{"id":"74290d85-ffcd-4b77-8152-2ee4ef09d6da","arxiv_id":"2507.02289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CineMyoPS segments myocardial scar and edema directly from contrast-free cine cardiac MRI by jointly learning motion and anatomy features, reaching test Dice scores of 0.53 for scar and 0.57 for edema.","lead":"A deep learning model called CineMyoPS tries to find heart muscle damage, specifically scar and edema, using only regular cine MRI scans that need no contrast dye. It combines motion tracking, heart anatomy segmentation, and time-series information, and is tested on 145 patients from three centers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fused LGE/T2 gold-standard labels registered to the cine ED frame via MvMM are never validated for alignment; if the registration is inaccurate, all reported Dice/HD numbers and baseline comparisons lose their meaning.","rationale":"The central claim has two parts: joint scar/edema segmentation from cine-only input, and 'promising performance' exceeding four baselines. Both rest on the validity of the labels used for training and evaluation. The label-generation pipeline is the least externally verifiable step: no registration error is quantified, and the paper's own ambiguity analysis (inter-observer Dice 0.69 for scar and 0.73 for edema) only addresses rater disagreement, not spatial alignment between modalities. I considered the Table III feature-label inconsistency, where the MyoPSIΦ and MyoPSΦL rows appear to have swapped checkmarks; it is a real reproducibility defect and could alter the feature-ablation conclusion, but it does not threaten the existence of a cine-only joint segmentation model or the main test-set comparison. I also considered the modest statistical significance of some baseline improvements, for example CineMyoPS versus 2D+1D Unet on scar Dice is not marked significant in Table V; this is a claim-strength issue rather than a correctness risk. The paper deserves credit for within-dataset comparison under a common training protocol, for reporting inter-observer variability, and for honestly discussing apical-slice failures. The conditional verdict is appropriate: if the registration concern is resolved by reporting alignment metrics, the central claim is credible; if not, the quantitative claims are not independently interpretable. Since the reader already flagged this exact assumption, the verdict should remain unchanged.","tokens_in":19322,"tokens_out":8153,"duration_ms":98136,"concrete_test":"On the Center-R test set, compute the Dice overlap between the myocardium masks propagated from LGE and T2w to the cine ED frame through the MvMM transforms and the cine ED myocardium contours (the same contours used by the anatomy module). Report mean, median, and worst-case overlap per center. If mean myocardium overlap is below 0.85 or any test case falls below 0.6, re-run the Table V comparison on the subset with verified alignment; if CineMyoPS no longer beats all four baselines on the primary Dice metric, the gold-standard registration is the load-bearing weakness. If mean overlap exceeds 0.9 with no severe outliers, the concern is mitigated and the conditional verdict can stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A states that gold-standard pathology labels were produced by registering rater-annotated LGE (scar) and T2w (edema) images to the cine ED frame with the MvMM tool, using structural information from all three modalities. No registration accuracy metric is reported anywhere in the paper. The inter-observer Dice values (0.69 for scar, 0.73 for edema) and the additional human observer's Dice (0.73/0.75) quantify label ambiguity among human raters on the fused labels, but they do not measure how accurately the LGE/T2 coordinates map to the cine ED space. CineMyoPS is trained and evaluated against these fused labels; systematic misregistration due to breath-hold differences, through-plane motion, or slice-thickness mismatch (e.g., Center-R LGE/T2 slice spacing of 10-17 mm versus cine 6 mm in Table II) would corrupt both training targets and test labels. In that case, the model would be learning to predict misaligned targets, and the reported test Dice of 0.53 for scar and 0.57 for edema, as well as the claimed superiority over the four baselines, would not reflect true pathology segmentation from cine. This is not a restatement of the segmentation claim; it is an unvalidated assumption about the evaluation target itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CineMyoPS, an end-to-end network that segments myocardial scar and edema from cine CMR only. It combines a motion estimation module, an anatomy segmentation module, and a pathology segmentation module, with a consistency loss to couple motion and anatomy features and a time-series aggregation over a subset of the cardiac cycle. Experiments on 145 patients from three centers (training and validation on Center-M/Center-Z, held-out test on Center-R) report Dice scores of 0.53 for scar and 0.57 for edema, together with comparisons against nnU-Net, OFSeg, ConvLSTM, and 2D+1D U-Net trained under the same protocol. The paper also includes ablation studies on frame proportion, feature combinations, the consistency loss, and a transmurality correlation analysis.","tokens_in":19561,"tokens_out":8747,"duration_ms":103334,"significance":"If the results hold, the paper provides a useful demonstration that cine-only CMR can support joint scar and edema segmentation in a multi-center setting, and the controlled same-protocol comparison with four baselines is a strength. The authors are also transparent about limitations, including the weak correlation in nonviable transmurality (R=0.22, p=0.17) and difficulties in apical slices. However, the significance is currently undercut by two issues: the evaluated model omits texture features that are part of the method description, and the gold-standard labels rely on an unvalidated multi-modal registration step. These are fixable but require clarification and additional evidence.","major_comments":[{"comment":"The method section describes the MyoPS module as fusing motion, anatomy, and texture features, with Eq. (9) concatenating [Φ_i, L^a_i⊗Φ_i, I_i⊗Φ_i]. However, the feature effectiveness study in Section III-C.2 concludes that MyoPSΦL (motion and anatomy, without texture) is the best variant and states that motion and anatomy features were adopted for MyoPS in the following sections. Consequently, the test-set results in Table V and the clinical quantification in Section III-E correspond to a different model than the one specified in Eq. (9) and Fig. 4. Please revise the method description to match the evaluated model, or report both variants, and state explicitly in Eq. (9) which inputs are used by the final CineMyoPS model.","section":"Section II-C, Eq. (9), Fig. 4, Section III-C.2"},{"comment":"The gold-standard pathology labels are generated by registering rater-annotated LGE and T2w images to the cine ED frame with the MvMM tool, but no registration accuracy metric is reported anywhere in the paper. The reported inter-observer Dice values (0.69 for scar, 0.73 for edema) quantify label variability, not the accuracy of the LGE/T2-to-cine mapping. Given the slice-thickness mismatch at Center-R (LGE/T2 slice spacing 10–17 mm versus cine 6 mm in Table II), misregistration would directly bias the training targets and every test metric in Table V. Please add quantitative registration validation, such as Dice of transformed myocardial contours, target registration error on anatomical landmarks, or a documented visual inspection protocol, and specify how the registered LGE and T2w labels are fused into the final gold standard.","section":"Section III-A, gold standard label generation"},{"comment":"The claim that CineMyoPS outperforms the four baselines should be tempered. In Table V, the edema Dice differences between CineMyoPS (0.57) and OFSeg, ConvLSTM, and 2D+1D U-Net (0.55, 0.56, and 0.56, respectively) are not statistically significant, and the scar Dice gap to 2D+1D U-Net (0.53 versus 0.50) is also not marked as significant. The significant gains are mainly against nnU-Net and on specific metrics. Please soften the wording in Section III-D.1 and in the abstract, or provide an error analysis that explains where the method genuinely improves over the strongest baselines.","section":"Section III-D, Table V"}],"minor_comments":[{"comment":"Center-R is listed with 50 subjects, but the TR/VA/TE row gives 0/0/45; please clarify the number of test subjects or correct the table.","section":"Table I"},{"comment":"In the last paragraph of the introduction, the stray word 'structure' appears before 'Overall'; this appears to be a typesetting artifact and should be removed.","section":"Section I"},{"comment":"The 'cos' operation in Eq. (8) is not explicitly defined; please define cosine similarity for the predicted probability maps and state whether it is computed per class or per channel.","section":"Section II-B, Eq. (8)"},{"comment":"The network is described as fully automatic, but the paper does not explain how the ED reference frame is selected during inference; if manual selection is required, please state this and discuss the possibility of automatic ED detection.","section":"Section II-A and Section III-A"},{"comment":"Table VI compares reported Dice values from different datasets and evaluation protocols; the text notes dataset differences, but a table footnote should explicitly warn that the literature numbers are not directly comparable to the numbers reported in this paper.","section":"Section III-D.2, Table VI"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript is within the scope of TMI and the empirical setup is largely sound. My recommendation of major revision hinges on the method/evaluated-model mismatch and the lack of registration validation; both are addressable with additional experiments and clarifications. I do not see grounds for rejection. The literature-comparison and 'first fully automatic' claims may also need moderation in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is the first to attempt joint scar and edema segmentation from cine CMR alone, and the within-dataset comparisons are credible. The four baselines are trained under the same protocol, and CineMyoPS beats them consistently (scar Dice 0.53 vs 0.42 for nnU-Net). That is a credible result for a new task.\n\nWhat the paper does well: the ablations are informative. The consistency loss helps (0.05 Dice scar), the time-series aggregation helps (0.08 scar), and the feature study shows motion is the most useful single feature and that texture actually hurts when added to motion plus anatomy. The authors are honest about the limitations: test Dice are low in absolute terms, apical slices are hard, and the nonviable-transmurality correlation is essentially zero (R=0.22, p=0.17). They do not oversell.\n\nThe biggest soft spot is a mismatch between the method as described and the method as evaluated. Equation (9) feeds texture (the original image intensity) into the MyoPS module, and Figure 4 says the same. But the Table III ablation shows that adding texture to motion and anatomy reduces performance, and the authors then say they adopted motion and anatomy features for the final model. So the evaluated system is not the one derived in the methods. On top of that, Table III has a checkmark/name inconsistency: the row labeled MyoPSIΦ shows checkmarks for Φ and L, and the row labeled MyoPSΦL shows checkmarks for I and Φ. It is hard to be sure which feature combination produced the reported 0.54 scar Dice. A referee should force them to clean this up.\n\nThe other issue is the gold standard. The LGE and T2 labels are registered to the cine ED frame with MvMM, and the paper reports no registration accuracy at all. Slice spacing for LGE/T2 in Center-R is 10–17 mm versus 6 mm for cine, so real misalignment is possible. The inter-observer Dice (0.69 scar, 0.73 edema) reflects label ambiguity, not coordinate alignment. If the registration is off, every Dice/HD number in the paper measures the wrong target. This is an unvalidated assumption about the evaluation, not a quibble.\n\nThere is also no code or data release yet, so the numbers cannot be independently checked.\n\nWho should read this: anyone working on contrast-free cardiac MRI. It is a genuinely new task and the experimental design is mostly sound. The paper deserves peer review but needs revision: reconcile the architecture description with the evaluated model, fix the table, and validate or at least discuss the registration accuracy.","headline":"First joint scar/edema segmentation from cine CMR with fair baselines, but the evaluated model doesn't match the described one and the label registration is unvalidated.","tokens_in":20176,"tokens_out":5243,"would_cite":true,"duration_ms":52070,"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":"CineMyoPS claims that a single contrast-free cine cardiac MR sequence is enough to jointly segment myocardial scar and edema.","keywords":["myocardial infarction","cine cardiac MRI","scar segmentation","edema segmentation","motion estimation","contrast-free CMR","deep neural network","time-series aggregation"],"falsifier":"An external test set whose scar and edema labels come from histology, or from an independent registration pipeline rather than the MvMM tool, would settle the claim: if CineMyoPS's Dice score against those labels falls to the level of the baselines, the reported advantage is an artifact of the fused gold standard.","tokens_in":19104,"feed_emoji":"🫀","tokens_out":6156,"duration_ms":67124,"temperature":0.7,"pith_summary":"CineMyoPS sets out to prove that myocardial scar and edema — two markers of heart attack damage normally seen only on contrast-enhanced LGE and T2-weighted MRI — can be segmented from cine cardiac MR alone, a rapid, contrast-free sequence. The paper argues that motion, anatomy, and time-series texture cues in a cine loop carry enough information to replace multi-sequence imaging for area-at-risk assessment. To show this, it trains an end-to-end network that estimates heart motion between frames, segments myocardial anatomy, and aggregates features across the cardiac cycle, with a consistency loss tying the two learned tasks together. On a multi-center test set, the method reaches Dice scores of 0.53 for scar and 0.57 for edema, higher than four baselines trained under the same protocol. If these results hold, contrast agents and long scanning sessions could be avoided in routine infarct assessment.","feed_headline":"One contrast-free cine MRI sweep maps scar and edema","feed_subtitle":"New end-to-end network CineMyoPS posts test Dice 0.53 for scar, 0.57 for edema, beating four baselines.","key_machinery":"The load-bearing mechanism is a three-module network: a motion estimation module built as a U-shaped registration subnetwork that predicts a dense displacement field from each frame to the end-diastolic reference, an anatomy segmentation module, and a MyoPS module that concatenates motion, anatomy, and texture features in the reference image space. Two devices carry the argument: a consistency loss, defined as cosine distance between the anatomy prediction warped by the motion field and the reference anatomy label, which co-trains motion and anatomy; and a time-series aggregation strategy that sums per-frame pathology predictions across the cardiac cycle and passes them through softmax(conv(·)). This lets the network exploit the interdependence of motion and structure while integrating information from multiple cardiac phases.","core_discovery":"The paper's central claim is that CineMyoPS is the first fully automatic network to segment both scars and edema from cine CMR images in an end-to-end fashion. It finds that explicit motion features — dense displacement fields produced by a registration subnetwork — are the single most informative cue for pathology segmentation, outperforming texture and anatomy features, and that combining motion with anatomy features gives the best overall results. The authors further report that the proposed consistency loss and time-series aggregation each produce statistically significant gains over ablated variants, and that on the held-out test center CineMyoPS outperforms nnU-Net, OFSeg, ConvLSTM, and 2D+1D U-Net under the same training protocol, with test Dice scores of 0.53 for scar and 0.57 for edema.","pith_inferences":["Editorial inference: the reported Dice ceiling sits near the inter-observer agreement (0.69 for scar, 0.73 for edema), so the practical limit of cine-only pathology segmentation may be partly set by label ambiguity rather than by the network; better label fusion could lift apparent performance without changing the model.","Editorial inference: since the model was trained on pre- and post-contrast cine but tested only on pre-contrast cine, the texture-feature drop may reflect a domain gap; training solely on pre-contrast cine, or explicit domain adaptation, is a natural next experiment.","Editorial inference: the apical-slice failures and the failure in nonviable-region transmurality suggest the method's errors concentrate in thin-wall, high-curvature regions; a slice-aware or shape-constrained extension is a testable route to improvement."],"forward_implications":["If correct, a full area-at-risk assessment for myocardial infarction could be obtained from a single rapid, contrast-free cine acquisition, removing gadolinium injection and roughly half the scan time associated with LGE.","Motion and anatomy features are the effective carriers of scar and edema information; texture alone adds noise once both are present, so future cine-based methods should prioritize deformation and structural cues.","Adding temporal frames helps up to a point: using four of six frames in the cardiac cycle reaches a plateau, giving a computational budget for future models.","The consistency loss means only the end-diastolic frame needs manual anatomy labeling during training, reducing annotation burden for cine-sequence segmentation.","Transmurality estimates correlate well with manual delineation in viable and intermediate regions but fail in fully nonviable myocardium, so the method supports screening but not yet full infarct-extent quantification."],"supporting_citations":[{"why":"Produces the gold-standard pathology labels by registering LGE and T2w raters' annotations to the cine end-diastolic frame.","marker":"[37]"},{"why":"Provides the U-shaped registration backbone that estimates the dense displacement fields used as motion features.","marker":"[31]"},{"why":"Supplies the U-Net backbone for the anatomy and pathology segmentation modules.","marker":"[26]"},{"why":"Supplies the joint segmentation and motion-estimation schema on which the consistency loss is built.","marker":"[33]"},{"why":"Establishes the motion-estimation-based approach to non-contrast scar segmentation that CineMyoPS extends.","marker":"[20]"},{"why":"Defines the nnU-Net baseline and the shared training strategy for all compared methods.","marker":"[38]"},{"why":"Provides the ConvLSTM spatio-temporal baseline used in the comparison study.","marker":"[23]"},{"why":"Provides the 2D+1D U-Net spatio-temporal baseline used in the comparison study.","marker":"[41]"}],"fun_headline_variants":["Cine MRI alone segments scar and edema using motion and anatomy","Contrast-free cine MRI maps scar and edema in one sweep","Motion cues drive cine MRI segmentation of scar and edema","Single cine MRI sweep identifies scar and edema without contrast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pathology labels used as ground truth are accurate enough to train and test on, even though they were created by registering rater annotations from LGE and T2w images onto the cine frames.","fun_headline_variants_meta":{"raw":{"variants":["Cine MRI alone segments scar and edema using motion and anatomy","Contrast-free cine MRI maps scar and edema in one sweep","Motion cues drive cine MRI segmentation of scar and edema","Single cine MRI sweep identifies scar and edema without contrast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00062,"raw_usage":{"total_tokens":2878,"prompt_tokens":950,"completion_tokens":1928,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1861}},"tokens_in":566,"tokens_out":1928,"duration_ms":17156,"temperature":1.0,"reasoning_tokens":1861,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:33:32.408820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An external test set whose scar and edema labels come from histology, or from an independent registration pipeline rather than the MvMM tool, would settle the claim: if CineMyoPS's Dice score against those labels falls to the level of the baselines, the reported advantage is an artifact of the fused gold standard.","supporting_citations":[{"cited_title":"Multivariate mixture model for myocardial segmentation combining multi-source images,","cited_arxiv_id":null,"evidence_quote":"Produces the gold-standard pathology labels by registering LGE and T2w raters' annotations to the cine end-diastolic frame."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net backbone for the anatomy and pathology segmentation modules."},{"cited_title":"Joint learning of motion estimation and segmentation for cardiac mr image sequences,","cited_arxiv_id":null,"evidence_quote":"Supplies the joint segmentation and motion-estimation schema on which the consistency loss is built."},{"cited_title":"Deep learning for diagnosis of chronic myocardial infarction on nonenhanced cardiac cine mri,","cited_arxiv_id":null,"evidence_quote":"Establishes the motion-estimation-based approach to non-contrast scar segmentation that CineMyoPS extends."},{"cited_title":"Convolutional lstm network: A machine learning approach for precipitation nowcasting,","cited_arxiv_id":null,"evidence_quote":"Provides the ConvLSTM spatio-temporal baseline used in the comparison study."},{"cited_title":"Learning spatio-temporal representation with pseudo-3d residual networks,","cited_arxiv_id":null,"evidence_quote":"Provides the 2D+1D U-Net spatio-temporal baseline used in the comparison study."}],"review_version":1}