{"id":"6a2a791b-daa3-4458-8c9b-3cf91601a082","arxiv_id":"2411.18249","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An end-to-end deep learning framework jointly trains adaptive k-space sampling, reconstruction, and deformable registration for dynamic MRI, improving registered-image similarity to a reference at 4x to 8x acceleration.","lead":"This paper presents an end-to-end deep learning pipeline that jointly learns where to sample k-space, how to reconstruct undersampled dynamic MRI images, and how to register the moving images to a static reference. The authors report improved registered-image similarity on cardiac cine and aorta datasets at 4x to 8x acceleration, but the evaluation does not directly validate the estimated deformation fields.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The registration-quality metric (Eq. 18) is also the training loss (Eq. 16), so joint-training gains may reflect reconstruction collapse toward the reference rather than genuine motion accuracy.","rationale":"The reader's weakest_assumption identifies precisely the load-bearing vulnerability: Eq. 18 is both the training loss and the evaluation metric, so the reported 'registration quality' can be inflated without true motion estimation. Internal evidence strengthens this concern: the alpha=0 ablation shows the metric can be trivially maximized by collapsing reconstruction to the reference, yielding 0.987 SSIM in Table C1 while reconstruction SSIM drops to 0.479 in Table C2. The paper is honest about this limitation in Sec. 5, but honesty does not remove the circularity; it locates it. Other potential concerns—adaptive sampling not beating a static learned mask, the registration comparison giving the proposed network all temporal frames, code not yet released—are real but secondary. They affect the strength and fairness of peripheral comparisons, whereas the metric issue directly undermines the headline claim of improved motion estimation. Since the proposed concrete test is cheap and explicitly endorsed by the authors, the appropriate verdict remains conditional: accept only if the deformation fields are validated on ground-truth moving images. I therefore agree with the reader and recommend no change to the CONDITIONAL verdict.","tokens_in":36194,"tokens_out":3099,"duration_ms":29947,"concrete_test":"Warp the fully-sampled ground-truth moving images xmov with the predicted deformation fields phi (from the trained joint and decoupled models, and for adaptive versus fixed sampling) and evaluate SSIM, PSNR, and NMSE against xref, exactly as the paper suggests in Sec. 5. If joint training still outperforms decoupled training on this objective metric of motion accuracy, the central claim survives; if the advantage shrinks, reverses, or vanishes, the reported registration-quality gains are metric artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that end-to-end joint training 'significantly improves registration quality' (Sec. 4.7)—is supported only by the similarity between the warped reconstructed moving image and the reference, computed by Eq. 18. This is the same objective optimized during training through Lreg in Eq. 16. Because the reconstruction module is unconstrained, the pipeline can maximize Eq. 18 by making |xhat_mov| approximate xref, yielding near-perfect warped similarity with trivial deformation fields. The paper itself documents this failure mode in Sec. 4.7 and Figures C7/C8: with alpha=0, beta=1, registration SSIM rises to 0.987 while reconstruction SSIM collapses to 0.479 (Tables C1/C2). This confirms the metric is gameable. The paper's Sec. 5 concession that 'our results reflect both registration and reconstruction quality' and that evaluating registered ground-truth moving images 'would provide a more objective assessment' is therefore not a minor caveat but the crux: without independent validation of the deformation fields, the reported advantages of joint training, adaptive sampling, and the proposed registration network over baselines may be artifacts of optimizing the evaluation metric rather than improvements in motion estimation. No other concern is as load-bearing: the adaptive-versus-optimized sampling gap is a scope limitation, and the comparison bias favoring the proposed registration network affects fairness but not the validity of the core claim as directly as the metric circularity does.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an end-to-end deep learning framework for dynamic MRI that jointly optimizes adaptive k-space sampling, reconstruction, and deformable registration. The pipeline consists of a learned adaptive sampling mask generator, the vSHARP unrolled reconstruction network, and a 2D U-Net registration network, trained with a weighted combination of reconstruction and registration losses. Experiments on CMRxRecon 2023 cardiac cine and CMRxRecon 2024 aorta datasets compare registration modules, reconstruction backbones, sampling strategies, loss-weight configurations, and joint versus decoupled training. The evaluation metric is the similarity between the warped reconstructed moving image and the reference image (Eq. 18). The authors report that joint training and the proposed components improve registration quality and motion estimation, with the framework marketed as the first end-to-end integration of adaptive sampling, reconstruction, and registration.","tokens_in":36474,"tokens_out":5056,"duration_ms":46744,"significance":"If substantiated, the proposed pipeline would be a practically useful contribution to dynamic MRI, where undersampling and motion are central challenges. The modular design allowing interchangeable reconstruction and registration networks is attractive, and the paper provides a substantial set of ablations on both in-distribution and out-of-distribution data. The forward models and loss functions are clearly specified, and the authors are transparent about several limitations, including the reliance on fully sampled ground truth and the conflation of reconstruction and registration quality in the current evaluation. However, the central claim of improved motion estimation rests on an evaluation metric that is identical in form to the training loss, and the paper itself documents a failure mode in which the metric is maximized by reconstructing the reference image rather than by estimating correct deformations.","major_comments":[{"comment":"The registration-quality evaluation metric in Eq. 18 is the same similarity measure used as the registration training loss in Eq. 16 (SSIM+L1 between the warped reconstructed moving image and the reference). Consequently, the reported improvements in 'registration quality'—including the central claim in Sec. 4.7 that joint training significantly improves registration over decoupled training—do not necessarily reflect more accurate deformation fields. The pipeline can maximize the metric by driving the reconstructed moving image toward the reference, yielding near-perfect warped similarity with trivial motion. This is not hypothetical: the alpha=0, beta=1 experiment in Tables C1 and C2 and Figures C7/C8 shows registration SSIM rising to 0.987 while reconstruction SSIM collapses to 0.479, exactly the collapse mode described. The paper's own Sec. 5 statement that 'our results reflect both registration and reconstruction quality' and the suggestion to evaluate registered ground-truth moving images are acknowledgements that the main claim is currently untested. The revision should include an independent evaluation of the deformation fields—for example, warping the fully sampled ground-truth moving images with the predicted fields, or using synthetic/annotated data with known deformations—before the motion-estimation claims can be supported.","section":"Sec. 4.6 and Fig. 5"},{"comment":"The comparison of registration modules (proposed U-Net, VoxelMorph, TransMorph, optical flow, DEMONS) is affected by the same confound. Because Eq. 18 is computed on the registered reconstructed images, differences between registration modules can arise from different interactions with the jointly trained reconstruction module rather than from the modules' intrinsic motion-estimation capability. The conclusion that the proposed lightweight model yields 'more accurate motion estimation' than the baselines is therefore not supported by the presented evidence, even if the numbers are correct. The independent metric proposed above should be reported for all registration baselines (or at least for the learned ones) to establish the claimed advantage.","section":"Sec. 5"},{"comment":"The paper explicitly concedes that 'our results reflect both registration and reconstruction quality' and that 'evaluating motion estimation independently by comparing registered ground truth moving images... would provide a more objective assessment.' This limitation is load-bearing, not a minor caveat; it undermines the abstract's and conclusion's statements that the framework enables 'accurate deformation field estimation' and 'robust motion estimation from undersampled dynamic data.' Given the authors themselves identify the remedy, the revision should implement that evaluation rather than defer it to future work. Until then, the central claim is not directly validated.","section":"Sec. 4.7"}],"minor_comments":[{"comment":"The symbol m is used both for the averaged metric in Eq. 18 and for the mask operator M in Eq. 2; renaming one would avoid confusion.","section":"Sec. 4.4, Eq. 18"},{"comment":"The evaluation metric does not include the smoothness term Lsmooth, so the reported numbers do not capture whether the estimated deformations are physically plausible; reporting a regularity metric (e.g., percentage of folding or mean Jacobian determinant) would strengthen the motion-quality assessment.","section":"Sec. 3.3, Eq. 17"},{"comment":"Qualitative results show warped images but never visualize the estimated deformation fields; showing vector or grid overlays of the predicted motion would help the reader assess the behavior of the registration component.","section":"Appendix C, Figs. C1-C10"},{"comment":"The comparison with VoxelMorph and TransMorph does not include information about the number of trainable parameters or training cost, which would be useful for judging the 'lightweight' claim.","section":"Sec. 4.6"},{"comment":"The heading 'Registration uantitative results' contains a typo ('uantitative' should be 'quantitative').","section":"Appendix C, Table C5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for a medical imaging venue and is generally well written, with detailed ablations and transparent reporting. The main technical concern—the circularity between the registration loss and the evaluation metric—is correctly identified by the paper itself, but the authors' proposed remedy (evaluating deformation fields on ground-truth moving images) is essential to the central claim and should not be left as future work. The revision should implement that evaluation or clearly downgrade the paper's claims to 'joint reconstruction-and-registration quality' rather than 'motion estimation accuracy.' I would not recommend rejection because the framework is well-posed and the fix is feasible, but the current evidence does not support the strongest claims in the abstract and conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is the joint end-to-end training of adaptive k-space sampling, reconstruction, and registration for dynamic MRI. No one else seems to have put those three pieces in one differentiable loop, and the modular design is genuine engineering. The ablation work is extensive, and the paper is unusually candid about its own limitations.\n\nWhere the paper does well: experiments cover two datasets, multiple acceleration factors, and several registration and reconstruction backbones. The authors even show the failure mode where weighting registration heavily makes the reconstruction collapse toward the reference. That transparency matters.\n\nSoft spots: the main claim—that joint training significantly improves registration quality—is only measured by the similarity between the warped reconstructed moving image and the reference, which is the same objective used as the registration loss. The alpha=0, beta=1 case in Table C1 confirms the metric can be gamed: registration SSIM soars to 0.987 while reconstruction SSIM collapses to 0.479. That is not a minor caveat. The paper's own Sec. 5 admits that comparing registered ground-truth moving images would be more objective. Without that, the reported gains over VoxelMorph/TransMorph and the joint-versus-decoupled difference might be artifacts of optimizing the evaluation metric. I'd also note the adaptive sampling does not beat a static optimized mask, and the registration network gets all temporal frames at once while the baselines are fed frame-by-frame—an unfair comparison, though not a fatal one.\n\nOn balance: the integration itself is a real contribution, and the paper is honest, but the central claim needs an independent test—warp ground-truth moving images with the predicted deformation fields and evaluate against a reference, or use landmark/segmentation overlap—before the registration results can be trusted.\n\nThis deserves peer review, but with a major-revision request for the deformation-field validation. A serious referee can push on that without much effort.","headline":"Useful end-to-end dynamic MRI pipeline, but the registration claim rests on a metric that is also the training loss; needs an independent deformation-field evaluation before the gains can be believed.","tokens_in":37012,"tokens_out":1324,"would_cite":false,"duration_ms":13422,"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":"This paper claims that an end-to-end deep learning pipeline that jointly learns adaptive k-space sampling, image reconstruction, and deformable registration for dynamic MRI improves motion estimation over training each stage separately.","keywords":["dynamic MRI","adaptive k-space sampling","undersampled MRI reconstruction","deformable image registration","motion estimation","end-to-end deep learning","cardiac cine MRI","MR-guided radiotherapy"],"falsifier":"Apply the predicted deformation fields to the fully sampled ground-truth moving frames instead of the reconstructions and measure their similarity to the reference; if joint end-to-end training no longer beats decoupled training on that ground-truth metric, the paper's central claim that joint training improves motion estimation itself would be refuted.","tokens_in":35965,"feed_emoji":"🫀","tokens_out":7910,"duration_ms":72248,"temperature":0.7,"pith_summary":"The paper is trying to establish that motion estimation from undersampled dynamic MRI is best done by one differentiable system that learns where to sample k-space, how to reconstruct images from those samples, and how to warp the reconstructed frames to a reference image, all trained together. It introduces a pipeline whose adaptive sampling module chooses case-specific k-space lines, whose reconstruction module is an unrolled optimization network, and whose registration module predicts deformation fields aligning the moving frames to a static reference. Controlled experiments on cardiac cine data and an unseen aorta dataset show that joint end-to-end training improves registration quality over training the same modules separately, and that learned phase-specific sampling helps motion estimation more than fixed sampling schemes. If the claim holds, the practical consequence is a single trained network that can take accelerated dynamic acquisitions and output motion estimates directly, which would support cardiac function assessment, motion tracking, and MR-guided radiotherapy.","feed_headline":"Joint training sharpens motion estimates from undersampled MRI","feed_subtitle":"Learning sampling, reconstruction, and registration in one network beats training them separately on cardiac and aorta scans.","key_machinery":"The load-bearing object is the full differentiable chain from k-space mask to deformation field. An adaptive sampling network produces a binary sampling mask through a straight-through estimator, a training trick that lets discrete sampling decisions pass gradients; the undersampled data then pass through the vSHARP reconstruction network, an unrolled ADMM optimizer that alternates learned denoising with data-consistency steps; and finally a 2D U-Net registration model outputs displacement fields that warp each reconstructed frame toward the reference. A combined loss $L = \\alpha L_{\\text{rec}} + \\beta L_{\\text{reg}}$, with the registration term measuring similarity between the warped reconstruction and the reference plus a smoothness penalty on the fields, is backpropagated through warping, reconstruction, and sampling selection, which is what makes the earlier stages optimize for motion estimation rather than image quality alone.","core_discovery":"On the paper's own terms, the central discovery is that coupling the three stages changes what each stage learns. The adaptive sampler is not merely maximizing reconstruction fidelity; it selects k-space lines that help the later registration align the moving frames to the reference. The reconstruction network is pushed to preserve the structures the registration needs, rather than only improving pixel-level image quality, and the registration module in turn provides the training signal that shapes both earlier stages. The paper reports that this joint optimization significantly improves registration quality compared with decoupled training, that phase-specific learned sampling beats unified and fixed patterns, that the unrolled vSHARP reconstruction supports registration better than the alternative VarNet, and that equal weighting of the reconstruction and registration losses gives the best balance between faithful images and accurate motion fields.","pith_inferences":["The reported registration scores are computed on warped reconstructed images, so part of the apparent registration gain may be reconstruction gain; the authors acknowledge that registering fully sampled ground-truth moving images would be a more objective test.","If the modularity claim is right, the registration block could be swapped for a segmentation-supervised or transformer-based model without retraining the sampler, an experiment the paper does not run.","Because the adaptive sampler is trained with fully sampled k-space ground truth, prospective deployment on scans where full sampling is impossible would require a self-supervised reformulation of the sampling loss, a step the authors flag as future work."],"forward_implications":["A single trained network can go directly from undersampled dynamic k-space to deformation fields, replacing separate acquisition-planning, reconstruction, and registration steps at inference.","Joint training makes the reconstruction subnetwork behave differently from a reconstruction-only network: it is shaped by what the registration needs, which is why decoupled training leaves registration quality on the table.","The choice of reconstruction backbone matters for motion estimation; the unrolled vSHARP design supports registration better than the tested VarNet configuration.","Phase-specific learned sampling patterns improve motion estimates over unified or fixed equispaced patterns, though equispaced initialization does not add a registration benefit.","The loss weights must be balanced: over-weighting registration can drive the reconstructor to imitate the reference image rather than the actual moving frame, inflating registration scores while corrupting the reconstruction."],"supporting_citations":[{"why":"Supplies the adaptive dynamic sampling approach that the proposed sampling module is based on, and acts as the main methodological predecessor.","marker":"[60]"},{"why":"Provides the vSHARP unrolled reconstruction model used as the default reconstruction module in the pipeline.","marker":"[57]"},{"why":"Provides the sensitivity-estimation and refinement method, and the VarNet architecture used as the alternative reconstruction baseline.","marker":"[39]"},{"why":"Provides the U-Net architecture used both for the registration model and for sensitivity-profile refinement.","marker":"[35]"},{"why":"Provides the VoxelMorph registration baseline and the differentiable warping operation used to compute registered images.","marker":"[5]"},{"why":"Provides the TransMorph transformer-based registration baseline that the proposed registration module is compared against.","marker":"[9]"},{"why":"Supplies the dataset-optimized learned sampling scheme used as a non-adaptive learned baseline in the sampling experiments.","marker":"[64]"},{"why":"Supplies the fixed non-adaptive sampling schemes, including equispaced patterns, used as baselines against learned sampling.","marker":"[58]"},{"why":"Provides the straight-through estimator used to binarize the adaptive sampling mask so gradients can flow end-to-end.","marker":"[7]"}],"fun_headline_variants":["Joint MRI pipeline beats separate training for motion tracking","End-to-end MRI network learns to sample, rebuild, and register","Coupled k-space sampling and reconstruction improve MRI motion estimation","Joint training of sampling, reconstruction, and registration sharpens MRI motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the similarity between the warped reconstructed moving image and the reference image is a faithful measure of deformation-field accuracy, even though that score also improves whenever the reconstructed image itself resembles the reference.","fun_headline_variants_meta":{"raw":{"variants":["Joint MRI pipeline beats separate training for motion tracking","End-to-end MRI network learns to sample, rebuild, and register","Coupled k-space sampling and reconstruction improve MRI motion estimation","Joint training of sampling, reconstruction, and registration sharpens MRI motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001258,"raw_usage":{"total_tokens":5167,"prompt_tokens":973,"completion_tokens":4194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":4125}},"tokens_in":589,"tokens_out":4194,"duration_ms":28055,"temperature":1.0,"reasoning_tokens":4125,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:22:47.328706+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the predicted deformation fields to the fully sampled ground-truth moving frames instead of the reconstructions and measure their similarity to the reference; if joint end-to-end training no longer beats decoupled training on that ground-truth metric, the paper's central claim that joint training improves motion estimation itself would be refuted.","supporting_citations":[{"cited_title":"Magnetic Resonance Imaging 115, 110266 (2025)","cited_arxiv_id":null,"evidence_quote":"Provides the vSHARP unrolled reconstruction model used as the default reconstruction module in the pipeline."},{"cited_title":"IEEE transactions on medical imaging 38(8), 1788–1800 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the VoxelMorph registration baseline and the differentiable warping operation used to compute registered images."},{"cited_title":"Medical image analysis 82, 102615 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the TransMorph transformer-based registration baseline that the proposed registration module is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dataset-optimized learned sampling scheme used as a non-adaptive learned baseline in the sampling experiments."},{"cited_title":"Magnetic Resonance Imaging 107, 33–46 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed non-adaptive sampling schemes, including equispaced patterns, used as baselines against learned sampling."}],"review_version":1}