{"id":"bc711e6d-ce03-4675-83c3-4bed1311bc6a","arxiv_id":"2510.05819","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised cardiac motion descriptor derived from deformable registration detects five cardiac keyframes in short-axis and four-chamber MRI, outperforming volume-based end-systole/end-diastole detection on most tested datasets.","lead":"The paper uses self-supervised image registration to build a one-dimensional motion curve of the heart from cardiac MRI cine frames, then detects five keyframes in each heartbeat with simple rules. It reports that this motion-based detection beats a volume-curve baseline on several public cardiac MRI datasets, especially in short-axis views.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Keyframe rule set in §2.3 is circular/inconsistent as written, so the reported cFD results are not reproducible from the method description.","rationale":"Read full text. The main empirical claim — motion-descriptor detection of ED/ES and five keyframes beats the LV-volume baseline — rests on a rule-based detector. In §2.3, the equations for ES, PF, ED, and MD use mutually referential cyclic intervals: ES is sought in [MS, PF] while PF is sought in [ES, MS]. Without an explicit ordering or disambiguation, the rule set cannot be instantiated. This is not a trivial typo because the intervals also constrain the search for ED and MD. The reader's conditional verdict noted ambiguity in these rules, but its weakest-assumption analysis focused on focus-point placement; I view the rule inconsistency as more load-bearing because it blocks verification of the primary numbers. The focus-point issue is a plausible concern for interpretability and 4CH performance, but the SAX advantage over baseline would survive even if some C_mse points sit near the myocardium; an unexecutable rule set invalidates the quantitative comparison until resolved. Concrete fix: re-implement from the published equations on the provided median curves; if it fails, the authors should release code or corrected pseudocode. Given this is a clarification/code-release issue rather than a demonstration that the hypothesis is false, the reader's CONDITIONAL verdict remains appropriate (UNCHANGED).","tokens_in":21079,"tokens_out":9444,"duration_ms":84917,"concrete_test":"Implement the §2.3 rules literally on the median α curves shown in Figures 4/5 (or on a synthetic crescent-shaped curve) and check whether the output is five ordered, non-duplicate keyframes in physiological sequence. If literal implementation fails or diverges from the paper's reported cFD values, inspect the GitHub repository's source code to recover the actual rule ordering; if no code or corrected pseudocode is provided, request clarification and treat the reported cFD numbers as unverified until the rule set is disambiguated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 defines keyframe detection rules that are internally inconsistent: ES is defined as max{α(t)=0, α′(t)>0} over t∈[MS, PF], while PF is simultaneously defined as min{α′(t)=0, α″(t)<0} over t∈[ES, MS]. Since PF appears in the interval for ES and ES in the interval for PF, neither can be computed without already knowing the other. ED and MD have analogous circular dependencies. If these intervals are interpreted literally, the rule set is not executable; if they are meant to be cyclic or to refer to subsequent/previous cardiac cycles, that ordering is not stated. Because the cFD values in Tables 3–4 and the reported superiority over the LV-volume baseline are produced by an algorithm that is not specified by these equations, the central claim cannot be independently verified from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a fully self-supervised method for detecting five cardiac keyframes (end-diastole, mid-systole, end-systole, peak flow, mid-diastole) in short-axis (SAX) and four-chamber (4CH) cine CMR. A deformable registration U-Net estimates dense displacement fields; a 1D motion descriptor α_t is computed by masking the field and aggregating voxel-wise cosine similarities between displacement vectors and position vectors pointing to a focus point. Five keyframes are detected from α_t via a rule set based on zero crossings and extrema. The method is evaluated on M&Ms-2, M&Ms, ACDC, and a tetralogy-of-Fallot cohort (GCN), and compared against a supervised LV-volume-based baseline. The central hypothesis is that motion-based phase detection outperforms volume-based detection.","tokens_in":21361,"tokens_out":9084,"duration_ms":70339,"significance":"If the results hold, the work would provide a label-free, scanner-generalizable alternative to volume-based cardiac phase detection, with additional keyframes enabling phase-aligned analysis. Its strengths include evaluation on multiple public multi-center, multi-disease datasets, including rare congenital heart disease, and the stated intent to release code and annotations. However, two issues currently block acceptance: the keyframe rule set is not an executable specification as written, and the abstract's 4CH improvement claim is not supported by the fully self-supervised results.","major_comments":[{"comment":"The keyframe rule set is circular and not executable as written. ES is defined as max{α(t)=0, α′(t)>0} over t∈[MS, PF], while PF is simultaneously defined as min{α′(t)=0, α″(t)<0} over t∈[ES, MS]. ED depends on PF and MD depends on ED. Therefore each equation references a keyframe that is itself defined by the same rule set, so the system cannot be evaluated without already knowing the solution. The text says the rules are applied as a sequence to the cyclic sub-sequence, but the ordering and cyclic-index convention are not stated. Since the reported cFD results in Tables 3–4 are produced by this rule set, the method is not reproducible from the manuscript alone. Please replace the equation block with an explicit, executable algorithm, e.g., pseudo-code that specifies the detection order (e.g., locate MS, then ES as the first zero crossing after MS, then PF as the first maximum after ES,","section":"§2.3"},{"comment":"The abstract claims 'improved detection accuracy by 30%–51% for SAX and 11%–47% for 4CH in ED and ES ... compared with the volume-based approach.' For the fully self-supervised focus point C_mse, Table 4 shows no improvement over the LV-volume baseline on M&Ms-2 test (ED: 0.94±1.11 vs 0.93±1.20; ES: 0.99±1.09 vs 0.91±1.20) and only marginal improvements on M&Ms-2 train. The 11% lower bound of the 4CH range does not appear to be supported by any C_mse result; it may stem from segmentation-derived focus points such as C_sept or C_lv. This overstatement concerns the central claim that the motion-descriptor method is more accurate than volume-based detection. Please restrict the summary claim to the specific focus-point variant used, or report the ranges separately for fully self-supervised and segmentation-derived focus points.","section":"Abstract / Table 4"}],"minor_comments":[{"comment":"The preprint header contains a typo: 'Medical Imaga Analysis' should be 'Medical Image Analysis'.","section":"Header"},{"comment":"There is an inconsistency in p-value reporting. The text says for M&Ms ED p<0.1e-3, but Table 3 shows '**' (p<0.01). Please standardize the reporting.","section":"§3 / Table 3"},{"comment":"The note 'In case of the M&Ms-2 test dataset, the mean cFD would be 0.82±0.89 if only ED and ES are considered like it is the case for base' does not state which focus point this refers to. Please specify (presumably C_mse) and clarify the comparison.","section":"Table 4"},{"comment":"The statement that the rule set 'has been further optimized to achieve optimal performance for healthy hearts' is not accompanied by a description of the optimization procedure, the data used, or whether labels were consulted. This is important for reproducibility and for assessing potential bias toward healthy subjects. Please describe the tuning process.","section":"Discussion"},{"comment":"The terms 'LAX' and '4CH' are used interchangeably (Abstract uses 'LAX', rest of the paper uses '4CH'). Please unify.","section":"Abstract"},{"comment":"The cFD formula can be simplified to min(|p_i−p̂_i|, T−|p_i−p̂_i|). Consider using this more readable form.","section":"§2.5 Eq. (12)"},{"comment":"The number of keyframe annotations per dataset and the inter-observer variability for the additional keyframes (MS, PF, MD) are not fully reported. Please add this information, as it is essential for judging the reliability of the 'all keyframes' results.","section":"§2.4 / Tables"}],"recommendation":"major_revision","confidential_remarks":"The SAX core result appears sound and is consistently supported across four datasets with paired significance tests. The main concerns are the ambiguous rule specification and the overstated 4CH claim in the abstract. Both are fixable in revision. I would encourage the editor to seek a clarified algorithm description and a corrected abstract before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The SAX story is solid: motion-descriptor ED/ES beats the LV-volume baseline on four datasets with paired Wilcoxon tests, and the five-keyframe extension is a real step beyond Koehler 2022a. The multi-dataset external validation, including the GCN tetralogy cohort, is the right way to test generalisation. Credit where due.\n\nBut the stress-test note is right. Section 2.3 defines ES on [MS, PF], PF on [ES, MS], ED on [PF, MS], MD on [PF, ED]. Those intervals are mutually referential. You cannot compute any of these without already knowing another. The paper never states whether the intervals are meant cyclically or in a different cardiac cycle, so the algorithm as printed is not executable. That is a reproducibility blocker. It is likely a presentational slip rather than a fundamental flaw — the figures show a sensible ordering — but as written the cFD numbers in Tables 3-4 are not tied to the equations.\n\nSecond, the abstract overclaims for 4CH. Using the fully self-supervised C_mse focus point on M&Ms-2 test, ED/ES cFD are 0.94/0.99 versus base 0.93/0.91 — no improvement. The 11–47% improvement only shows up with anatomical focus points (C_sept/C_lv) or on the GCN dataset. The abstract's wording hides that.\n\nThird, there is post-hoc selection. Mask thresholds (T_norm, T_Δα) were picked empirically, the rule set was \"further optimized\" for healthy hearts, and the best focus point is reported per dataset. Without a validation split or pre-registered choices, the headline numbers are optimistic. The 4CH story is also weaker because C_mse often lands in the atria; the authors attribute the degraded performance to atrial/ventricular motion overlap, which is plausible but means the descriptor's physiological meaning is view-dependent.\n\nNone of this kills the core SAX claim. The SAX ED/ES results are consistent and within or near inter-observer variability. The addition of MS, PF, MD is useful for phase-resolved analysis and aligns with the authors' earlier strain work. I would send this to peer review, but reviewers should insist on a corrected, unambiguous rule order, a proper validation split for the threshold choices, and a recalibrated abstract. If those are fixed, this is a worthwhile contribution.\n\nWho is it for: anyone working on cardiac phase detection, temporal alignment, or self-supervised motion analysis in CMR. It will be cited.","headline":"Solid SAX evidence and a useful five-keyframe extension, but the §2.3 rule equations are circular as written and the 4CH abstract claim overreaches — fixable before this should be trusted.","tokens_in":21846,"tokens_out":3497,"would_cite":true,"duration_ms":28784,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A motion descriptor built from deformable image registration detects five cardiac keyframes in cine CMR and beats left-ventricular volume curves for end-diastole and end-systole.","keywords":["cardiac keyframe detection","deformable image registration","self-supervised learning","motion descriptor","cine CMR","cardiac magnetic resonance imaging","end-diastole and end-systole","multi-view cardiac imaging"],"falsifier":"A controlled experiment that fixes the registration model but perturbs the focus point across chamber boundaries would settle the load-bearing assumption: if keyframe error stays flat when the focus point moves outside the left ventricle, the interpretable contraction/relaxation sign is not what is carrying the accuracy; if error rises, the assumption is confirmed.","tokens_in":20977,"feed_emoji":"🫀","tokens_out":8744,"duration_ms":57816,"temperature":0.7,"pith_summary":"The paper proposes that cardiac phase can be read directly from the direction of myocardial motion rather than from changes in ventricular volume. It builds dense displacement fields between consecutive cine frames using an unsupervised deformable registration network, then compresses each field into a single curve that is negative while the heart contracts and positive while it relaxes. From zero-crossings and extrema of this curve, the method flags five keyframes: end-diastole, mid-systole, end-systole, peak filling, and mid-diastole. Across several multi-centre, multi-disease datasets plus a rare congenital-heart-defect cohort, the method detects ED and ES with 30–51% lower error than a supervised volume-based baseline in short-axis views, and it still finds the extra keyframes within about 1.3 frames in short-axis and 1.7 frames in long-axis views. If these results hold, cardiac MRI analysis could be temporally aligned across patients and cycles without manual labels or ECG.","feed_headline":"One motion curve finds five cardiac phases in MRI","feed_subtitle":"Self-supervised motion curves find five keyframes, beating the volume-curve baseline without labels or ECG.","key_machinery":"The carrying object is the deformable motion descriptor, a 1D curve computed from the dense vector field produced by a convolutional registration network. Per voxel, the cosine similarity is taken between the displacement vector and a position vector pointing from that voxel to a focus point; a rule-based mask keeps only voxels with persistent large motion and strong directional change, and the descriptor is the masked mean. The focus point is computed self-supervised as the temporal mean centre of mass of the masked field, so the whole pipeline needs no labels. A simple rule set converts the curve's shape into keyframes: the global minimum is mid-systole, zero crossings with appropriate slo","core_discovery":"The central claim is that myocardial displacement fields, not chamber-volume curves, are the right signal for timing the cardiac cycle. The paper shows this by defining a one-dimensional descriptor that averages, over a motion-masked region, the cosine similarity between each voxel's displacement and a vector pointing toward a focus point; negative values mean contraction, positive values mean relaxation. The smoothed descriptor reliably crosses zero at end-diastole and end-systole, and its global minimum plus derivative extrema identify mid-systole, peak filling, and mid-diastole. On large multi-centre, multi-disease evaluations, this self-supervised rule-based method detects ED and ES with","pith_inferences":["Editorial extension: the continuous descriptor curve could serve as a cardiac-phase clock, so any intermediate phase could be interpolated from its shape rather than limited to five named keyframes.","Editorial extension: if atrial/ventricular motion overlap is the main cause of weaker long-axis results, separate focus points for ventricles and atria could isolate chamber-specific contraction and improve accuracy.","Editorial extension: the same descriptor idea may transfer to other dynamic imaging modalities with a contracting focus, such as echocardiography, where the smaller field of view would need a different focus-point strategy.","Editorial extension: because the masking thresholds were chosen empirically, a per-patient adaptive threshold is a direct testable extension for low-resolution or artefact-heavy acquisitions."],"forward_implications":["Keyframe detection becomes possible without manual labels or ECG, using only cine images and a registration model trained on unlabelled frames.","End-diastole and end-systole are found within about one frame in short-axis cine MRI across unseen centres, comparable to inter-observer variability, making automated temporal alignment practical.","The same one-dimensional curve yields three additional physiologically meaningful keyframes, enabling phase-resolved analysis beyond the traditional ED–ES pair.","Because the registration is self-supervised and the phase rules are fixed, the method transfers to unseen scanners, pathologies, and a rare congenital heart defect cohort without retraining.","Replacing volume curves with motion curves addresses a known blind spot: iso-volumetric contraction and relaxation change myocardial motion without changing chamber volume.","The approach enables inter- and intra-patient comparison of cardiac dynamics at aligned phases regardless of cycle length or starting phase."],"fun_headline_variants":["Motion curves beat volume curves for cardiac phase detection","Self-supervised motion curves find five cardiac MRI keyframes","Deformable registration spots 5 heart phases without labels","MRI motion descriptor times cardiac cycle better than volume","Five keyframes from motion, not volume, in cardiac MRI"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes the automatically chosen focus point sits inside the left ventricle, so that motion toward it means contraction; in several long-axis cases it lands in the atria, and then the sign of the descriptor no longer cleanly separates contraction from relaxation.","fun_headline_variants_meta":{"raw":{"variants":["Motion curves beat volume curves for cardiac phase detection","Self-supervised motion curves find five cardiac MRI keyframes","Deformable registration spots 5 heart phases without labels","MRI motion descriptor times cardiac cycle better than volume","Five keyframes from motion, not volume, in cardiac MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1477,"prompt_tokens":890,"completion_tokens":587,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":509}},"tokens_in":634,"tokens_out":587,"duration_ms":5422,"temperature":1.0,"reasoning_tokens":509,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:14:21.711961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that fixes the registration model but perturbs the focus point across chamber boundaries would settle the load-bearing assumption: if keyframe error stays flat when the focus point moves outside the left ventricle, the interpretable contraction/relaxation sign is not what is carrying the accuracy; if error rises, the assumption is confirmed.","supporting_citations":[],"review_version":1}