{"id":"9ee3f6bf-e6f2-4827-82eb-220ebe02fc08","arxiv_id":"2507.18677","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"HeartUnloadNet maps end-diastolic left-ventricular meshes to unloaded geometries in milliseconds on synthetic finite element data, but the reported DSC and sample-efficiency numbers are internally inconsistent.","lead":"HeartUnloadNet predicts the unloaded, zero-pressure left ventricular geometry from a diastolic mesh using a graph neural network with physiological conditioning and cycle-consistent training. On 20,700 synthetic FE simulations it reports sub-millimeter accuracy and 0.02-second inference, but several headline numbers are internally inconsistent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'DSC ≥ 0.986' claim depends on a nonstandard node-level 0.1 mm threshold, and the '97% DSC with 200 samples' claim is numerically inconsistent with Table 4; both should be recomputed before the central accuracy claims are accepted.","rationale":"The reader correctly flags the under-constrained inverse FE ground truth and the internal metric inconsistencies. I agree with CONDITIONAL, but I identify the load-bearing concern slightly differently: the most actionable weakness is the thresholded DSC definition combined with the irreproducible 200-sample claim. The inverse FE ground-truth caveat is real and is explicitly acknowledged in Section 3.3.1, yet within the synthetic experimental scope, the network could still be a valid surrogate for the FE pipeline. The metric issue, however, directly affects the abstract's quantitative claims and can be settled by recomputation. I do not see an internal inconsistency in the cycle-consistency training: Algorithm 1's loss is a standard L2 reconstruction loss and is a plausible weak-supervision mechanism. The architecture, ablations, and code release are credible evidence that the method works on the synthetic distribution; the concern is about the precision and interpretability of the reported numbers, not about whether the method is fundamentally flawed. Hence CONDITIONAL is appropriate: the paper should be accepted only after the metric definition and low-supervision numbers are corrected and clarified.","tokens_in":17879,"tokens_out":1641,"duration_ms":17854,"concrete_test":"Recompute the low-supervision experiment with an exact 200-sample supervision count (and, separately, exact 209-sample count as mentioned in the introduction) and report DSC using the standard voxelized Dice on the unloaded meshes, alongside the current node-level 0.1 mm threshold. If standard DSC at 200 samples falls below 0.97, the abstract's data-efficiency claim is unsupported. Also compute the test meshes' median edge length and the fraction of nodes with error in [0.05, 0.15] mm to determine whether the 0.1 mm threshold is stable or a saturation artifact.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline evidence that HeartUnloadNet is 'significantly more accurate than traditional inverse FE solvers' rests on a node-level DSC defined in Table 2 as the percentage of mesh nodes with Euclidean error < 0.1 mm, in a coordinate frame that is globally normalized to zero mean and unit scale. A 0.1 mm threshold on a scale-normalized mesh may be much smaller than the typical nodal spacing, making DSC a near-saturation count rather than a robust overlap measure; the reported HD of 0.083 cm (0.83 mm) is over eight times the 0.1 mm threshold, so the error distribution is not captured by the thresholded metric alone. The central 'sub-millimeter accuracy' phrasing is also potentially circular: the DSC threshold itself is 0.1 mm, so 'average DSC 0.986' does not independently establish sub-millimeter accuracy. Additionally, the abstract and conclusion state '97% DSC with as few as 200 training samples,' but Table 4 reports 3% supervision with DSC 0.971 and 1% supervision (≈139 samples) with DSC 0.619; 3% of the 13,890 training samples is ≈417, not ≈200, so the claimed 200-sample result is not reproducible from the reported supervision-ratio rows. Because the architecture and cycle-consistency mechanism are otherwise plausible, these metric and numeric inconsistencies are the weakest load-bearing link in the central performance claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HeartUnloadNet is a graph-attention network that predicts the unloaded (zero-pressure) left-ventricular mesh from an end-diastolic mesh and physiological parameters (pressure, stiffness, helix angles). The model is trained on 20,700 paired FE simulations generated by an inverse unloading–reloading pipeline, uses a cycle-consistency loss for weak self-supervision, and is evaluated against GCN, PointNet++, and inverse FE baselines. The paper claims sub-millimeter accuracy (DSC 0.986, HD 0.083 cm), an inference time of 0.023 s (over 10^5 times faster than inverse FE), and that cycle consistency maintains 97% DSC with only ~200 training samples.","tokens_in":18169,"tokens_out":5024,"duration_ms":55343,"significance":"If the reported accuracy holds, this would be a valuable contribution: it is, to the authors' knowledge, the first deep-learning surrogate for unloaded cardiac geometry, it handles variable mesh topology without template registration, and it explicitly conditions on biophysical parameters. The release of code and the construction of a 20,700-case FE simulation dataset are also strengths, as is the systematic ablation study of architectural components. However, the central accuracy claims rest on a metric definition and on a data-efficiency claim that are internally inconsistent, and the ground truth is generated by an inverse FE pipeline whose non-uniqueness is acknowledged. These issues must be resolved before the performance claims can be accepted.","major_comments":[{"comment":"The definition of DSC as the percentage of nodes with Euclidean error below 0.1 mm (Table 2) is inconsistent with the reported MD and SD. For HeartUnloadNet, MD is 0.028 ± 0.010 cm (0.28 ± 0.10 mm) and SD is 0.013 cm (0.13 mm) (Table 3). A mean error of 0.28 mm with 98.6% of nodes below 0.1 mm would require a strongly non-Gaussian error distribution with extreme outliers; the reported SD alone does not support such a distribution without an explicit description. Moreover, the training protocol normalizes all mesh coordinates to zero mean and unit scale (Section 3.4), so it is unclear whether the 0.1 mm threshold is applied to de-normalized physical coordinates or to normalized coordinates. The authors should report the full error distribution (e.g., percentiles), clarify the coordinate frame used for DSC, and avoid implying that a high DSC independently establishes sub-millimeter accuracy when the threshold itself is 0.1 mm.","section":"Table 2, Table 3, Section 3.4"},{"comment":"The claim that the model 'maintains a DSC of 97% even with as few as 200 training samples' is not supported by Table 4. Section 4.2 states that 3% labeled data corresponds to 'around 200 cases,' but 3% of the 13,890 training samples (42 shapes) is approximately 417 cases. The 1% row, corresponding to about 139 cases, yields DSC 0.619 ± 0.251 for the full model (A0). A 200-sample setting would fall between 1% and 3% supervision, with an expected DSC far below 0.97 based on the reported values. The abstract (200 samples), contribution 3 (209 training cases), and Section 4.2 (3%, ~200 cases) are mutually inconsistent. The authors must specify the exact number of labeled samples used in each supervision-ratio experiment and reconcile the text with the numbers in Table 4.","section":"Section 4.2, Table 4, Abstract, Contribution 3"},{"comment":"The ground-truth unloaded geometry is produced by the authors' inverse FE unloading–reloading procedure, which the paper itself notes is 'typically under-constrained and can have non-unique solution.' The network therefore learns to reproduce the output of this particular inverse FE pipeline, not an independently validated zero-pressure state. The claim that HeartUnloadNet is 'significantly more accurate than traditional inverse FE solvers' (Section 4.1) compares the network against a different inverse FE baseline (Finsberg et al.), but both the network's training target and the baseline are outputs of inverse FE methods. This does not make the comparison circular in a formal sense, but it means the reported accuracy is relative to one chosen inverse FE solution, and the clinical-transferability claims are weaker than stated. The authors should explicitly acknowledge this limitation in the Limitations section and, if possible, validate on an independent forward-FE test or clinical data.","section":"Section 3.3.1, Abstract, Section 4.1"}],"minor_comments":[{"comment":"In the sentence 'The only changed global parameters parameters allowed to vary', the word 'parameters' is duplicated; please proofread the manuscript throughout for similar typos.","section":"Section 3.3.1"},{"comment":"The description of weak supervision does not specify how the labeled subset is selected (e.g., random sampling, stratified by parameter) or whether the cycle-consistency loss is applied to all samples or only to unlabeled ones. This information is needed to interpret the supervision-ratio experiments.","section":"Section 4.2"},{"comment":"The LOVO table reports only a subset of the parameter values (three of six for pressure and stiffness). Please state why these particular values were chosen, or report all held-out values, so that the reader can assess the claimed interpolation capability.","section":"Table 7"},{"comment":"The inference time '0.023±0.002 s' in Table 3 differs from the abstract's '0.02 seconds' and the text's '~0.02 s'; please use consistent rounding or state the exact value in all places.","section":"Section 4.1"},{"comment":"Reference [24] appears to be a self-citation to an arXiv preprint; please verify that this citation is appropriate and that all external references are correctly formatted.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The metric inconsistency (Major Comment 1) and the supervision-ratio contradiction (Major Comment 2) are serious enough that the central numerical claims cannot be accepted as written. However, these issues appear fixable by recomputing the metrics with a conventional overlap-based DSC, reporting the full error distribution, and correcting the sample-count descriptions. I recommend major revision rather than rejection, provided the authors can produce internally consistent numbers and clearly qualify the ground-truth dependence on the inverse FE pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper builds a graph-attention network that predicts unloaded LV geometry from ED meshes, conditioned on pressure, stiffness, and fiber angles, with cycle-consistent bidirectional decoders. That is a sensible direction and, as far as I know, the first learning-based unloading for the heart, so the novelty is real. The ablations are extensive and show the architectural choices matter. The speed claim (0.02 s vs hours) is credible for GNN inference.\n\nThe main problems are the metrics and the numbers. The DSC is defined as the percentage of nodes within 0.1 mm, yet MD is 0.028 cm (0.28 mm) and HD is 0.083 cm (0.83 mm). If the mean node error is 0.28 mm, a 0.1 mm threshold should pass very few nodes under a normal-looking error distribution; 0.986 is mathematically implausible unless the distribution is extremely skew-concentrated at zero, which the SD (0.013 cm) does not suggest. The stress-test is right: the headline 'sub-millimeter accuracy' is not established by a thresholded metric at 0.1 mm. Recommend recomputing DSC at a meaningful threshold (e.g., 1 mm) or replacing it with IoU on voxelized masks.\n\nSecond, the abstract and conclusion say '97% DSC with 200 training samples,' but Table 4 shows 3% supervision giving 0.971 and 1% supervision (about 139 samples) giving 0.619. Three percent of 13,890 is about 417, not 200. The claim does not match the table. This is easy to fix but must be corrected before the paper is citable.\n\nThird, the inverse FE baseline looks weak: Finsberg et al.'s method with first-order approximations, not the current state of the art. The claim that the network is 'significantly more accurate than traditional inverse FE solvers' should be hedged to 'than the particular solver we compared.' The PCA baseline is coarse but harmless as an extra.\n\nThe ground-truth generation uses inverse FE unloading-reloading, which the paper admits is under-constrained. The network learns to reproduce that particular solution. That is acceptable for a surrogate benchmark, but the paper should say plainly that accuracy is relative to this synthetic reference, not to an independently verified zero-pressure state. The limitation on healthy shapes only is stated honestly.\n\nOverall: a useful surrogate with solid ablations, but the central performance claims are not yet trustworthy as written. The fixes are small: correct the numbers, redefine DSC, and soften the baseline comparison. I would send it to review with a request for revision, and treat the headline numbers with skepticism until recomputed.","headline":"A genuinely novel GNN surrogate for cardiac unloading with thorough ablations, but the headline DSC and 200-sample claims are numerically inconsistent and need correction before the accuracy numbers can be trusted.","tokens_in":18702,"tokens_out":1853,"would_cite":true,"duration_ms":20546,"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":"HeartUnloadNet claims the zero-pressure left-ventricular geometry can be predicted straight from the end-diastolic mesh in about 0.02 seconds with sub-millimeter accuracy, outperforming inverse finite-element solvers on 20,700 synthetic…","keywords":["cycle-consistent graph networks","cardiac inverse modeling","weakly-supervised learning","zero-pressure geometry","biomechanical shape prediction","left ventricular mechanics","graph attention network","finite element surrogate"],"falsifier":"Take a heart or high-fidelity model for which the unloaded geometry is known independently, for example an excised ventricle imaged at zero cavity pressure and then loaded to a known pressure, and compare the network's predicted unloaded mesh with that measured state; if the node-level overlap or mean distance is substantially worse than the reported 0.986 and 0.028 cm, the claim that the network predicts true unloaded geometry fails.","tokens_in":17640,"feed_emoji":"🫀","tokens_out":9859,"duration_ms":110314,"temperature":0.7,"pith_summary":"The paper tries to show that recovering the unloaded, zero-pressure shape of the left ventricle, a step normally done by slow inverse finite-element optimization, can be replaced by a single forward pass of a graph neural network. On synthetic test geometries it reports a node-level overlap score of 0.986 and an average maximum surface deviation of 0.083 cm, with inference around 0.02 seconds per case, about five orders of magnitude faster than the inverse FE baseline and more accurate. The network accepts an end-diastolic mesh of any topology plus physiological parameters such as pressure, stiffness, and fiber orientation, and outputs the displacement to the unloaded state. A cycle-consistency loop lets it train with far fewer labeled examples, keeping overlap near 0.97 with roughly 200 training cases. If correct, this makes a clinically unobservable reference state available in real time for personalized cardiac biomechanics.","feed_headline":"Network predicts zero-pressure heart shape in 0.02 s","feed_subtitle":"A graph neural network maps loaded ventricles to their unloaded reference with sub-millimeter accuracy.","key_machinery":"The load-bearing mechanism is a graph-attention encoder coupled with two deformation decoders that share the encoder and run in opposite directions. The encoder builds per-node features from mesh coordinates by aggregating neighbor information with learned attention weights, pools a global mesh representation, and fuses it with an encoded vector of the physiological parameters through cross-attention. One decoder predicts end-diastolic-to-unloaded nodal displacements; the other reconstructs the end-diastolic mesh from the predicted unloaded mesh. The training loss is the supervised squared error on the unloaded mesh plus a weighted cycle term that penalizes mismatch between the reconstructed and original end-diastolic mesh, enforcing reversibility of the deformation and providing weak self-supervision.","core_discovery":"HeartUnloadNet's central claim is that inferring the unloaded left-ventricular configuration is a learnable graph-to-graph regression: given an end-diastolic tetrahedral mesh and four physiological parameters, the network outputs the displacement field that returns the myocardium to its unloaded state. On a held-out test set of 6,810 synthetic LV cases drawn from 20,700 finite-element simulations, the reported accuracy is a node-level DSC of 0.986 ± 0.023 and a Hausdorff distance of 0.083 ± 0.028 cm, with a mean node error of 0.028 ± 0.010 cm, at 0.023 seconds per case. This beats the inverse FE baseline on both accuracy and speed, and it also beats graph-convolution and point-cloud baselines. The authors attribute the data efficiency to a cycle-consistent bidirectional decoder: with about 200 supervised samples the model retains a DSC near 0.97, while removing the cycle term drops the same low-supervision setting to about 0.71. They present the method as the first learning-based approach to estimate unloaded cardiac geometry with explicit biophysical priors.","pith_inferences":["An implication the paper leaves implicit is that the network is effectively a fast emulator of its own FE data-generation pipeline, so clinical value hinges on how faithfully that pipeline represents real myocardial tissue rather than on the reported synthetic accuracy alone.","A testable extension is to train the same cycle-consistent architecture on end-systolic or full-cycle meshes with active contraction parameters, since the loading and unloading decoders already form a bidirectional deformation model.","A practical deployment check is to run the trained network on image-derived patient meshes with pathologies; the current experiments cover healthy anatomical variation only, so behavior on scarred, dilated, or hypertrophic ventricles remains open."],"forward_implications":["Inverse FE unloading can be replaced by one forward pass for meshes inside the trained distribution, cutting per-case cost from hours to roughly 0.02 seconds.","Cycle consistency enables a weak-supervision regime: about 200 labeled FE simulations are enough to keep node-level overlap near 0.97, reducing the cost of building training datasets.","Because the network is topology-agnostic and conditioned on pressure, stiffness, and fiber angles, it can process new meshes without template registration and interpolate to unseen physiological parameter values.","A fast, reliable zero-pressure reference state would make patient-specific stress and strain analysis and intervention planning practical in clinical time scales."],"supporting_citations":[{"why":"Supplies the cycle-consistency training principle that gives the network its weak self-supervision and data efficiency.","marker":"[40]"},{"why":"Defines the backward-displacement inverse FE solver used as the main accuracy and speed baseline.","marker":"[9]"},{"why":"Provides the population-derived statistical shape model used to generate the synthetic LV geometries for FE ground truth.","marker":"[23]"},{"why":"Supplies the exponential passive material law used in the FE pipeline that creates ground-truth unloaded and end-diastolic pairs.","marker":"[14]"},{"why":"The reduced-order principal-component inverse FE approach that HeartUnloadNet is compared against for deformation accuracy.","marker":"[36]"},{"why":"Supplies the graph-convolution baseline and the graph-based rationale for handling arbitrary mesh topology.","marker":"[17]"},{"why":"Meshing tool used to convert LV surface shapes into the tetrahedral volumetric meshes the network operates on.","marker":"[11]"}],"fun_headline_variants":["Unloaded heart shape from diastolic mesh in 0.02 s","Cycle-consistent graph net: zero-pressure LV in 20 ms","Deep learning beats inverse FE on unloaded heart geometry","Sub-mm unloaded LV geometry, 100k× faster inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ground-truth unloaded geometry is the one produced by the paper's inverse-FE unloading-reloading pipeline, which the paper itself notes is typically under-constrained and can have non-unique solution, so a network trained to reproduce that pipeline may not recover the true zero-pressure state if the constitutive model or inverse solution is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Unloaded heart shape from diastolic mesh in 0.02 s","Cycle-consistent graph net: zero-pressure LV in 20 ms","Deep learning beats inverse FE on unloaded heart geometry","Sub-mm unloaded LV geometry, 100k× faster inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000954,"raw_usage":{"total_tokens":4158,"prompt_tokens":1124,"completion_tokens":3034,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":2962}},"tokens_in":740,"tokens_out":3034,"duration_ms":24363,"temperature":1.0,"reasoning_tokens":2962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:32:29.190514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a heart or high-fidelity model for which the unloaded geometry is known independently, for example an excised ventricle imaged at zero cavity pressure and then loaded to a known pressure, and compare the network's predicted unloaded mesh with that measured state; if the node-level overlap or mean distance is substantially worse than the reported 0.986 and 0.028 cm, the claim that the network predicts true unloaded geometry fails.","supporting_citations":[{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversarial networks, in: Proceedings of the IEEE international conference on computer vision, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the cycle-consistency training principle that gives the network its weak self-supervision and data efficiency."},{"cited_title":"Efficientestimationofpersonalizedbiventricular mechanicalfunctionemployinggradient-basedoptimization.Internationaljournalfornumericalmethodsinbiomedicalengineering34,e2982","cited_arxiv_id":null,"evidence_quote":"Defines the backward-displacement inverse FE solver used as the main accuracy and speed baseline."},{"cited_title":"Left ventricular shape variation in asymptomatic populations: the multi-ethnic study of atherosclerosis","cited_arxiv_id":null,"evidence_quote":"Provides the population-derived statistical shape model used to generate the synthetic LV geometries for FE ground truth."},{"cited_title":"Passive material properties of intact ventricular myocardium determined from a cylindrical model","cited_arxiv_id":null,"evidence_quote":"Supplies the exponential passive material law used in the FE pipeline that creates ground-truth unloaded and end-diastolic pairs."},{"cited_title":"Efficientestimationofload-freeleftventriculargeometryandpassivemyocardialpropertiesusingprincipalcomponentanalysis","cited_arxiv_id":null,"evidence_quote":"The reduced-order principal-component inverse FE approach that HeartUnloadNet is compared against for deformation accuracy."}],"review_version":1}