{"id":"d12ba241-119b-4da3-b0df-28578e825d38","arxiv_id":"2411.15233","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A volumetric neural deformable model with point transformers recovers dense 3D heart wall motion from 2D apparent motion cues, beating iterative baselines on synthetic data.","lead":"This paper trains a neural network to turn sparse 2D motion cues from tagged MRI planes into dense 3D heart wall motion, using synthetic data it generates itself. The idea could speed up cardiac motion analysis, but the results are not yet tested on real MRI images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation is confined to the authors' own υNDM-based simulation, whose hand-set temporal scalars (Supp. Eqs. 13-15) match the network's output parameterization, so the reported MAE may reflect self-consistency with the simulator rather than transfer to real tagged MRI.","rationale":"The reader's CONDITIONAL verdict is appropriate, and the identified weakest assumption is the same one that is most load-bearing. The paper is internally consistent: the hybrid point transformer is well-motivated, the ablations support the design choices (Tables 2 and 3), and the comparison with MDM and NDM is reasonable given the lack of learning-based baselines. However, the evaluation is effectively a self-consistency test within the authors' simulation framework. Ground-truth motion is generated by Eq. 6 with hand-set per-coordinate scalars, and the network's output representation (global a1,a2,a3,τ plus local diffeomorphic flow) can express that exact interpolation. SPAMM datapoints are computed by mesh-plane clipping with exact correspondences, so the inputs are idealized and the reported accuracy is in-distribution. The supplementary material itself acknowledges the potential deviation of the synthetic data from real distributions. Therefore the central claim about recovering motion from multi-planar tagged MRI is conditional on simulation realism; a transfer test on an independently generated temporal profile would isolate whether the model learned the specific hand-set scalars, and a real-data study, even without ground truth, using physiological plausibility and comparison to existing methods would address the remaining gap. No internal inconsistency or architecture flaw was identified; the concern is external validity. Thus the reader's conditional acceptance remains the correct verdict.","tokens_in":16944,"tokens_out":5801,"duration_ms":57596,"concrete_test":"Generate a second test set with the same geometry and SPAMM extraction pipeline but with temporal scalars s(t_q) replaced by an independent physiological profile (e.g., derived from a published LV volume curve or a biomechanical model), then evaluate the already-trained network without any fine-tuning. If MAE rises from 0.954 mm to roughly 1.5-2 mm or higher, the model has exploited the specific hand-set scalars in Eqs. 13-15 rather than learning general 3D motion recovery. If MAE stays near 0.954 mm, the temporal interpolation is not the limiting factor, and the remaining concern shifts to the idealization of SPAMM cues, which could then be tested by adding realistic noise, outliers, and missing correspondences to the input S(t_q), S(t_q+1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest assumption is not merely the absence of real data; it is that the synthetic ground-truth motion is generated by the same parametric model family the network is designed to invert. Step 3 (Eq. 6) linearly interpolates ED/ES υNDMs using hand-set per-coordinate temporal scalars s_x, s_y, s_z (Supp. Eqs. 13-15). The network's output q_g=(a1,a2,a3,τ) plus local diffeomorphic flow q_d can represent exactly this kind of interpolation, and the training loss is measured against those interpolated M(t_q). Moreover, the apparent motion cues S(t_q) and S(t_q+1) are exact mesh-plane intersections with known one-to-one correspondences, free of tracking error, tag fading, noise, and through-plane artifacts. The test set is drawn from the same 500-subject simulation pipeline, so the reported MAE of 0.954 mm is an in-distribution fitting result. Supplementary Section 8 explicitly concedes that missing intermediate cardiac phases in the CMR dataset can make synthetic data deviate from real data distributions, and the only QC is radiologist review, not quantitative validation. Consequently, the central claim in the title and abstract that the method recovers 3D motion from multi-planar tagged MRI is not yet supported. The paper demonstrates a simulation-to-simulation proof of concept; external validity on actual tagged MRI remains unestablished.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learning-based framework, volumetric neural deformable models (υNDMs), to recover dense 3D heart wall motion from sparse 2D apparent motion cues in multi-planar tagged MRI. The heart wall is represented as a volumetric deformable model with global deformation parameters (ellipsoid aspect ratios and twisting angle) and a local diffeomorphic point flow. A hybrid point transformer with cross-attention between material points and SPAMM datapoints, followed by self-attention encoder–decoder layers, is trained to map apparent motion cues to true 3D material point trajectories. Since no real tagged-MRI ground truth is available, the authors construct a synthetic dataset by fitting two-layer υNDMs to ED and ES segmentations from a public CMR dataset, interpolating intermediate frames with hand-set temporal scalars, and computing SPAMM datapoints by mesh-plane clipping. On a synthetic test set of 200 subjects the method reports MAE 0.954 mm (SI 0.096), outperforming a reimplemented meshless deformable model (MDM) and NDM. Ablations examine the contributions of global/local deformations, k-NN number, training stages, and separated vs. mixed cue fusion.","tokens_in":17262,"tokens_out":4735,"duration_ms":46028,"significance":"If the synthetic results transfer to real tagged MRI, the approach would be a valuable contribution: it is fast (0.147 s per time point), provides dense correspondence-preserving motion fields, and introduces a new architecture for fusing multi-planar apparent motion cues. The simulation framework itself could be useful for benchmarking if its realism is established. However, the current evidence is only a simulation-to-simulation proof of concept; the central claim that the method recovers 3D motion from multi-planar tagged MRI is not yet supported by quantitative results on real data. The paper is honest about the synthetic nature of its evaluation, but the strength of the claim in the title and abstract exceeds the evidence.","major_comments":[{"comment":"The evaluation is confined to a synthetic dataset generated by the authors' own simulation, and the ground-truth motion lies within the same parametric family the network is designed to invert. In Step 3, intermediate frames are produced by linearly interpolating ED and ES υNDMs with hand-set temporal scalars s_x(t_q), s_y(t_q), s_z(t_q) (Supplementary Eqs. 13–15). The network outputs global parameters (a1, a2, a3, τ) plus a local diffeomorphic flow, which can represent exactly this type of interpolation. The test set is drawn from the same 500-subject simulation pipeline used for training. Consequently, the reported MAE of 0.954 mm (Table 1) is an in-distribution fitting result and does not establish that the network recovers true physiological heart wall motion. Supplementary Section 8 explicitly concedes that missing intermediate cardiac phases in the CMR dataset can make the synthetic data deviate from the real distribution. The paper provides only radiological QC of the simulated shapes, no quantitative comparison to any real tagged-MRI motion measurements. The title and abstract claim recovery from multi-planar tagged MRI; that claim is not supported by the current experiments.","section":"§3.2 Step 3, Eq. (6) and Supplementary §7, §8; Table 1"},{"comment":"The apparent motion cues fed to the network are exact SPAMM datapoints computed by mesh-plane clipping, with one-to-one correspondences and no noise, tag fading, tracking errors, or through-plane artifacts. In real tagged MRI, SPAMM datapoint extraction is noisy and correspondences between time frames are not perfect; in practice these signals are obtained by tracking algorithms such as those cited in the introduction (e.g., DeepTag). The paper does not evaluate robustness to these realistic degradations. Since the stated task is to recover motion from multi-planar tagged MRI, the evaluation should include either real tagged-MRI data or a realistic noise/correspondence-error model applied to the synthetic inputs. Without such evidence, the method's performance on clean, exactly-corresponding synthetic datapoints does not predict its clinical utility.","section":"§3.2 Step 4 and §3.3"},{"comment":"The comparison to MDM may be confounded by the fact that MDM is a reimplementation whose hyperparameters (e.g., phyxel kernel radius) are acknowledged to be tuned on the training dataset, with no sensitivity analysis or cross-validation reported. The paper states that tuning on training data cannot guarantee optimal test performance, but it does not quantify how sensitive the MDM results are to its parameters. This is a secondary issue, but it weakens the claim that the proposed method outperforms conventional deformable-model approaches; the heading 'state-of-the-art' for the reimplemented MDM is not substantiated.","section":"§4.3, Table 1"}],"minor_comments":[{"comment":"The MAE metric excludes the starting time point, which is reasonable, but the caption of Table 1 says 'averaged over a whole cardiac cycle' without mentioning this exclusion; please clarify the caption or the metric definition.","section":"§4.2, Eq. (12)"},{"comment":"The color scale in Fig. 8 for error maps is not defined in the caption (the same color bar appears to cover 0–5 mm); a clear color bar and a statement of the displayed cardiac phases would improve readability.","section":"§4.5.1, Fig. 8"},{"comment":"The notation 'we take M as e in Eq. (2)(3)' is confusing; the distinction between the undeformed material points M and the ellipsoid primitive e should be spelled out when defining the global deformation step M'(t_q) = M(t_q) ∘ q_g.","section":"§3.3, Eq. (8)"},{"comment":"The claim that this is the first learning-based approach for 3D true heart wall motion recovery from 2D apparent motion cues is a novelty assertion that is difficult to verify and not essential; consider softening it to 'a learning-based approach' and citing related learned motion-recovery works where applicable.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":"The core methodological idea is interesting and the synthetic framework is clearly described, but the evidence is currently insufficient for the claims made. I strongly encourage the authors to either add a real tagged-MRI evaluation (even on a small cohort, with manual or tracked landmarks as reference) and an analysis of robustness to correspondence noise, or to explicitly reposition the paper as a simulation-based proof of concept with a correspondingly scaled-back title and abstract. The absence of such a change would not justify acceptance, as the central claim of recovering motion from multi-planar tagged MRI is currently unsubstantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about this paper is that it proposes the first learning-based method for recovering 3D heart wall motion from sparse 2D apparent cues, and it works on the authors' synthetic benchmark. But every result is on data generated by their own simulator, and that simulator uses the same deformable model family the network is designed to invert. So the reported MAE of about 0.95 mm is a self-consistency result, not a demonstration on real tagged MRI.\n\nWhat's genuinely new: the volumetric extension of neural deformable models, with twisting parameters, and a hybrid point transformer that uses cross-attention to fuse apparent cues from SAX and LAX planes and self-attention to refine them into dense true motion. The simulation framework itself is also a contribution, though its realism rests on one radiologist's qualitative review. The paper does solid engineering: clean ablations, comparison against reimplemented MDM and NDM, and a two-stage training scheme that handles accumulation error. On the synthetic test set, the method beats both baselines by a clear margin and is much faster than MDM.\n\nWhere it's soft: the external validity is unestablished. The ground-truth motion is created by linearly interpolating ED/ES models with hand-set per-coordinate scalars (Supp. Eqs. 13-15), and the apparent cues are exact mesh-plane intersections with known correspondences. No tag fading, no tracking noise, no through-plane ambiguity, no respiratory motion. The authors acknowledge in the supplementary that missing intermediate cardiac phases in the CMR dataset can make synthetic data deviate from real distributions, but they only offer the radiologist's opinion as QC. So the central claim in the title and abstract — recovery from multi-planar tagged MRI — is a statement of intent, not of demonstrated performance.\n\nThat said, this is not a fatal methodological flaw. For a first learning-based method in this niche, synthetic validation with known ground truth is a reasonable starting point, and the paper is transparent about it in the abstract. The weak spot is the size of the gap between what they claim and what they measure. If a referee pushes back, the fix is either real tagged MRI validation or a substantially more realistic simulation (noise, tag fading, partial correspondences). I'd probably require that for acceptance in a strong venue, but I wouldn't desk-reject it.\n\nBottom line: worth a serious referee. The method is novel, the engineering is careful, and the problem matters. The evaluation is the one thing that keeps it from being a slam dunk.\n\nRecommendation: send to peer review, but be explicit with the authors that validation on real tagged MRI — or at least a realistic synthetic challenge — is a condition for acceptance.","headline":"A genuinely new learning-based method for 3D heart wall motion recovery, but the evaluation is entirely synthetic, so the title's promise about real tagged MRI remains unproven.","tokens_in":17799,"tokens_out":3201,"would_cite":true,"duration_ms":28691,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A volumetric neural deformable model with a hybrid point transformer maps sparse 2D apparent-motion cues from multi-planar tagged MRI into dense 3D true heart-wall motion, reporting a mean error of 0.954 mm on a synthetic cardiac test set.","keywords":["tagged MRI","cardiac motion recovery","volumetric neural deformable model","apparent motion fusion","point attention","left ventricle wall motion","synthetic cardiac motion simulation"],"falsifier":"Take a real multi-planar tagged MRI sequence of a heart or a mechanical phantom with known 3D motion, run the network trained on the synthetic dataset, and measure the distance between predicted and known material-point trajectories; if the error is much larger than the 0.954 mm reported here, or the deformed mesh shows frequent self-intersections, the synthetic distribution does not represent real apparent motion.","tokens_in":1980,"feed_emoji":"🫀","tokens_out":2194,"duration_ms":94871,"temperature":0.7,"pith_summary":"When a heart wall is imaged with tagged MRI, the visible tag intersections are not the same material points over time: points move out of the imaging plane, so the 2D apparent motion seen in the image is an incomplete and biased sample of the true 3D motion. This paper tries to show that this gap can be bridged by learning, not iterating: a volumetric neural deformable model (υNDM) represents the wall as a layered mesh with global deformation parameters and a diffeomorphic local flow, and a hybrid point transformer learns to map sparse apparent-motion pairs from short-axis and long-axis planes into dense true 3D motion for every material point. To train and test this mapping, the authors build a simulation that derives heart-wall geometry from a large cardiac MRI dataset and synthesizes a full cardiac cycle, including twisting, with known ground truth. On that synthetic test set they report a mean absolute error of 0.954 mm, below the 1.180 mm and 1.532 mm of the two comparison methods, with a self-intersection ratio of 0.096. The practical payoff, if true, is that dense regional 3D wall motion could be recovered from standard 2D multi-planar tagged MRI in a single forward pass, without slow iterative optimization.","feed_headline":"Model turns 2D MRI tag motion into dense 3D heart-wall motion","feed_subtitle":"It fuses short- and long-axis tag cues into dense wall trajectories, averaging under 1 mm error on synthetic tests","key_machinery":"The central object is the volumetric neural deformable model (υNDM), an extension of a two-surface neural deformable model to a full wall with multiple layers, combining global deformation parameter functions $q_g = (a_1, a_2, a_3, \\tau)^\\top$ with a diffeomorphic local point flow $q_d$. The carrying mechanism is the hybrid point transformer: a point cross-attention layer treats each material point as a query and nearby SPAMM datapoints as keys and values, so sparse apparent-motion cues from short-axis and long-axis planes are upsampled and fused into dense true-motion hints; stacked point self-attention encoder-decoder layers then refine the hints into a latent motion code, from which MLPs predict the global parameters and a conditional neural ODE predicts the local diffeomorphic flow. Marginal-space learning gradually increases deformation freedom during training, and a two-stage scheme with a longer sequential rollout reduces accumulated drift during inference. The full mapping is written as $\\Psi(M(t_q), A(t_q, t_{q+1})) \\mapsto (\\hat{M}(t_{q+1}), q^\\upsilon_N(t_q, t_{q+1}))$.","core_discovery":"The paper's central claim is that dense 3D true heart-wall motion can be recovered from sparse multi-planar 2D apparent motion cues by learning a volumetric neural deformable model. The model represents the left-ventricular wall as a layered ellipsoidal mesh whose geometry is governed by global deformation parameter functions (aspect ratios and twisting) and whose local deformation is a diffeomorphic point flow integrated by a neural ODE. A hybrid point transformer maps the observed SPAMM datapoint pairs $\\{S(t_q), S(t_{q+1})\\}$ into per-material-point motion: cross-attention with material points as queries and SPAMM datapoints as keys and values upsamples and fuses the sparse cues from SAX and LAX views into dense true-motion hints, and hierarchical self-attention refines these hints into a latent motion code that the network decodes into global and local deformation parameters. The authors report that this learned mapping recovers 3D material-point trajectories with a mean absolute error of 0.954 mm over a cardiac cycle, compared with 1.180 mm for the iterative meshless deformable model and 1.532 mm for the geometry-only neural deformable model, while keeping the self-intersection face ratio at 0.096 on the synthetic test set.","pith_inferences":["Editorial inference: the cross-attention fusion design is not specific to cardiac tagging; the same query-key-value pattern could be applied to any 2D-to-3D motion task where two orthogonal image-plane families provide complementary displacement components, such as tissue phase mapping or DENSE imaging, provided synthetic training data with ground truth can be generated.","Editorial inference: the paper's simulation uses hand-set temporal scalars and linear interpolation between two cardiac phases; a natural stress test would be to generate training data with a learned cardiac motion model or with noise consistent with real SPAMM tag fading, then check whether the reported accuracy survives.","Editorial inference: because SPAMM datapoints do not coincide with material points and the method fuses all coordinates from both plane families, the paper implicitly claims that learned attention can replace the hand-coded separation of in-plane and through-plane cues used by classical deformable models; a direct test would be to compare on real data where through-plane motion is large."],"forward_implications":["A fixed set of material points is deformed sequentially through a cardiac cycle, so material correspondence across all 20 phases is established automatically without a separate tracking or registration step.","The learned fusion of SAX and LAX apparent-motion cues with mixed coordinates outperforms the classical practice of feeding only in-plane components from SAX and through-plane components from LAX, as shown in ablation A4.","Global parameters capture twisting and regional shape change while the diffeomorphic local field keeps the deformed mesh nearly fold-free, with a self-intersection ratio of 0.096 on the synthetic test set.","Two-stage training with a rollout length of five reduces accumulation drift, making sequential inference over a full cardiac cycle stable.","The simulation framework generates full-cardiac-cycle 3D wall motion with known ground truth, which the authors use both to train the network and to quantify recovery accuracy, and the same tool could be reused to evaluate future methods."],"supporting_citations":[{"why":"Defines the deformable-model/SPAMM framework for recovering 3D LV wall motion from apparent cues, including the twisting values and correspondence handling that this method learns.","marker":"[39]"},{"why":"Introduces the neural deformable model (NDM) with diffeomorphic point flow local deformation, which υNDM extends from two surfaces to a full wall and also serves as a baseline.","marker":"[55]"},{"why":"Supplies the point attention mechanism used for the hybrid cross-attention and self-attention transformer.","marker":"[61]"},{"why":"Provides the meshless deformable model (MDM) baseline and the global/local deformation parameter split that the motion recovery network predicts.","marker":"[51]"},{"why":"Large 3D cardiac cine MRI dataset whose ED/ES segmentations are the geometric source for the synthetic heart-wall motion simulation.","marker":"[5]"},{"why":"Neural ODE solver used to integrate the diffeomorphic point flow for the local deformation field.","marker":"[9]"},{"why":"Introduces the ellipsoidal deformable model with parameter functions and twisting that υNDM adopts for global deformation.","marker":"[40]"},{"why":"Mesh-plane clipping algorithm used to compute SPAMM datapoints from the volumetric mesh in the simulation.","marker":"[16]"}],"fun_headline_variants":["AI recovers 3D heart motion from 2D MRI tags","Volumetric neural model maps 2D tags to 3D heart wall motion","Deep learning turns sparse MRI tags into dense heart trajectories","Neural deformable model reconstructs 3D heart motion from 2D cues"],"cache_read_input_tokens":19840,"weakest_assumption_plain":"The whole evaluation rests on the synthetic data being realistic: ground-truth motion is produced by linearly interpolating between end-diastolic and end-systolic wall geometries with manually chosen time-varying scalars, and the apparent-motion cues are computed exactly by clipping the mesh with imaging planes, with only a radiologist's visual review rather than quantitative comparison to real tagged MRI.","fun_headline_variants_meta":{"raw":{"variants":["AI recovers 3D heart motion from 2D MRI tags","Volumetric neural model maps 2D tags to 3D heart wall motion","Deep learning turns sparse MRI tags into dense heart trajectories","Neural deformable model reconstructs 3D heart motion from 2D cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":3082,"prompt_tokens":1090,"completion_tokens":1992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":1911}},"tokens_in":706,"tokens_out":1992,"duration_ms":11955,"temperature":1.0,"reasoning_tokens":1911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:09:16.462617+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real multi-planar tagged MRI sequence of a heart or a mechanical phantom with known 3D motion, run the network trained on the synthetic dataset, and measure the distance between predicted and known material-point trajectories; if the error is much larger than the 0.954 mm reported here, or the deformed mesh shows frequent self-intersections, the synthetic distribution does not represent real apparent motion.","supporting_citations":[{"cited_title":"Analysis of left ventricular wall motion based on volumetric deformable models and mri-spamm","cited_arxiv_id":null,"evidence_quote":"Defines the deformable-model/SPAMM framework for recovering 3D LV wall motion from apparent cues, including the twisting values and correspondence handling that this method learns."},{"cited_title":"Neural deformable models for 3d bi- ventricular heart shape reconstruction and modeling from 2d sparse cardiac magnetic resonance imaging","cited_arxiv_id":null,"evidence_quote":"Introduces the neural deformable model (NDM) with diffeomorphic point flow local deformation, which υNDM extends from two surfaces to a full wall and also serves as a baseline."},{"cited_title":"Point transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the point attention mechanism used for the hybrid cross-attention and self-attention transformer."},{"cited_title":"Meshless deformable models for 3d cardiac motion and strain analysis from tagged mri","cited_arxiv_id":null,"evidence_quote":"Provides the meshless deformable model (MDM) baseline and the global/local deformation parameter split that the motion recovery network predicts."},{"cited_title":"A bi-ventricular cardiac atlas built from 1000+ high resolution mr images of healthy subjects and an anal- ysis of shape and motion","cited_arxiv_id":null,"evidence_quote":"Large 3D cardiac cine MRI dataset whose ED/ES segmentations are the geometric source for the synthetic heart-wall motion simulation."},{"cited_title":"Deformable models with parameter functions for car- diac motion analysis from tagged mri data","cited_arxiv_id":null,"evidence_quote":"Introduces the ellipsoidal deformable model with parameter functions and twisting that υNDM adopts for global deformation."},{"cited_title":"Computer Graphics: Principles and Prac- tice","cited_arxiv_id":null,"evidence_quote":"Mesh-plane clipping algorithm used to compute SPAMM datapoints from the volumetric mesh in the simulation."}],"review_version":1}