{"id":"db996de8-9492-4ae6-8042-adb2c0fadcab","arxiv_id":"1908.09560","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A learned motion-prediction model, trained on an initial k-space center acquisition, accelerates motion-aware 4D MRI by about 2x in scan time and 40x in reconstruction time while keeping image quality comparable.","lead":"MRI scans of the chest and abdomen are easily blurred by breathing motion, and existing motion-aware methods are too slow for routine use. This paper learns a patient's breathing pattern from a short training phase and then predicts motion from a tiny k-space patch, cutting scan time roughly in half and reconstruction time from hours to minutes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation protocol trains motion model on both ends of the time series, so reported accuracy is interpolation; the deployed accelerated method must extrapolate from the first 100 time points. The key generalization and quality claims are therefore not causally tested.","rationale":"The strongest claim is that motion can be predicted from a short initial training phase, enabling near-halved acquisition time and two-orders-of-magnitude faster reconstruction at equivalent or higher quality. The acquisition and reconstruction time reductions are concrete and credible: 11.1 vs 5.8 minutes and 2 hours vs 3 minutes, supported by sequence parameters in Table 1. The soft spot is the motion-prediction validation. Training on both leading and last time points measures interpolation over known motion states, not extrapolation to unseen drift or amplitude changes. Since the deployed accelerated protocol is purely causal, this mismatch is precisely where the claim could fail. The paper's own limitation sentence confirms the mechanism. The qualitative image comparison in Figures 4-5 is encouraging but is explicitly not a matched comparison, and TV was only used to validate shift correction. So the conditional verdict is appropriate; the missing piece is a causal retest of prediction accuracy and, ideally, a matched quantitative reconstruction comparison. I agree with the reader's weakest-assumption identification.","tokens_in":6201,"tokens_out":3069,"duration_ms":32997,"concrete_test":"Re-run the motion-prediction experiment on the 12 standard acquisitions with a strictly causal split: fit the cubic/PCA model on time points 1-100 only, predict time points 101-1500, and report mean and 95th-percentile displacement error as a function of time. If prediction error after the training phase exceeds the reported <2 mm mean / <4 mm 95th percentile, or grows with respiratory drift, the generalization assumption underlying the accelerated sequence is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4's motion-prediction experiment does not test the accelerated protocol as deployed. For the 12 standard acquisitions, the training phase was \"split into the leading and last time points to account for organ drift,\" so the reported sub-2 mm mean error is an interpolation result: the model has seen data from the end of the series. The actual accelerated sequence trains only on the first 100 time points and predicts 1400 later points (Table 1). The paper's own conclusion concedes that \"changes in amplitude of the motion after the training phase may compromise the motion prediction though.\" The reconstruction-quality claim is similarly under-tested: Section 4 states that \"we cannot make a direct comparison between the shift-corrected and accelerated reconstruction method because they are applied to different acquisitions,\" and the only quantitative quality metric (TV) is used for the shift-correction component, not for the accelerated-versus-standard comparison. Thus the central claim of equivalent or higher reconstruction quality under a nearly 2x faster acquisition is not backed by a valid causal evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an accelerated version of a motion-aware 4D MR imaging method. In the standard method, the k-space center is sampled repeatedly to estimate non-rigid motion fields that correct peripheral k-space patches, but this requires long acquisition and reconstruction times. The proposed method adds an initial training phase in which larger center patches are acquired, learns a cubic regression model from PCA scores of these patches to motion fields, and during the inference phase predicts motion from much smaller center patches. This reduces the acquisition time from 11.1 min to 5.8 min and the reconstruction time from 2 h to 3 min. The authors also introduce a systematic temporal shift correction based on quadratic interpolation of the motion fields. Experiments on 12 volunteers (6 with the accelerated sequence) report average motion-prediction errors below 2 mm and qualitative reconstruction results that are claimed to be equivalent or better than the standard approach.","tokens_in":6382,"tokens_out":3384,"duration_ms":38104,"significance":"If the claimed gains are validated, the contribution is practically important: a roughly two-fold reduction in acquisition time and two-orders-of-magnitude reduction in reconstruction time would make motion-aware 4D abdominal/thoracic MRI substantially more clinically usable. The acquisition-time and reconstruction-time reductions are precisely stated and consistent with the sequence parameters in Table 1. The shift-correction idea is clearly presented and its quantitative TV-based evaluation is a useful first step. However, the two load-bearing claims, that motion can be predicted forward in time after the training phase and that reconstruction quality is equivalent or higher than the standard method, are not established by the reported experiments. The paper does not provide code or a fully reproducible pipeline, but the experimental setup is described in sufficient detail that the missing validation could in principle be added.","major_comments":[{"comment":"The motion-prediction validation does not test the deployed accelerated protocol. The text states that 'we split the training phase into the leading and last time points to account for organ drift,' so the reported sub-2 mm mean error in Figure 3 is an interpolation result: the model has seen data from both ends of the time series. The actual accelerated sequence trains only on the first 100 time points and predicts the remaining 1400 time points (Table 1), which is forward extrapolation. The paper's own conclusion concedes that 'changes in amplitude of the motion after the training phase may compromise the motion prediction though,' and Volunteer 9 is reported to show exactly such a change. As it stands, the motion-prediction claim is therefore not causally validated for the proposed acquisition workflow.","section":"§4, 'Motion Prediction'"},{"comment":"The abstract's claim of 'equivalent to higher reconstruction quality' is not supported by a quantitative or controlled comparison. The text explicitly says 'we cannot make a direct comparison between the shift-corrected and accelerated reconstruction method because they are applied to different acquisitions,' and Figures 4 and 5 are selected qualitative slices. The only quantitative image-quality metric, total variation, is used exclusively in the shift-correction experiment, not in the accelerated-versus-standard comparison. A valid evaluation would need, for example, a forward-split simulation on the standard acquisitions, a quantitative sharpness/artifact metric on the accelerated acquisitions, or a blinded reader study; none is provided.","section":"§4, 'Accelerated Motion-Aware MR Imaging'"},{"comment":"The regression model in Eq. (2) is trained and evaluated on the same acquisition, so the description as 'motion prediction' overstates what is demonstrated. The training phase provides both the PCA basis and the least-squares weights Ψ for a single subject's scan, and the reported error is computed after training on both temporal ends of that same acquisition. The manuscript should clearly state that the model is a per-acquisition fit and that no cross-subject or cross-session generalization is claimed; the current wording implies a more general predictive model than the experiments establish.","section":"§3.1, Eq. (2)"}],"minor_comments":[{"comment":"The relation between the training-phase duration (76 s), inference-phase duration (4.5 min), and total accelerated acquisition time (5.8 min) is consistent, but a sentence spelling out that the accelerated time includes the training phase would help readers avoid misreading the 5.8 min as purely inference time.","section":"Table 1"},{"comment":"The TV-based evaluation is reported as a statistically significant increase with a large effect size, but TV is a proxy for sharpness and can increase with noise or artifacts; a brief discussion of this limitation, or an additional metric, would strengthen the claim that the shift correction improves image quality.","section":"§4, 'Systematic Shift Correction'"},{"comment":"There is a typo in the first sentence: 'anaccelerated' should be 'an accelerated'. Also, the discussion of amplitude changes affecting prediction would benefit from a reference to the quantitative error observed for Volunteer 9 in Figure 3.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The core acceleration numbers appear credible and the proposed method is well motivated, but the validation protocol does not test the method as deployed. The motion-prediction experiment trains on both ends of the time series and therefore measures interpolation, while the deployment requires forward extrapolation from the first 100 time points; the quality comparison is explicitly acknowledged as non-comparable. These are fixable with additional experiments on the already-acquired data, so I recommend major revision rather than rejection. I did not find issues with citation or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new here is a practical acceleration of the authors' earlier motion-aware 4D MRI framework. Instead of continuously sampling large k-space center patches, the sequence learns a motion model from an initial 100-time-point training phase and then predicts motion from tiny 9-point patches, cutting acquisition time from 11.1 to 5.8 minutes and reconstruction from 2 hours to 3 minutes. That is a real engineering gain, and the timing numbers are clearly reported with sequence parameters. The systematic shift correction via quadratic interpolation is also a sensible, well-motivated fix, and the TV-based sharpness improvement (p = 0.002, d = 0.86) supports that component.\n\nThe soft spots are exactly where the reader and stress-test point. The motion-prediction experiment trains the regression on both the leading and last time points of the standard acquisitions, so the reported sub-2 mm mean error is an interpolation result, not a test of the deployed protocol, which must extrapolate from the first 100 time points to 1400 later ones. The paper even concedes that amplitude changes after training may compromise prediction. That concession is honest, but it does not fix the evaluation. The reconstruction-quality claim is equally under-tested: the paper states it cannot make a direct quantitative comparison between shift-corrected and accelerated reconstructions because they use different acquisitions, and the TV analysis applies to shift correction, not to accelerated versus standard. So the headline claim of \"equivalent or higher reconstruction quality\" rests on qualitative inspection of a few coronal slices.\n\nHaving said that, the timing improvements are solid and reproducible from the sequence parameters, and the acceleration idea is worth pursuing. The flaws are in the validation, not in the core method. A clean causal split—train only on the first 100 time points, then evaluate on the rest—could substantially improve confidence. A quantitative image-quality metric on matched acquisitions, even a simple one like structural similarity on the overlapping region, would strengthen the quality claim. As it stands, this is a useful engineering contribution with a validation gap.\n\nThis paper deserves a serious referee. It is not a desk reject; it is a solid MICCAI-style submission that needs a revised validation before the central claims can be accepted. I would send it to peer review, with the expectation of major revisions.","headline":"Practical acceleration of motion-aware 4D MRI with real time savings, but the key prediction-generalization and quality claims are not causally validated.","tokens_in":6950,"tokens_out":1268,"would_cite":false,"duration_ms":15080,"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-aware 4D MRI method learns breathing motion from a short initial k-space-center training phase and then predicts motion from tiny center patches, cutting acquisition time almost in half and reconstruction from 2 hours to 3…","keywords":["motion-aware MRI","k-space center","motion prediction","4D MRI","non-rigid motion correction","free-breathing MRI","cubic regression","temporal shift correction"],"falsifier":"Scan a volunteer under free breathing with the accelerated sequence, then have them change breathing depth or cough after the 100-time-point training phase, and compute the motion prediction error against motion fields recovered from full center patches. If the 95th-percentile error consistently exceeds the roughly 4 mm observed in the paper, or if the reconstructed images show visible blurring at the diaphragm, the assumption that trained motion generalizes to later time points is falsified.","tokens_in":5990,"feed_emoji":"🧲","tokens_out":5833,"duration_ms":51243,"temperature":0.7,"pith_summary":"Motion during MRI is a major source of image artifacts, especially for free-breathing scans of the chest and abdomen. This paper claims that a motion-aware 4D MRI scan can be accelerated by learning the patient's characteristic motion during a short initial training phase and then predicting the motion fields from tiny k-space center patches instead of measuring the full center every time. The result is an acquisition time reduced from 11.1 minutes to 5.8 minutes and a reconstruction time reduced from 2 hours to 3 minutes, with reconstruction quality equivalent to or better than the slower approach. The authors tested the method on 12 volunteers scanning lungs and abdomen during free breathing, with average motion prediction error below 2 mm. If it holds, the work would bring motion-corrected 4D MRI closer to clinical use.","feed_headline":"Motion prediction halves scan time and cuts reconstruction 40-fold","feed_subtitle":"A brief training phase learns lung motion from k-space center patches, then predicts it for the rest of the scan.","key_machinery":"The central mechanism is a learned motion-prediction model built from k-space center patches. In a training phase, full-size center patches are converted to motion fields by non-rigid image registration; these motion fields are then regressed, using a cubic polynomial in the PCA scores of tiny center patches, onto the tiny patches that will be sampled during the accelerated scan. The predicted motion fields are applied to correct peripheral k-space patches before accumulation, and a quadratic-interpolation shift correction with shift $\\Delta = 0.5$ compensates for the systematic temporal offset between center and peripheral patch sampling. This model is what lets the sequence replace most full-center measurements with tiny patches, producing the time savings.","core_discovery":"On its own terms, the paper establishes that the motion fields needed for motion-compensated 4D MRI can be inferred from a small subset of k-space center data. During an initial training phase of 100 time points, about 76 seconds, full center patches are acquired and registered to recover motion fields; a cubic regression model is then fit between PCA scores of tiny center patches and these motion fields. In the inference phase, only tiny center patches are sampled and the model predicts the non-rigid motion for the remaining 1400 time points, allowing the peripheral k-space data to be spatially corrected. The paper further introduces a systematic temporal shift correction, shifting motion fields by half a time step using quadratic interpolation to account for the delay between center and peripheral patch acquisition. Experiments on 12 free-breathing volunteers show prediction errors below 2 mm on average and qualitatively equivalent or sharper reconstructions than the standard 11-minute approach.","pith_inferences":["The same train-then-predict structure could be extended to cardiac or combined respiratory-cardiac motion, provided a reliable periodic trigger or a longer training phase captures the faster dynamics; the paper only demonstrates respiratory motion.","The model's reliance on a fixed training phase suggests an online adaptation scheme, periodically re-estimating the regression weights from recently acquired center patches, could make the method robust to drift and amplitude changes without lengthening the scan.","A direct testable extension would be to train on one volunteer and predict motion for a different scan session or for instructed deep-breathing phases; the paper's training setup uses the same session, so cross-session generalization is not established."],"forward_implications":["Acquisition time for the motion-aware 4D sequence drops from 11.1 to 5.8 minutes, making the protocol more feasible for routine clinical scans of thorax and abdomen.","Reconstruction time drops from about 2 hours to 3 minutes because full 3D image registration is only needed during the 100-time-point training phase.","The shift correction increases image sharpness: average total variation rises significantly ($p=0.002$, Cohen's $d=0.86$) across the 12 standard and 6 accelerated acquisitions.","Motion prediction accuracy stays below 2 mm on average, so the method can track respiratory motion with sub-voxel precision in free-breathing volunteers.","Because the accelerated method yields equivalent or higher reconstruction quality in less time, motion-aware 4D MRI becomes a plausible option for time-resolved imaging without gating or binning."],"supporting_citations":[{"why":"Supplies the base motion-aware acquisition and reconstruction that this paper accelerates, and the comparison standard for reconstruction quality.","marker":"[7]"},{"why":"Provides the non-rigid image registration used to derive training motion fields from k-space center patches and to compute ground-truth motion.","marker":"[10]"},{"why":"Motivates the problem by showing that motion artifacts are a leading cause of repeated clinical abdominal MRI acquisitions.","marker":"[12]"}],"fun_headline_variants":["Motion prediction from k-space center: 2x faster scan, 100x faster reconstruction","Predict motion from k-space center patches: halve scan time, 100x faster reconstruction","Train on k-space center, predict motion: scan in half, reconstruct 100x quicker","K-space center motion prediction speeds up MRI 2x and reconstruction 100x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the motion observed during the initial training phase is representative of the motion that will occur throughout the rest of the scan; if breathing amplitude, phase, or organ drift changes afterwards, the predicted motion fields will be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Motion prediction from k-space center: 2x faster scan, 100x faster reconstruction","Predict motion from k-space center patches: halve scan time, 100x faster reconstruction","Train on k-space center, predict motion: scan in half, reconstruct 100x quicker","K-space center motion prediction speeds up MRI 2x and reconstruction 100x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2917,"prompt_tokens":939,"completion_tokens":1978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1885}},"tokens_in":555,"tokens_out":1978,"duration_ms":16155,"temperature":1.0,"reasoning_tokens":1885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:08:07.065952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan a volunteer under free breathing with the accelerated sequence, then have them change breathing depth or cough after the 100-time-point training phase, and compute the motion prediction error against motion fields recovered from full center patches. If the 95th-percentile error consistently exceeds the roughly 4 mm observed in the paper, or if the reconstructed images show visible blurring at the diaphragm, the assumption that trained motion generalizes to later time points is falsified.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the base motion-aware acquisition and reconstruction that this paper accelerates, and the comparison standard for reconstruction quality."},{"cited_title":"AirLab: Autograd Image Registration Laboratory","cited_arxiv_id":"1806.09907","evidence_quote":"Provides the non-rigid image registration used to derive training motion fields from k-space center patches and to compute ground-truth motion."},{"cited_title":"Abdominal Radiology 42(1), 306–311 (2017)","cited_arxiv_id":null,"evidence_quote":"Motivates the problem by showing that motion artifacts are a leading cause of repeated clinical abdominal MRI acquisitions."}],"review_version":1}