{"id":"78bfd391-3fbb-4515-a5b0-e306dd4c942a","arxiv_id":"2507.17678","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MCM uses a bidirectional Mamba encoder and motion decoder to estimate cardiac motion from sequential cine MRI frames, improving Dice and smoothness over single-frame-pair baselines.","lead":"This paper introduces MCM, a Mamba-based neural network that tracks heart motion in MRI by using a short sequence of frames around each target, not just single image pairs. It reports smoother and slightly more accurate motion fields than existing methods on two public cardiac datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal-consistency advantage rests on qualitative evidence; the paper's own discussion shows only +0.5% Dice when Nf grows from 1 to 5.","rationale":"The reader's weakest assumption is exactly the load-bearing weakness: quantitative evaluation covers only ED-to-ES on three slices, while the method's claimed improvement is temporal consistency across the cardiac cycle. The manuscript itself flags this limitation in the Discussion (ground truth only at ED and ES; +0.5% Dice when Nf is increased; temporal consistency shown only in Fig. 4). I do not see an internal inconsistency or a flaw that would justify rejection. The architecture is plausible, and the ED-ES results are consistent, though without significance tests the magnitude of gains is uncertain. A full-cycle quantitative benchmark, even synthetic, would settle whether the sequence-based design provides a real temporal advantage. Since the reader already issued CONDITIONAL with the same condition, the verdict should remain unchanged.","tokens_in":9523,"tokens_out":4421,"duration_ms":51128,"concrete_test":"Create a ground-truth full-cycle benchmark by taking ACDC/M&Ms ED slices and applying a known periodic radial contraction/expansion deformation to generate a synthetic cine sequence with known motion at every phase. Train MCM and the strongest baseline (TransMorph or MambaMorph) under the same protocol, then compute per-frame endpoint error and temporal smoothness of tracked trajectories across all phases. If MCM does not reduce intermediate-frame error or trajectory jerkiness relative to single-pair baselines, the temporal-consistency advantage is unsupported. As a secondary check on the published ED-ES numbers, run a paired bootstrap test on subject-level Dice differences between MCM and TransMorph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that sequence-based target frames yield smooth, temporally consistent cardiac motion. All quantitative results (Table 1, Table 2) are computed only for the ED-to-ES transition on three slices, because manual segmentations exist only at ED and ES. The only evidence for whole-cycle temporal consistency is Fig. 4, a qualitative sampling-line visualization; it is not compared against any baseline and carries no error bars. Moreover, the Discussion concedes that increasing Nf from 1 to 5 yields only modest quantitative gains (+0.5% Dice) while producing 'visibly smoother' fields. Thus the reported gains on ED-ES Dice could come from increased model capacity or input diversity rather than from temporal consistency, and the central advantage of MCM is not quantitatively established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes MCM, an unsupervised deep network for 2D short-axis cine MRI cardiac motion tracking. MCM registers the end-diastolic reference frame I0 to a target frame It using, in addition to I0 and It, a sequence of K neighboring target frames (Nf = 2K+1; K = 2 in the experiments). The paired input sequence is processed by a hierarchical encoder of bi-directional Mamba blocks, whose bi-directional scanning Mamba (BiSM) applies forward and backward SSM scans, followed by a motion decoder with a dual-path fusion head (DFH) that combines forward/backward Conv3D fusion with 2D convolution to emit the deformation field Φt. The model is trained with an MSE similarity loss plus a gradient-based smoothness regularizer (Eq. 6). Experiments on ACDC and M&Ms compare MCM against dDemons, VoxelMorph, TransMorph and MambaMorph on three short-axis slices (basal, mid, apical) for ED-to-ES tracking, reporting Dice, |J|<0% and ||J|−1|. The paper reports higher Dice and lower Jacobian-based regularity metrics for MCM on both datasets, ablation results for BiSM, DFH, Nf and λ, and inference-time/VRAM measurements. The code is released.","tokens_in":9613,"tokens_out":12171,"duration_ms":118541,"significance":"If the reported results hold, this is a useful incremental contribution to cardiac motion tracking: it demonstrates that a Mamba-based architecture consuming a short window of neighboring target frames achieves slightly higher ED-ES Dice and markedly lower |J|<0% than strong registration baselines on two public datasets, with essentially flat inference time as Nf grows. The paper's strengths are that all baselines are run from their official implementations, the ablations isolate the contributions of the bi-directional scanning and the fusion head, the λ sensitivity is reported, and the code is released, which makes the main quantitative claims reproducible. The main limitation is that the paper's signature claim of smooth, temporally consistent tracking across the whole cardiac cycle is supported only by a qualitative figure (Fig. 4), while the quantitative evaluation covers only the ED-ES transition; the paper's own Discussion acknowledges that increasing Nf from 1 to 5 improves Dice by only about 0.5 percentage points.","major_comments":[{"comment":"The central claim that sequence-based tracking yields 'smooth and temporally consistent' motion (abstract and contribution 1) is not quantitatively established. All quantitative comparisons in Tables 1 and 2, and the ablations in Table 2, evaluate only the ED-to-ES transition on three slices, because ground-truth contours exist only at ED and ES. The only whole-cycle evidence is the qualitative sampling-line visualization in Fig. 4, which contains no error bars, no comparison with any baseline method, and no statement of what quantity is plotted; the Discussion itself acknowledges that increasing Nf from 1 to 5 yields only modest quantitative gains (e.g., +0.5% Dice) while producing 'visibly smoother' fields. This leaves open the alternative explanation that the Dice gains come from increased model capacity or input diversity rather than from temporal modeling, so the paper's central differentiator is not supported by the presented evidence. I would ask for a quantitative temporal-consistency evaluation across all frames of the cycle, for example temporal smoothness statistics of the displacement magnitude or Jacobian determinant, or contour-propagation consistency against the ED and ES annotations, with the Nf=1 configuration and at least one single-pair baseline plotted alongside MCM in Fig. 4.","section":"Discussion; Sec. 3 (Comparison study); Fig. 4"},{"comment":"The Dice gains over the strongest baselines are around one percentage point with overlapping standard deviations (e.g., ACDC mid: Ours 84.6(4.9) vs TransMorph 83.7(4.9); M&Ms apical: Ours 77.6(11.5) vs TransMorph 76.4(11.7)), and the paper reports no statistical significance tests. It is also not stated whether the mean and standard deviation in Table 1 are computed over subjects or over individual slices, which directly affects the effective sample size in a per-slice analysis. A paired non-parametric test over subjects (e.g., Wilcoxon signed-rank or bootstrap) is needed to support the claim in Sec. 3 that the proposed method 'outperforms all baseline methods'.","section":"Table 1"},{"comment":"The abstract's claim that the method learns the continuous dynamics of the myocardium 'without increasing computational complexity' is at odds with Table 3, which reports GPU VRAM growing from 3.2 GB (Nf=1) to 7.8 GB (Nf=3) to 12.4 GB (Nf=5), roughly a fourfold increase, and which lists only inference-time VRAM for the baselines (1.5-3.6 GB), not training VRAM. The statement that is actually supported by Table 3 is that inference time is roughly constant as Nf grows (16.3 to 17.1 ms); the abstract and contribution (1) should be qualified accordingly, and the Discussion's acknowledgment of 'the cost of increased GPU memory usage' should be reflected in the paper's framing.","section":"Abstract and contribution (1); Table 3"}],"minor_comments":[{"comment":"The citation labels in Table 3 are inconsistent with Table 1 and the reference list: 'VM [23]' should be 'VM [1]' and 'TM [1]' should be 'TM [7]'.","section":"Table 3"},{"comment":"Reference [2] attributes 'A New Approach to Linear Filtering and Prediction Problems' to T. Basar; this is R.E. Kalman's 1960 paper, and the citation should be corrected.","section":"References [2]"},{"comment":"The description of the bi-directional scanning does not specify the exact token ordering fed to the SSMs, namely whether the sequence is over the Nf frames for each spatial position, or a flattened Np×Nf sequence, and in which interleaving order; the paper should state this precisely.","section":"Sec. 2.1 (BiSM)"},{"comment":"The configurations of the 'Without BMBs' and 'without DFH' ablations are not described; the reader cannot tell what replaces the Mamba blocks and the fusion head in those variants, respectively.","section":"Table 2 and Fig. 4(b)"},{"comment":"In Eq. (4) the symbol FM is reused for the fused feature on the left-hand side after being used as the input on the right-hand side; a distinct symbol for the output of the averaging step would avoid the self-referential notation.","section":"Eq. (4)"},{"comment":"The caption of Fig. 4 should state what quantity is plotted against the temporal axis in panels (b) and (c), and the panels should include a baseline trace (at least the Nf=1 variant) so the claimed improvement in smoothness is visible rather than implicit.","section":"Fig. 4"},{"comment":"In the comparison-study paragraph, 'comparing the wrapped segmentation' should read 'warped segmentation'; the paper should also state the number of test subjects and slices behind the reported means and standard deviations.","section":"Sec. 3 (Comparison study)"}],"recommendation":"major_revision","confidential_remarks":"This is a transparent, well-organized empirical study with released code and honest reporting of its main weakness: the Discussion explicitly states that increasing Nf from 1 to 5 yields only about +0.5% Dice while giving 'visibly smoother' fields. My major-revision recommendation is driven by the gap between the abstract's central claims (temporally consistent, smooth tracking; no increase in computational complexity) and the quantitative evidence, and by the absence of significance testing for roughly one-percentage-point Dice differences over strong baselines. If the authors add a quantitative whole-cycle temporal-consistency evaluation (Major Comment 1) and statistical tests (Major Comment 2), I would view the paper as publishable at a specialist venue; without them, the incremental value of the sequence design over a single-pair Mamba baseline is small and the main claimed benefit remains unverified. I see no integrity concerns; the citation misattribution in reference [2] and the swapped labels in Table 3 appear to be proofreading issues rather than signs of anything else."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what to know: MCM is a sensible application of Mamba to cardiac motion tracking that feeds a short target-frame sequence into a bi-directional scanning Mamba encoder. The architecture is new to this problem and the authors compare against official implementations of VoxelMorph, TransMorph, and MambaMorph on two public datasets. That part is solid.\n\nThe real innovation is modest but genuine: treating the target as a small temporal neighborhood rather than a single frame, with a dual-path fusion head. The ablations show the Mamba blocks and the bi-directional scanning help, and the code is public.\n\nThe soft spots are about the central claim. The paper says sequence input gives smooth, temporally consistent motion across the cardiac cycle, but every quantitative result is for the ED-to-ES transition on three SAX slices—the only places with manual segmentations. Whole-cycle smoothness is supported only by qualitative sampling-line plots in Fig. 4, without baseline comparisons or error bars. The discussion itself concedes that increasing Nf from 1 to 5 adds only +0.5% Dice, so it is hard to rule out that the ED–ES gain comes from added capacity or input diversity rather than faithful temporal modeling. There are also no significance tests, and the improvements over TransMorph/MambaMorph are around one point in Dice, which could be noise. The computational-complexity claim is also a bit slippery: inference time stays flat, but VRAM goes from 3.2 to 12.4 GB as Nf grows.\n\nNone of this is fatal. The method is plausible and the evaluation is standard for the field. It's just that the headline advantage—temporal consistency—is not yet quantitatively demonstrated. The authors are transparent about the segmentation limitation, which I credit.\n\nWho is this for? Anyone tracking cardiac motion with 2D CMR who wants a strong baseline with code. It deserves a serious referee, but I'd ask for full-cycle or intermediate-frame validation (even indirect measures), statistical testing, and a more careful computational claim before accepting.\n\nRecommendation: send to peer review; expect revision.","headline":"Plausible Mamba-based cardiac motion tracker whose sequence-input advantage is asserted more strongly than the quantitative evidence supports.","tokens_in":10178,"tokens_out":2590,"would_cite":false,"duration_ms":27385,"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":"The paper claims that feeding a short sequence of neighboring cardiac MRI frames, instead of a single target frame, into a Mamba-based bi-directional state-space network produces smoother and more accurate deformation fields for ED-to-ES…","keywords":["cardiac motion tracking","Mamba","state space model","deformable image registration","cine MRI","temporal consistency","myocardial strain","sequential images"],"falsifier":"Warp the end-diastolic segmentation to every intermediate frame, not just end-systole, and compute Dice against manually traced or propagated segmentations on those frames; if a single-pair baseline such as TransMorph or MambaMorph matches or beats the sequence-based MCM on those intermediate frames, the core temporal-consistency advantage would be refuted.","tokens_in":9309,"feed_emoji":"🫀","tokens_out":3152,"duration_ms":34810,"temperature":0.7,"pith_summary":"Cardiac motion tracking in cine MRI is usually done by registering one reference frame to one target frame, which ignores the heart's continuous motion. This paper argues that including the target frame's neighbors in the input gives the network temporal context, leading to smoother and more consistent motion fields. The proposed MCM network processes a five-frame target sequence with bi-directional Mamba scanning and a dual-path fusion head, and on the ACDC and M&Ms datasets it achieves higher Dice overlap when warping end-diastolic segmentations to end-systole, plus lower percentages of negative Jacobian determinants and lower volume distortion than dDemons, VoxelMorph, TransMorph, and MambaMorph. A sympathetic reader would care because temporal consistency is exactly what strain analysis and cardiac function assessment need, and the method promises it without a large increase in inference time.","feed_headline":"Mamba network tracks heart motion from frame sequences","feed_subtitle":"Feeding Mamba a short sequence of MRI frames, not one pair, yields smoother cardiac deformation fields and beats four baselines on ACDC…","key_machinery":"The load-bearing components are (1) the bi-directional scanning Mamba (BiSM) block, which splits the deformation feature map into spatial positions, orders those positions along the temporal axis in both forward and backward directions, feeds each direction through its own state-space model (SSM), and sums the two outputs to capture temporal dynamics from both sides; and (2) the dual-path fusion head (DFH), which applies forward and backward 3D convolutions along the sequence dimension and averages the results before a 2D convolution outputs the motion field Φt. Together they let the network integrate neighboring frames without explicitly learning a recurrence over the whole cardiac cycle, and the paper attributes the gains in temporal consistency and smoothness to this design.","core_discovery":"The paper's central claim is that motion estimation for cine cardiac MRI improves when the network sees a sequence of target frames centered on the frame of interest, not just an isolated pair. The network takes the end-diastolic reference frame I0 together with St = {I_{t−K},...,I_t,...,I_{t+K}} (K=2, five frames total), pairs each target with I0, and processes the pairs through hierarchical bi-directional Mamba blocks. Each block scans the sequence both forward and backward in time, feeding the spatially ordered tokens into parallel state-space models and summing their outputs; a dual-path fusion head later combines forward and backward 3D convolutions on the fused feature map. On three short-axis slices of the ACDC dataset, the method reaches 84.6% mid-ventricular Dice versus 83.7% for TransMorph and 83.7% for MambaMorph, and it reports the lowest |J|<0% and ||J|−1| values on all three slices, which the paper interprets as smoother, more diffeomorphic, and more volume-preserving deformation fields. The same pattern holds on the M&Ms dataset. The paper also shows that removing the bi-directional blocks or the fusion head reduces performance, and that increasing the sequence length from one to five frames yields visibly smoother motion traces across the cardiac cycle.","pith_inferences":["A direct test the authors leave implicit is to evaluate motion fields on intermediate cardiac frames using propagated segmentations or tracked landmarks; if the temporal smoothing does not improve accuracy there, the central advantage over single-pair methods would be weakened.","Because the Mamba recurrence is linear in sequence length, pushing K beyond 2 (e.g., Nf = 7 or 9) might further smooth the motion traces, provided GPU memory, which the paper reports grows with Nf, can be managed.","The same sequence-pairing idea could be extended to 3D cardiac volumes or to across-slice temporal tracking in short-axis stacks, where the additional temporal context may reduce slice-to-slice inconsistencies.","One could probe the contribution of the dual-path fusion head separately with a diagnostic that compares forward-only, backward-only, and averaged fusion on intermediate frames, rather than only on the ED-to-ES endpoints."],"forward_implications":["If the central claim is correct, cardiac motion tracking systems can adopt sequence-aware architectures to obtain smoother deformation fields for the same registration loss, without changing the training objective or requiring paired intermediate-frame segmentations.","The bi-directional Mamba block and dual-path fusion head are modular; they could be inserted into other deformable registration pipelines that currently use single frame pairs.","Lower negative-Jacobian percentages and lower ||J|−1| suggest the estimated fields are closer to diffeomorphic and volume-preserving, which is directly relevant to strain and myocardial thickness measurements in clinical workflow.","The inference time stays around 17 ms even with five input frames, so the temporal-context benefit does not obviously trade away real-time usability.","The paper's loss function is a standard intensity-similarity plus smoothness regularization, so the sequence design can be combined with other losses such as segmentation-guided or biomechanics-informed terms."],"supporting_citations":[{"why":"dDemons, the conventional diffeomorphic registration baseline that the method must beat in Dice and diffeomorphism metrics.","marker":"[23]"},{"why":"VoxelMorph, the widely used unsupervised learning baseline for deformable registration that the method compares against.","marker":"[1]"},{"why":"TransMorph, the transformer-based registration baseline that provides the strongest Dice competition on ACDC mid-ventricle.","marker":"[7]"},{"why":"MambaMorph, the prior Mamba-based registration framework that the proposed bi-directional scanning directly extends.","marker":"[11]"},{"why":"ACDC, one of the two public cine CMR datasets used for training and evaluation.","marker":"[5]"},{"why":"M&Ms, the multi-centre, multi-vendor dataset used to demonstrate generalization beyond a single acquisition protocol.","marker":"[6]"},{"why":"Mamba, the selective state-space model whose linear-time recurrence underlies the bi-directional scanning mechanism.","marker":"[10]"}],"fun_headline_variants":["Mamba reads MRI sequence for smoother heart motion tracking","Bi-directional Mamba yields smoother cardiac motion from sequences","Sequence vs pair: Mamba improves cardiac MRI motion tracking","Smoother cardiac motion via Mamba on sequential MRI frames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative evidence only spans the end-diastolic to end-systolic transition on three short-axis slices, so the paper's broader claim of temporally consistent motion across the whole cardiac cycle relies on visual inspection of motion traces rather than measurements on intermediate frames.","fun_headline_variants_meta":{"raw":{"variants":["Mamba reads MRI sequence for smoother heart motion tracking","Bi-directional Mamba yields smoother cardiac motion from sequences","Sequence vs pair: Mamba improves cardiac MRI motion tracking","Smoother cardiac motion via Mamba on sequential MRI frames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3770,"prompt_tokens":1059,"completion_tokens":2711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":2644}},"tokens_in":675,"tokens_out":2711,"duration_ms":22738,"temperature":1.0,"reasoning_tokens":2644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:18:37.836117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Warp the end-diastolic segmentation to every intermediate frame, not just end-systole, and compute Dice against manually traced or propagated segmentations on those frames; if a single-pair baseline such as TransMorph or MambaMorph matches or beats the sequence-based MCM on those intermediate frames, the core temporal-consistency advantage would be refuted.","supporting_citations":[{"cited_title":"In: MICCAI (2007)","cited_arxiv_id":null,"evidence_quote":"dDemons, the conventional diffeomorphic registration baseline that the method must beat in Dice and diffeomorphism metrics."},{"cited_title":"IEEE Trans Med Imaging38(8), 1788–1800 (2019)","cited_arxiv_id":null,"evidence_quote":"VoxelMorph, the widely used unsupervised learning baseline for deformable registration that the method compares against."},{"cited_title":"Med Imag Anal82, 102615 (2022) 10 J","cited_arxiv_id":null,"evidence_quote":"TransMorph, the transformer-based registration baseline that provides the strongest Dice competition on ACDC mid-ventricle."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ACDC, one of the two public cine CMR datasets used for training and evaluation."},{"cited_title":"IEEE Trans Med Imaging 40(12), 3543–3554 (2021)","cited_arxiv_id":null,"evidence_quote":"M&Ms, the multi-centre, multi-vendor dataset used to demonstrate generalization beyond a single acquisition protocol."}],"review_version":1}