the same label-propagation protocol is needed for the stated comparison.","section":"Section 3.1, Table 2"},{"comment":"Algorithm 1 repeatedly computes \"classification accuracy acc\" without specifying which split is used, and Table 4 reports \"Validation\" and \"Testing\" columns although no validation set is defined in the dataset description (which gives 100 training and 50 test cases). If the feature selection or model selection used the test set, the reported accuracies would be optimistically biased. The authors must specify the exact data splits used for selection and for reporting Table 4 and Figure 4, and ideally perform nested cross-validation.","section":"Algorithm 1 and Section 3.2"}],"minor_comments":[{"comment":"The abstract reports Dice scores of 0.945, 0.908, and 0.905 for LV, RV, and myocardium, but Table 2's multi-frame results are 0.974/0.956, 0.912/0.857, and 0.960/0.922; the abstract numbers do not match any row of the table.","section":"Abstract and Table 2"},{"comment":"Headings such as \"phi 1,3 ratio ED\" and \"V all ratio ES\" are not defined; a feature glossary or formula list is needed to understand which quantities are being ablated.","section":"Table 4"},{"comment":"Calling 100% test accuracy on 50 patients \"unprecedented\" is overstated; the corresponding confidence interval is wide, and the claim should be tempered.","section":"Section 3.2"},{"comment":"The abstract states that local strains are estimated, but the methods only compute Neo-Hookean energy terms; the strain tensor is never defined or reported, so the wording should be aligned with the actual computation.","section":"Section 2.2.3"},{"comment":"The paper does not provide code, learning rates, patch sizes, the number of cascaded networks, or the training epochs used for the ACDC experiment; Section 2.1 refers to prior work, but the specific configuration for this study should be reported.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The stress-test concern is valid: the headline classification accuracy depends on ground-truth labels at test time, and the biomechanical features are rescaled versions of the regularizer. These are not local presentation issues; they affect the central claims of automatic, explainable diagnosis. Substantial new experiments with automatic segmentation and a redefinition of the features would be needed before the paper could support its conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chris, here's the quick take on Comte et al. The headline number is 100% test accuracy on ACDC, and it's not reliable in the way it reads. The classifier gets ground-truth segmentations at test time for the volumetric features and for the label regions where the biomechanical features are computed. So the method is oracle-assisted, not a fully automatic pipeline from raw cine-MRI. The paper also has a concrete internal inconsistency: the abstract reports Dice 0.945/0.908/0.905, but Table 2's multi-frame numbers are 0.974/0.912/0.960 and 0.956/0.857/0.922, and no line matches the abstract.\n\nWhat's genuinely new: the combination of Neo-Hookean-regularized cascaded registration with subsequent strain-based classification. Prior work either registers with biomechanical constraints without classifying, or classifies from static/radiomic features. The registration framework seems sensible, and the idea of using strain-energy terms as features is worth exploring. The local moduli derivation is a bit odd—they define mu(p) and kappa(p) by rescaling the energy terms, so the features are algebraically tied to the regularizer—but it's not crazy as a feature construction.\n\nThe soft spots are significant. The registration comparison to nnU-Net is apples-to-oranges: they compare label propagation (using ground-truth labels on the target) to automatic segmentation. The feature selection is a greedy forward-backward on the training set, and with 100 training samples and high-dimensional features, the perfect test accuracy smells like selection bias; they don't report cross-validation or confidence intervals on the 100%. The Dice mismatch is disqualifying for the abstract.\n\nBottom line: the paper is a reasonable proof-of-concept for the registration part, but the classification claim as stated is not supported. A serious referee should see it, mainly to push the authors to either run automatic segmentation at test time, match baselines properly, and fix the reporting. I'd like to see a revised version.","headline":"The 100% test accuracy is not what it seems: the classifier is fed ground-truth segmentations at test time, and the paper's own Dice numbers don't match the abstract.","tokens_in":12082,"tokens_out":2545,"would_cite":false,"duration_ms":27945,"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 biomechanics-informed registration pipeline extracts local shear and bulk moduli from cine-MRI and, combined with volumetric features, classifies five cardiac conditions with 100% accuracy on the ACDC test set.","keywords":["cardiac cine-MRI","deformable image registration","Neo-Hookean biomechanics","cardiovascular disease classification","strain-based features","explainable AI","ACDC dataset"],"falsifier":"Replace the ground-truth segmentations on the ACDC test set with segmentations produced by an automatic model, then rerun the same feature extraction and logistic-regression pipeline; if test accuracy drops from 100%, the perfect classification depends on test-time ground-truth masks rather than on the biomechanical features alone.","tokens_in":11141,"feed_emoji":"🫀","tokens_out":10612,"duration_ms":103208,"temperature":0.7,"pith_summary":"The paper tries to establish that cardiac disease can be classified from the motion of the heart rather than only from its shape. It builds a deep-learning registration pipeline that tracks how cine-MRI frames deform between end-diastole and end-systole, regularizes the deformation field with a Neo-Hookean strain-energy term, and converts the resulting deformation into local estimates of shear modulus, bulk modulus, and strain. These biomechanical markers, combined with volumetric features from expert segmentations, are fed to a logistic-regression classifier. On the ACDC benchmark, the pipeline reaches 98% training accuracy and 100% held-out test accuracy across five cardiac conditions. If that result survives independent testing, it would mean motion-derived mechanical features carry diagnostic information that static anatomy alone misses, and that the predictions come with a biomechanical explanation doctors can inspect.","feed_headline":"Heart-motion biomechanics hit 100% test accuracy across five diseases","feed_subtitle":"Physics-informed registration extracts tissue strain markers; logistic regression then aces the held-out ACDC test set.","key_machinery":"The load-bearing object is the Neo-Hookean strain-energy regularizer $\\Phi(\\varphi)=\\frac{\\mu}{2}(\\hat I_1^C-3)+\\frac{\\kappa}{2}(J-1)^2$, where $\\hat I_1^C$ is the deviatoric first invariant of the right Cauchy-Green tensor and $J=\\det(F)$ is the local volume change. From the predicted deformation field $\\varphi$, the model forms the deformation gradient $F$, splits its energy into distortional and volumetric parts, and inverts the strain-energy expression voxel-wise to produce local shear modulus $\\mu(p)$ and bulk modulus $\\kappa(p)$ estimates. Those per-voxel moduli, summarized by mean, standard deviation, and 10th/90th percentiles inside six anatomical labels, plus volumetric features from ground-truth segmentations, form the classification feature set. The regularizer does double duty: it constrains the registration to nearly volume-preserving, physically plausible motion and simultaneously produces the mechanical features used downstream for diagnosis.","core_discovery":"The central claim is that biomechanics-informed image registration converts cardiac cine-MRI into physically interpretable features that separate normal hearts, myocardial infarction, dilated cardiomyopathy, hypertrophic cardiomyopathy, and abnormal right ventricle with perfect held-out accuracy on the ACDC dataset. The authors assert that embedding Neo-Hookean hyperelastic energy as a regularizer in a cascaded deformable-registration network yields physiologically plausible deformation fields (Dice 0.974/0.956 for LV, 0.912/0.857 for RV, 0.960/0.922 for myocardium under multi-frame propagation), and that local shear and bulk moduli estimated voxel-wise from those fields, added to volumetric features, give logistic regression a training accuracy of 0.98 and test accuracy of 1.0. They further claim the selected biomechanical features are not redundant with volume: removing them lowers accuracy, and each contributes above-chance accuracy alone, which supports the paper's thesis that cardiac mechanics carries classification-relevant signal.","pith_inferences":["The paper leaves untested whether the 100% test figure survives when ground-truth segmentations are replaced by automatic ones; the volumetric features encode disease-related chamber geometry, so a segmenter that errs on diseased hearts could lower the real-world ceiling.","The fixed material constants $\\mu = 2$ kPa and $\\kappa/\\mu = 50$ are a hidden sensitivity knob; varying them could reveal whether classification relies on relative strain patterns or on absolute moduli, and the authors' proposed learnable, data-driven parameters could turn the pipeline into a stiffness-mapping diagnostic.","The same registration-plus-strain pipeline could be pointed at other dynamic imaging tasks, such as left-atrial strain in atrial fibrillation or motion tracking in other organs, wherever a deformation field can be estimated and tissue mechanics are disease-relevant; the paper stops at cardiac ventricles.","Because feature selection was performed on the same 100-subject training set whose accuracy is reported, the specific feature ranking may be tuned to this benchmark; an external-cohort replication is needed before treating the list as general."],"forward_implications":["If the 100% test accuracy holds up, biomechanical features derived from registered motion can serve as diagnostic markers for five cardiac conditions with a transparent mechanism: each prediction is traceable to local stiffness and strain patterns rather than to a black-box image statistic.","The registration results imply that a single physics-regularized network can propagate segmentations across cardiac phases with Dice scores comparable to or better than a dedicated segmentation network for the left ventricle and myocardium, enabling temporal label propagation without retraining.","Physicians could inspect which anatomical region and which modulus drove a given classification, because the selected features are named mechanical quantities, such as the 10th percentile of $\\mu$ in a specific myocardial segment at end-systole, rather than latent activations.","Because the model maintains accuracy above 0.9 with training sets as small as 60 subjects, the feature set may transfer to smaller clinical cohorts where deep classifiers usually overfit."],"supporting_citations":[{"why":"Supplies the ACDC benchmark with 150 patients, five disease classes, and expert segmentations used for training and test.","marker":"Bernard et al. (2018)"},{"why":"Contributes the cascaded registration architecture that the paper adapts with the Neo-Hookean regularizer.","marker":"Comte et al. (2025)"},{"why":"Provides the local cross-correlation similarity loss that the registration objective combines with biomechanical regularization.","marker":"Balakrishnan et al. (2019)"},{"why":"Motivates the volumetric/distortional decomposition of the deformation gradient that defines the Neo-Hookean strain energy.","marker":"Bonet et al. (2000)"},{"why":"Gives the literature-based shear modulus $\\mu = 2$ kPa used to set the biomechanical regularization constants.","marker":"Pislaru et al. (2014)"},{"why":"Pioneers biomechanics-informed registration of cardiac MRI, the line of work this paper extends toward disease classification.","marker":"Qin et al. (2020)"},{"why":"Provides the prior ACDC classification pipeline whose static-feature accuracy this paper aims to beat.","marker":"Isensee et al. (2018)"},{"why":"Supplies the nnU-Net segmentation benchmark that the multi-frame label-propagation Dice scores are compared against.","marker":"Isensee et al. (2021)"}],"fun_headline_variants":["Physics-informed heart motion model achieves 100% test accuracy","Biomechanics AI reads cardiac strain, scores perfect on ACDC","Explainable AI uses tissue mechanics to hit 100% on ACDC","Heart strain features predict CVD with perfect test score"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's claimed 100% test accuracy assumes that expert segmentations of the heart chambers are available at test time, because the volumetric features and the six anatomical regions used to summarize the biomechanical features are computed from ground-truth masks.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed heart motion model achieves 100% test accuracy","Biomechanics AI reads cardiac strain, scores perfect on ACDC","Explainable AI uses tissue mechanics to hit 100% on ACDC","Heart strain features predict CVD with perfect test score"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000696,"raw_usage":{"total_tokens":3187,"prompt_tokens":1028,"completion_tokens":2159,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":2089}},"tokens_in":644,"tokens_out":2159,"duration_ms":16144,"temperature":1.0,"reasoning_tokens":2089,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:18:32.491309+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the ground-truth segmentations on the ACDC test set with segmentations produced by an automatic model, then rerun the same feature extraction and logistic-regression pipeline; if test accuracy drops from 100%, the perfect classification depends on test-time ground-truth masks rather than on the biomechanical features alone.","supporting_citations":[{"cited_title":"Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Transactions on Medical Imaging 2018;37(11):2514--2525","cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC benchmark with 150 patients, five disease classes, and expert segmentations used for training and test."},{"cited_title":"Deep cascaded registration and weakly-supervised segmentation of fetal brain MRI","cited_arxiv_id":null,"evidence_quote":"Contributes the cascaded registration architecture that the paper adapts with the Neo-Hookean regularizer."},{"cited_title":"VoxelMorph: a learning framework for deformable medical image registration","cited_arxiv_id":null,"evidence_quote":"Provides the local cross-correlation similarity loss that the registration objective combines with biomechanical regularization."},{"cited_title":"Finite element analysis of air supported membrane structures","cited_arxiv_id":null,"evidence_quote":"Motivates the volumetric/distortional decomposition of the deformation gradient that defines the Neo-Hookean strain energy."},{"cited_title":"Viscoelastic properties of normal and infarcted myocardium measured by a multifrequency shear wave method: comparison with pressure-segment length method","cited_arxiv_id":null,"evidence_quote":"Gives the literature-based shear modulus $\\mu = 2$ kPa used to set the biomechanical regularization constants."},{"cited_title":"Biomechanics-informed neural networks for myocardial motion tracking in MRI","cited_arxiv_id":null,"evidence_quote":"Pioneers biomechanics-informed registration of cardiac MRI, the line of work this paper extends toward disease classification."},{"cited_title":"nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the nnU-Net segmentation benchmark that the multi-frame label-propagation Dice scores are compared against."}],"review_version":1}