{"id":"6f8f0183-7ac7-4152-be7c-c64b3c37c095","arxiv_id":"2508.20734","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CardioMorphNet estimates cardiac motion by learning to warp predicted heart-shape masks through a recurrent Bayesian network, reporting top results on UK Biobank cine MRI.","lead":"The paper introduces CardioMorphNet, a deep learning system that estimates heart motion from cardiac MRI by aligning predicted heart-shape maps instead of raw image intensities. It reports more accurate motion estimates and lower motion-field uncertainty than prior methods on UK Biobank scans.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Primary comparison is biased: CardioMorphNet's training loss explicitly optimizes the same automated-mask overlap used as the evaluation metric, while baselines are not given this supervision.","rationale":"I read the paper in good faith: the Bayesian recurrent formulation is clearly specified, the architecture is detailed, the code link is provided, and the ablation study shows each component contributes. The central claim, however, is that CardioMorphNet outperforms state-of-the-art methods in cardiac motion estimation. That claim is currently supported only by a comparison on mask-overlap metrics at ED/ES, where the proposed method's training loss (Eq. 7) optimizes exactly that overlap, while the baselines are not given the same mask supervision. This is a form of evaluation bias that can inflate results independent of actual motion quality. The reader's weakest_assumption focused on the accuracy of the automated masks from [54]; I agree that is an issue, but the more immediate methodological flaw is the training-evaluation alignment: even if the masks are perfect, the comparison is unfair because the proposed method is directly optimizing the evaluation metric. This distinction makes my agreement 'partial.' I do not reject the paper; the method is plausible and could be valuable. The verdict should remain conditional: the authors must provide a fair comparison (e.g., train baselines with the same mask loss) or validate on an independent ground truth (expert contours, phantom motion, or the M&M dataset claimed in the abstract). My concrete test targets exactly this and would settle whether the central claim survives.","tokens_in":18506,"tokens_out":7527,"duration_ms":70233,"concrete_test":"Retrain VoxelMorph (or SegMorph) with an additional supervised Dice/cross-entropy loss on the same ED/ES automated masks (analogous to Eq. 7), keeping its other losses intact, and re-evaluate on the same test split. If the DSC/HD95 gap in Table 2 shrinks to non-significance, the reported advantage is explained by training-objective alignment. As a secondary check, compare CardioMorphNet's warped masks against a small set of manually traced expert contours at ED/ES; if the advantage does not persist, the automated masks themselves are biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of superior cardiac motion estimation rests on Section 4.3 (Table 2), where DSC/JAC/HD95/MSD are computed between warped masks and 'ground-truth' masks at ED/ES. These masks are generated by the automated tool in [54] (Section 4.1). Critically, the training loss includes L_sup-shape (Eq. 7), a supervised cross-entropy between these same ground-truth masks and the warped masks at ED/ES. Thus CardioMorphNet is explicitly trained to maximize the evaluation metric, while all comparison methods (VoxelMorph, DragNet, SegMorph, etc.) are trained on intensity-based similarity or other objectives, not on mask overlap. This training-evaluation circularity means the reported improvements may reflect optimization alignment rather than genuine motion-estimation accuracy. The paper provides no independent evaluation: the abstract claims validation on M&M, but the experiments report only UK Biobank, and no manual expert contours or alternative motion benchmarks are used. This is the weakest load-bearing link: without a fair comparison or external ground truth, the headline superiority claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CardioMorphNet, a recurrent Bayesian deep network for 3D cardiac motion estimation from cine short-axis CMR sequences. The model combines a recurrent variational autoencoder, a segmentation network (SegNet), and a registration network (DeformNet). It estimates segmentation masks, DVFs, and latent variables jointly, with a loss derived from an ELBO plus a bending-energy regularizer. The key design feature is that registration is guided by warping segmentation maps rather than by an intensity-based image similarity loss. The method is evaluated on UK Biobank data, comparing mask overlap (DSC, JAC, HD95, MSD) with VoxelMorph, DragNet, SegMorph, JMS, SyN, and LCC-Demons, and also reports uncertainty maps and an ablation study.","tokens_in":18839,"tokens_out":3427,"duration_ms":38729,"significance":"If the claims are substantiated, the paper makes a useful contribution by demonstrating that shape-guided, recurrent Bayesian registration can produce accurate cardiac DVFs with explicit uncertainty estimates, without intensity-based similarity losses. The code is made available, and the method is compared against several strong baselines. The uncertainty maps are a valuable addition over deterministic registration methods. However, the evaluation currently has a serious circularity issue: the same automated masks are used both as training supervision at ED/ES and as the evaluation target, and the abstract claims M&M validation that does not appear in the experiments. These issues must be resolved before the central claim of superior performance can be accepted.","major_comments":[{"comment":"The evaluation is circular with respect to the supervision. Section 4.1 states that ED/ES time steps and ground-truth masks are obtained with the automated tool of [54]. These same masks are used in Eq. (7) as supervised training labels for the warped masks, and in Table 2 as the evaluation target for DSC/JAC/HD95/MSD. CardioMorphNet is therefore explicitly trained to maximize the evaluation metric, while the comparison methods (VoxelMorph, DragNet, SegMorph, etc.) are not trained with this mask-overlap supervision. The reported improvements may reflect optimization alignment rather than genuine motion-estimation accuracy. The paper provides no independent check, such as manual expert contours, an alternative motion benchmark, or M&M results. This is the load-bearing weakness of the paper; please add an independent evaluation or retrain baselines with the same mask supervision.","section":"§4.1, §4.3, Eq. (7), Table 2"},{"comment":"The claim that the loss is 'derived' from the Bayesian formulation is not supported by the derivation. The factorization in Eq. (3) is simplified to Eq. (4) by strong conditional-independence assumptions that are neither justified nor derived. Then the ELBO in Eq. (6) introduces hand-set weights λ1, λ2, λ3, and ρ, with values chosen empirically (Section 3.2). These ad hoc weights mean the objective is not a pure variational bound but a weighted sum of terms. This weakens the paper's central methodological claim. Please provide the full ELBO derivation, state the exact independence assumptions, and clarify how the λ weights are justified.","section":"§3.2, Eqs. (3)–(6)"},{"comment":"The semi-supervised shape loss is self-referential. Lsemi-shape minimizes the cross-entropy between Ψ, the SegNet posterior q(Mt|It), and θ, the warped previous mask p(Mt|Mt−1, Dt). At non-ED/ES frames, Ψ is the model's own output and is also being optimized. Therefore the DVF is trained against a moving target produced by the same network, which can trivially reduce this term by making SegNet copy its own predictions. This is not a fixed supervisory signal. Please clarify whether Ψ is detached or fixed, and discuss the implications for the registration objective.","section":"§3.3, Eq. (8)"},{"comment":"The abstract states that the method is 'Validated on the UK Biobank and M&M datasets', but the experiments in Section 4 only report UK Biobank. There is no M&M dataset description, preprocessing, results table, or analysis anywhere in the paper. Either add the M&M experiments or remove the claim from the abstract. This is a factual inconsistency in the headline claim.","section":"Abstract, §4"}],"minor_comments":[{"comment":"The text states that a paired t-test was employed, but no p-values or confidence intervals are reported in Table 2 or the surrounding text. Please report statistical significance for the main comparisons.","section":"§4.3"},{"comment":"Since the automated tool of [54] is used for ground-truth masks, it would be helpful to report the tool's segmentation accuracy (e.g., Dice vs. manual contours) on the UK Biobank subset used here.","section":"§4.1"},{"comment":"There are typos: 'DrageNet' should be 'DragNet', and 'SegMorh' should be 'SegMorph'.","section":"§5"},{"comment":"In Eq. (9), the notation Tr(CDj) is unclear because CD is a voxel-wise 3×3 covariance, not a matrix. Please define the trace operation on this tensor.","section":"Eq. (9)"},{"comment":"The column header 'NoDeformNet' is confusing; it appears to mean that DeformNet is always enabled. Consider renaming to 'DeformNet' with ✓/✗ like the other columns.","section":"Table 3"},{"comment":"The uncertainty maps in Figure 6 use a color scale but no numerical colorbar; adding one would make the visualization more interpretable.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible extension of the authors' prior DragNet/SegMorph line of work, and the code availability is a plus. The main concern is the evaluation circularity and the missing M&M validation claimed in the abstract. These are fixable with additional experiments, but without them the headline superiority claim is not established. I would support acceptance after a substantive revision that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CardioMorphNet is a serious attempt at cardiac motion estimation, but its central superiority claim rests on an unfair comparison. The authors train their model to maximize overlap with a set of automated masks at ED/ES and then evaluate against those same masks, while the baselines get no such supervision. That alone should make you cautious about the numbers.\n\nWhat's actually new: the framework replaces intensity-based registration similarity with recursive registration of segmentation maps, derived from a Bayesian ELBO, and it produces voxel-level DVF uncertainty from an explicit covariance. The recurrent VAE + ConvLSTM structure is a sensible extension of their own SegMorph/DragNet lines. The ablation shows both SegNet and RVAE contribute, and the code is linked. Credit where due: the method is well-specified and the idea of grounding registration in anatomy rather than raw intensities is worth pursuing.\n\nSoft spots, in order of severity. First, the evaluation circularity. The ED/ES 'ground truth' masks come from the automated tool in [54], and the training loss includes L_sup-shape, which is exactly the cross-entropy between those masks and the warped masks at those same frames. The metric reported in Table 2 is overlap between warped masks and those same masks. So the model is directly optimizing the evaluation target. Baselines (VoxelMorph, DragNet, SegMorph, etc.) are trained with intensity-based or other objectives, not mask overlap. The reported margin may reflect optimization alignment rather than better motion estimation. Also, the abstract claims validation on M&M and mentions clinical indices, but the body reports only UK Biobank and has no clinical indices section. Either the abstract is ahead of the paper or the experiments are missing; either way it needs fixing.\n\nSecond, the Bayesian derivation is more asserted than shown. Equation (3) to (4) relies on a factorization with assumptions that are stated but not justified, and the lambdas in Eq. (6) are 'chosen empirically' — fine, but it undercuts the rigor the prose claims.\n\nThird, there is no independent check on the automated masks. A small expert-annotation subset or a secondary benchmark would go a long way.\n\nWho's it for: someone working on cardiac motion or joint segmentation-registration. The method itself is plausible and the ablations are informative. But don't take the superiority claim at face value. If you cite it, cite it as a method, not as proof that shape-guided registration beats the field.\n\nRecommendation: this deserves peer review, but the reviewers should ask for a fair comparison — e.g., give the baselines the same mask supervision or evaluate on an independent ground truth — and for the M&M claim to be backed or dropped.","headline":"Plausible shape-guided registration method, but the headline comparison is biased: the model trains on the same automated masks it is then scored against.","tokens_in":19282,"tokens_out":4641,"would_cite":false,"duration_ms":41770,"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":"CardioMorphNet claims cardiac motion is learned more accurately by recursively registering segmentation maps than by matching image intensities, and reports superior shape-registration metrics and lower DVF uncertainty on a large population","keywords":["Bayesian Modelling","Cardiac Motion Estimation","Deformable Image Registration","Cine CMR","Recurrent Variational Autoencoder","Shape-Guided Registration","Uncertainty Quantification","Population Cardiac MRI"],"falsifier":"On a held-out set with expert manual contours (or a digital phantom with known ground-truth deformation), compare CardioMorphNet's warped masks and DVFs against independent ground truth rather than the automated tool's masks. If its advantage over intensity-based baselines disappears, the claim that shape-guided mask registration yields superior cardiac motion estimation is refuted.","tokens_in":18448,"feed_emoji":"🫀","tokens_out":9608,"duration_ms":102129,"temperature":0.7,"pith_summary":"CardioMorphNet aims to show that cardiac motion can be estimated more accurately by registering anatomical segmentation maps than by matching image intensities. The authors build a recurrent Bayesian network in which a segmentation network proposes the current frame's heart shape and a deformable registration network proposes a displacement field from two consecutive SAX volumes and a recurrent latent variable; the prior warps the previous shape with that field. The variational loss derived from the model's ELBO contains supervised cross-entropy at the two annotated frames (end-diastole and end-systole) and semi-supervised cross-entropy between the proposed and warped shapes at every other frame, with no intensity-based registration term. On a large population cardiac-MRI cohort, the paper reports that this approach yields higher Dice and Jaccard scores, lower surface distances, and lower DVF uncertainty inside the heart than several intensity-based and joint segmentation-registration baselines. If correct, the work shows that anatomy-level supervision can replace pixel-level appearance matching for motion tracking and can provide per-voxel uncertainty from an explicit posterior covariance.","feed_headline":"Shape-guided network tops pixel-based rivals on cardiac motion","feed_subtitle":"It aligns heart segmentations over the cycle, giving sharper deformation fields and lower uncertainty in cardiac regions.","key_machinery":"The load-bearing machinery is a recurrent variational autoencoder with two posterior networks and one warping prior. SegNet computes q(Mt|It), the per-voxel distribution over cardiac labels for the current frame; DeformNet computes q(Dt|It, It−1, Zt), a Gaussian displacement field with mean and covariance sigma-squared-I plus V-transpose-V; and the prior p(Mt|Mt−1, Dt) is implemented by warping the previous segmentation map with the current DVF. A ConvLSTM hidden state carries spatio-temporal context into the latent variable Zt. The ELBO loss combines a supervised and semi-supervised shape cross-entropy term that replaces intensity similarity, KL divergences for the displacement field and th","core_discovery":"The central claim, stated on the paper's own terms, is that a recurrent variational Bayesian framework can learn cardiac motion by recursively registering segmentation maps, and doing so is enough to beat methods that optimize intensity-based image similarity. In the proposed generative model, the joint probability of the sequence factorizes so that the current mask depends on the previous mask and the current displacement field; the posterior estimates masks, displacement fields, and latent variables from the current and previous SAX volumes and a recurrent hidden state. Minimizing the resulting ELBO loss aligns the warped previous mask with the current segmentation at every time step, supe","pith_inferences":["The evaluation uses the same automated contour tool for training supervision and for scoring; a reader should treat the reported accuracy as agreement with that tool's segmentations. An expert-annotation or synthetic-deformation benchmark would test whether the shape-guided loss truly captures cardiac motion rather than fitting the tool's contours.","The abstract asserts more accurate clinical-index estimation, but the body reports only mask-overlap, uncertainty, and diffeomorphism evaluations; that assertion lacks a demonstrated result in the presented text.","The recursive mask-registration loss is not cardiac-specific in principle; the same two-frame-supervision scheme could transfer to other organs or imaging modalities where reliable contours exist at two timepoints and intensity matching is unreliable.","Because the model uses only short-axis volumes and discrete masks, through-plane motion and rotation that do not change mask overlap may be under-constrained; combining with long-axis views or mesh/tagging evidence would be a natural stress test."],"forward_implications":["DVFs are trained by mask overlap across the whole cardiac cycle, so only the ED and ES frames need manual or automated labels; intermediate frames are held together by recursive shape consistency.","The framework produces per-voxel DVF uncertainty directly from the learned covariance, without dropout or test-time augmentation.","On the reported cohort, shape-guided registration gives higher DSC/JAC and lower HD95/MSD than intensity-based and joint registration baselines in the LV, LV myocardium, and RV.","The DVF uncertainty in the cardiac region is lower than in DragNet and SegMorph, which the paper interprets as higher confidence in the estimated heart motion."],"supporting_citations":[{"why":"Supplies the automated ED/ES ground-truth masks used both to supervise training and to score registration accuracy.","marker":"[54]"},{"why":"Provides the variational lower-bound (ELBO) formulation from which the training loss is derived.","marker":"[52]"},{"why":"Defines the standard deep-learning registration approach based on intensity similarity that the paper argues against and compares itself to.","marker":"[39]"},{"why":"The closest joint segmentation-registration baseline, whose intensity-based loss the paper seeks to outperform with shape-guided loss.","marker":"[51]"},{"why":"Recurrent variational registration baseline used for DVF uncertainty comparison.","marker":"[49]"},{"why":"Earlier joint motion-segmentation baseline that motivates the shape-guided design.","marker":"[50]"},{"why":"Probabilistic diffeomorphic registration model whose latent-space idea is extended with recurrence and mask supervision.","marker":"[48]"},{"why":"Conventional diffeomorphic registration baseline used to demonstrate improvement over non-learning methods.","marker":"[41]"}],"fun_headline_variants":["Shape-guided Bayesian net trumps pixel-based cardiac motion","Segment alignment beats intensity matching in heart motion AI","Bayesian model lowers uncertainty in cardiac motion tracking","No intensity losses: AI maps heart motion via shape recursion"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The automated tool that supplies the end-diastolic and end-systolic masks is accurate enough to serve as both training labels and the evaluation standard; if those masks are biased, the reported scores mainly measure agreement with the tool, not true cardiac motion.","fun_headline_variants_meta":{"raw":{"variants":["Shape-guided Bayesian net trumps pixel-based cardiac motion","Segment alignment beats intensity matching in heart motion AI","Bayesian model lowers uncertainty in cardiac motion tracking","No intensity losses: AI maps heart motion via shape recursion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3344,"prompt_tokens":769,"completion_tokens":2575,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2513}},"tokens_in":513,"tokens_out":2575,"duration_ms":22800,"temperature":1.0,"reasoning_tokens":2513,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:51:43.419271+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set with expert manual contours (or a digital phantom with known ground-truth deformation), compare CardioMorphNet's warped masks and DVFs against independent ground truth rather than the automated tool's masks. If its advantage over intensity-based baselines disappears, the claim that shape-guided mask registration yields superior cardiac motion estimation is refuted.","supporting_citations":[{"cited_title":"Automated car- diovascular magnetic resonance image analysis with fully convolutional networks","cited_arxiv_id":null,"evidence_quote":"Supplies the automated ED/ES ground-truth masks used both to supervise training and to score registration accuracy."},{"cited_title":"Voxel- morph: a learning framework for deformable medical image registration","cited_arxiv_id":null,"evidence_quote":"Defines the standard deep-learning registration approach based on intensity similarity that the paper argues against and compares itself to."},{"cited_title":"Segmorph: Concurrent motion estimation and segmentation for cardiac mri sequences","cited_arxiv_id":null,"evidence_quote":"The closest joint segmentation-registration baseline, whose intensity-based loss the paper seeks to outperform with shape-guided loss."},{"cited_title":"Dragnet: Learning-based de- formable registration for realistic cardiac mr sequence generation from a single frame","cited_arxiv_id":null,"evidence_quote":"Recurrent variational registration baseline used for DVF uncertainty comparison."},{"cited_title":"Joint learning of motion estimation and segmentation for cardiac mr image sequences","cited_arxiv_id":null,"evidence_quote":"Earlier joint motion-segmentation baseline that motivates the shape-guided design."},{"cited_title":"Learn- ing a probabilistic model for diffeomorphic registration","cited_arxiv_id":null,"evidence_quote":"Probabilistic diffeomorphic registration model whose latent-space idea is extended with recurrence and mask supervision."},{"cited_title":"A reproducible evaluation of ants similarity metric performance in brain image registra- tion","cited_arxiv_id":null,"evidence_quote":"Conventional diffeomorphic registration baseline used to demonstrate improvement over non-learning methods."}],"review_version":1}