{"id":"8ccce631-b7a0-43ac-9640-cd634b85765d","arxiv_id":"2507.23224","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"EMORe uses expectation-maximization to refine soft motion-bin assignments and reject outlier readouts, improving 5D cardiac MRI sharpness and artifact suppression over compressed sensing.","lead":"EMORe is a new MRI reconstruction algorithm that corrects imperfect motion binning and discards motion-corrupted data during 5D cardiac MRI, using an expectation-maximization loop. It improved image sharpness and reduced artifacts over standard compressed sensing in simulated and volunteer scans, especially when volunteers coughed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EMORe's claimed robustness depends on a strong, phantom-tuned binning prior (αg=0.85) that requires a ~670:1 likelihood ratio to reassign a readout; no sensitivity analysis explores initial self-gating error or αg, leaving the central generalizability claim unsecured.","rationale":"The reader's weakest assumption and my independent read converge: the least secure part of EMORe's argument is the interaction between the informative prior θ and the initial self-gating binning. The quantitative 670:1 likelihood threshold, derived from Eq. (4) and Table I, makes the concern concrete rather than hypothetical. The phantom study is otherwise well designed, and the in vivo results, while modest, are directionally consistent with the central claim. I also note two secondary issues that do not change the verdict: (1) Algorithm 1 line 7 writes the stopping condition with a logical AND, while the text says 'whichever is achieved first'; as written, η is ineffective and all runs continue to J, so the reported runtimes do not correspond to the stated criterion. (2) The Brier score in Section II-B is computed only over the K valid bins, excluding the outlier bin; for readouts rejected as outliers this can make the reported 'bin assignment accuracy' improvement partly an artifact of soft or thresholded weights rather than correct inter-bin reassignment. Both are fixable in revision. Keeping the reader's CONDITIONAL verdict is appropriate, with the added explicit condition that the authors supply the αg and initial-label-corruption sensitivity analysis described above.","tokens_in":13309,"tokens_out":8174,"duration_ms":102806,"concrete_test":"Using the released MRXCAT pipeline, add a controlled label-corruption layer before computing θ in Eq. (4): randomly reassign 5%, 10%, 15%, and 20% of readouts to a neighboring cardiac or respiratory bin, independently of the existing bulk-motion outliers, and sweep αg over {0.5, 0.65, 0.8, 0.95} at each corruption level. If EMORe's PSNR/SSIM advantage over CS persists at ≥15% label corruption and across the αg sweep, the initialization concern is not load-bearing; if the advantage collapses or reverses, the method's robustness claim must be conditioned on self-gating accuracy or on per-site tuning of αg.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that EMORe's inter-bin correction is governed by the hand-set prior θ in Eq. (4), and the paper provides no evidence that this prior transfers to the range of self-gating inaccuracies encountered in practice. With αg=0.85, αo=0.05, and K=80, every non-SG prior is (1−0.85−0.05)/79 ≈ 0.00127; for a readout to move from its initial SG bin to an alternative bin, the likelihood ratio must exceed roughly 0.85/0.00127 ≈ 670. Thus EMORe can confidently reverse only relatively gross misassignments, while subtle but systematic SG errors—the regime that motivates inter-bin correction—may be locked in by the prior. The authors acknowledge EM's sensitivity to initialization in Section II-A.3 and tune αg on a single phantom dataset, but they never vary αg or degrade the initial SG labels. Without such a sensitivity analysis, the paper's central claim that EMORe is a generally robust, drop-in correction for imperfect self-gating is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EMORe, an expectation-maximization framework for 5D cardiac MRI reconstruction that iteratively refines probabilistic bin assignments and rejects motion-corrupted readouts into a dedicated outlier bin. The E-step computes posterior participation weights from the current image estimate and a bin-assignment prior, and the M-step re-estimates images via sparsity-regularized least squares. The authors validate EMORe against standard compressed sensing in a simulated MRXCAT phantom with controlled bulk motion and in 13 in vivo free-breathing scans, reporting improved PSNR, SSIM, edge sharpness, Brier score, and blinded reader scores.","tokens_in":13551,"tokens_out":4305,"duration_ms":49229,"significance":"If the claims hold, EMORe is a practically valuable drop-in refinement for free-running, free-breathing self-gated 5D cardiac MRI, and the public availability of source code and sample data strengthens reproducibility. The EM derivation is internally consistent, the phantom study provides a quantitative benchmark with ground-truth bin assignments, and the in vivo evaluation includes blinded expert scoring. The main risks are that several load-bearing implementation details are underspecified, in particular the noise-variance estimate used in the E-step, the sensitivity of results to the hand-set binning prior, and a contradiction between the stopping-criterion text and pseudocode. These are fixable with additional analysis and clarification rather than being fundamental flaws.","major_comments":[{"comment":"The E-step depends critically on the noise standard deviation σ, but the manuscript never states how σ is estimated. Since Table I sets τ = 3σ, the outlier-rejection boundary is determined by σ, and the posterior weights scale as exp(-1/(Lσ²) ||A x - y||²). For in vivo data, especially after coil compression, the noise level is not reported or characterized. The authors should specify a concrete noise-estimation procedure (e.g., from background regions or k-space corners) and, ideally, report sensitivity of the results to σ, because without this the in vivo E-step is not fully reproducible.","section":"II-A.1, Eq. (1b)-(1c), Table I"},{"comment":"With αg = 0.85, αo = 0.05, and K = 80, every non-self-gating prior probability equals (1 - 0.85 - 0.05)/79 ≈ 0.00127, so a readout requires a likelihood ratio exceeding roughly 0.85/0.00127 ≈ 670 to be reassigned from its initial self-gating bin. The authors acknowledge in Section II-A.3 that EM is sensitive to initialization, and all hyperparameters were tuned on a single phantom dataset, yet no experiment varies αg or degrades the initial self-gating labels. This leaves the central generalizability claim—that EMORe robustly corrects realistic self-gating inaccuracies—unsecured. A sensitivity analysis over αg and over initial binning accuracy is needed.","section":"II-A.3, Table I"},{"comment":"The pseudocode stops when the normalized squared image difference is less than η AND t ≥ J, but the text states the stopping criterion is either the maximum number of iterations J or the threshold η, whichever is achieved first. With the AND condition as written, the loop would not terminate at J if the normalized difference has not fallen below η, contradicting the text and potentially affecting runtime and convergence behavior. The condition should be an OR (or the text should be revised to match the pseudocode).","section":"Algorithm 1, line 7; Section II-A.3"},{"comment":"The Brier score is defined as a sum over the K valid bins, while the algorithm assigns posterior mass to K+1 bins including the outlier bin. The manuscript does not specify how corrupted readouts (whose true class is the outlier bin) are represented in the true participation weights w̃, or whether they are excluded from the Brier computation. Without this clarification, the claimed bin-assignment accuracy improvement is ambiguous and may not fully reflect the method's outlier-rejection behavior.","section":"II-B, Brier score definition"}],"minor_comments":[{"comment":"There is a typo in the Conclusion: 'implmented' should be 'implemented'.","section":"V. Conclusion"},{"comment":"The text states that blind scoring was performed on 26 cine pairs, while Table II reports per-reviewer means. Please clarify whether the paired t-test was performed across the 26 pairs or across the 13 volunteers, and specify how the per-reviewer values in Table II were aggregated.","section":"III-B, Table II"},{"comment":"The claim of robustness to coughing-induced motion is based on only three volunteers. The authors should state this small-sample limitation explicitly when discussing the in vivo coughing results, rather than presenting the qualitative evidence from Fig. 6 as the primary support.","section":"IV. Discussion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for IEEE TMI and the phantom study is a solid quantitative benchmark. The main concerns are the underspecified noise-variance estimation, the lack of sensitivity analysis for the binning prior, and the pseudocode/text contradiction; these are fixable with additional experiments and clarifications. I would not reject the paper, but I would require the sensitivity analysis and noise-estimation description before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EMORe is a solid engineering contribution. The EM-guided soft binning with a dedicated outlier bin is genuinely new for self-gated 5D cardiac MRI, and the authors validate it the right way: a phantom study with ground truth and controlled outlier levels up to 70%, plus an in vivo study with instructed coughing. The code and data are public. That is the kind of reproducibility I like to see.\n\nThe phantom results are convincing: EMORe improves PSNR, SSIM, edge sharpness, and Brier score over standard CS, and the improvement grows with outlier fraction. The in vivo edge sharpness and blind scores are also statistically significant, though the effect is modest (edge sharpness 0.722 vs 0.694). The method is a drop-in upgrade rather than a paradigm shift, but that is a fair contribution.\n\nThe soft spots are real but addressable. The biggest is the informative prior in Eq. (4). With αg=0.85, αo=0.05, and K=80, a readout needs roughly a 670:1 likelihood ratio to be moved out of its initial SG bin. So EMORe will correct gross misassignments but may be almost locked to the initial self-gating for subtle, systematic errors. The authors tune αg on one phantom and never explore its sensitivity or degrade the initial SG labels. Given the paper's claim of robustness to imperfect SG, this missing analysis is the main thing standing between this and a stronger result. The stress-test note's arithmetic checks out.\n\nSmaller issues: Algorithm 1's stopping criterion uses AND where the text says 'either...or' (OR); that needs fixing. The in vivo σ is never stated as estimated. And the phantom study reports many paired t-tests without a multiple-comparison correction; the in vivo section uses Bonferroni, so the authors know better.\n\nOverall, the central mechanism is sound and the validation is honest. The authors acknowledge EM's sensitivity to initialization and note degradation at >40% outliers. This is a subfield paper, not a breakthrough, but it is a useful one. I would send it to peer review and ask for a sensitivity analysis on αg and initial binning accuracy, plus the clarifications above. It will be a better paper after that.","headline":"EMORe is a well-validated EM-based binning correction for 5D cardiac MRI that deserves peer review, but its robustness claim needs a sensitivity analysis on the binning prior.","tokens_in":14101,"tokens_out":2795,"would_cite":true,"duration_ms":31317,"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":"EMORe: a reconstruction method that corrects motion-bin assignments and rejects outlier readouts in self-gated 5D cardiac MRI.","keywords":["5D cardiac MRI","self-gating","expectation-maximization","binning correction","outlier rejection","compressed sensing","motion artifacts","free-breathing MRI"],"falsifier":"Run EMORe on a phantom dataset in which a controlled fraction of readouts (e.g., 20–40%) is deliberately assigned to the wrong motion bin by corrupting the self-gating signal, and compare PSNR and Brier score against compressed sensing; the central claim would be refuted if EMORe's advantage disappears or if its final bin assignments remain as wrong as the initialization. A minimal version is to lower the prior weight $\\alpha_g$ from $0.85$ to $0.5$ and show that EMORe's output quality collapses once the true bins are no longer protected.","tokens_in":1920,"feed_emoji":"🪀","tokens_out":2232,"duration_ms":109533,"temperature":0.7,"pith_summary":"This paper proposes EMORe, a reconstruction method for free-running, free-breathing self-gated 5D cardiac MRI that treats imperfect retrospective motion binning as a correctable error rather than a fixed input. The authors aim to show that by iteratively re-estimating each k-space readout's probability of belonging to each of 80 cardiorespiratory motion bins, plus an explicit outlier bin for motion-corrupted data, while simultaneously improving the image estimate, residual blur and bulk-motion artifacts can be substantially reduced. If true, this matters because 5D cardiac MRI currently relies on self-gating binning that is corrupted by inaccurate signal extraction, irregular breathing, and sporadic patient motion such as coughing, and these errors limit diagnostic confidence. The paper validates the claim in simulated MRXCAT phantom studies and in 13 in vivo scans, reporting statistically significant gains over compressed sensing in image sharpness, structural similarity, and blinded artifact scores. The method is a drop-in reconstruction upgrade at a modest increase in computation time.","feed_headline":"EMORe sharpens cardiac MRI by fixing misassigned motion data","feed_subtitle":"In 50 phantom and 13 volunteer scans, EMORe beat compressed sensing on sharpness and artifact scores.","key_machinery":"The central mechanism is an expectation-maximization loop in which the true motion-bin assignment of every readout is a latent variable. In the E-step, the posterior probability that readout $n$ belongs to bin $k$ is computed by Bayes' rule from a Gaussian k-space residual likelihood and an informative self-gating prior $\\theta(n,k)$; the $(K+1)$-th 'bin' is an outlier state whose likelihood is a constant $\\exp(-\\tau^2/\\sigma^2)$ with $\\tau=3\\sigma$. In the M-step, the images for all $K$ motion states are updated by minimizing a weighted least-squares data-fidelity term plus anisotropic total-variation regularization along spatial, cardiac, and respiratory dimensions, solved approximately with ADMM. The binning prior $\\theta(n,k)$ assigns probability $0.85$ to the initial self-gating assignment, $0.05$ to the outlier bin, and spreads the remainder over the other bins, which stabilizes the high-dimensional EM loop but caps how far a readout can move from its original bin.","core_discovery":"The paper's central claim is that residual motion artifacts in self-gated 5D cardiac MRI arise from two recoverable causes: valid readouts assigned to the wrong motion bin, and readouts corrupted by bulk motion that belong to no valid bin. By treating each readout's true bin assignment as a latent variable, EMORe uses an expectation-maximization loop that alternates between computing soft posterior probabilities of bin membership and updating the reconstructed images. The outlier bin, the (K+1)-th state, collects readouts inconsistent with all valid motion states, so corrupted data are rejected rather than blurring the image. In the MRXCAT phantom study, which includes 50 simulated scans spanning 0% to 70% bulk-motion corruption, EMORe outperforms standard compressed sensing in peak signal-to-noise ratio, structural similarity index, edge sharpness, and bin-assignment accuracy as measured by Brier score. In 13 in vivo scans, including three with instructed coughing, EMORe significantly improves blood-myocardium edge sharpness and expert artifact scores relative to compressed sensing.","pith_inferences":["A boundary condition the paper does not test: the informative prior pins 85% of each readout's prior probability to the initial self-gating bin, so if a clinical sequence produces systematic binning errors above roughly 15%, EMORe's corrections would be throttled; a natural extension is to make $\\alpha_g$ adaptive per readout based on self-gating signal confidence.","EMORe corrects discrete bin membership but does not model continuous motion within a bin; combining it with intra-bin non-rigid motion estimation could address residual blur that discrete reassignment cannot fix.","The outlier bin could double as a diagnostic signal: the fraction of readouts assigned to it over time is a data-driven record of motion corruption that might flag segments of the acquisition for re-scan or guide prospective gating in future sequences.","The Brier-score improvement reported even at 0% simulated outliers suggests soft binning can serve as a post-hoc quality metric for self-gating accuracy, independent of image content."],"forward_implications":["If EMORe works as claimed, 5D cardiac MRI can tolerate sporadic bulk motion such as coughs, twitches, and deep breaths without requiring navigator echoes, breath holds, or rescanning.","Valid but misassigned readouts are corrected rather than thrown away, so the effective acceleration rate is not increased by outlier rejection; only genuinely corrupted readouts land in the outlier bin.","The method is a post-acquisition software change: the same k-space data, trajectory, and self-gating signals feed both compressed sensing and EMORe, so existing 5D MRI protocols can adopt it without changing the scan.","In the in vivo results, the gains are largest where clinical need is greatest, namely irregular breathing and instructed coughing, so patients with such motion may benefit most.","Because the M-step is a generalized EM with only a few ADMM iterations, the framework can be ported to other motion-resolved MRI settings, such as 4D flow, with similar regularization."],"supporting_citations":[{"why":"Supplies the expectation-maximization framework, including the generalized-EM partial M-step and the sensitivity-to-initialization caveat used in the algorithm design.","marker":"[32]"},{"why":"Earlier EM-based outlier rejection in fetal cardiac MRI that discards misassigned valid data; EMORe's contribution is to add inter-bin correction instead.","marker":"[33]"},{"why":"Standard compressed-sensing reconstruction serves as the baseline that EMORe must beat in every phantom and in vivo comparison.","marker":"[37]"},{"why":"Prior intra-bin respiratory-motion correction in 5D MRI that assumes perfect self-gating; it marks the state of the art EMORe extends by relaxing that assumption.","marker":"[27]"},{"why":"Provides the free-running self-gated acquisition and signal-extraction pipeline on which the phantom and in vivo data are modeled.","marker":"[11]"},{"why":"MRXCAT numerical phantom supplies the ground-truth anatomies, cardiac and respiratory dynamics, and simulated k-space data for the phantom study.","marker":"[41]"},{"why":"Edge-sharpness assessment by parametric modeling is the primary quantitative outcome measure used in both phantom and in vivo studies.","marker":"[39]"},{"why":"Brier score quantifies the bin-assignment accuracy improvement that is central to the claim that EMORe corrects misbinning.","marker":"[40]"}],"fun_headline_variants":["EMORe outsmarts motion artifacts in 5D cardiac MRI","Motion-corrupted MRI data? EMORe bins them out","EMORe upgrades 5D heart MRI with smart binning","Soft bin assignment sharpens 5D cardiac MRI","EMORe beats compressed sensing on motion-robust MRI"],"cache_read_input_tokens":16256,"weakest_assumption_plain":"The method's correction power rests on the initial self-gating binning being correct for the large majority of readouts, because the prior pins $0.85$ of each readout's prior probability to its original bin; if systematic self-gating errors exceed roughly $15\\%$, the EM loop is biased toward the wrong bins and may not recover them, a sensitivity the authors themselves note.","fun_headline_variants_meta":{"raw":{"variants":["EMORe outsmarts motion artifacts in 5D cardiac MRI","Motion-corrupted MRI data? EMORe bins them out","EMORe upgrades 5D heart MRI with smart binning","Soft bin assignment sharpens 5D cardiac MRI","EMORe beats compressed sensing on motion-robust MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000131,"raw_usage":{"total_tokens":1182,"prompt_tokens":1050,"completion_tokens":132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":45}},"tokens_in":666,"tokens_out":132,"duration_ms":2321,"temperature":1.0,"reasoning_tokens":45,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:56:06.384973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EMORe on a phantom dataset in which a controlled fraction of readouts (e.g., 20–40%) is deliberately assigned to the wrong motion bin by corrupting the self-gating signal, and compare PSNR and Brier score against compressed sensing; the central claim would be refuted if EMORe's advantage disappears or if its final bin assignments remain as wrong as the initialization. A minimal version is to lower the prior weight $\\alpha_g$ from $0.85$ to $0.5$ and show that EMORe's output quality collapses once the true bins are no longer protected.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the expectation-maximization framework, including the generalized-EM partial M-step and the sensitivity-to-initialization caveat used in the algorithm design."},{"cited_title":"Fetal cardiac cine imaging using highly accelerated dynamic MRI with retrospective motion correction and outlier rejection,","cited_arxiv_id":null,"evidence_quote":"Earlier EM-based outlier rejection in fetal cardiac MRI that discards misassigned valid data; EMORe's contribution is to add inter-bin correction instead."},{"cited_title":"Intra-bin correction and inter-bin compensation of respiratory motion in free-running five-dimensional whole-heart mag- netic resonance imaging,","cited_arxiv_id":null,"evidence_quote":"Prior intra-bin respiratory-motion correction in 5D MRI that assumes perfect self-gating; it marks the state of the art EMORe extends by relaxing that assumption."},{"cited_title":"Fully self-gated whole-heart 4D flow imaging from a 5-minute scan,","cited_arxiv_id":null,"evidence_quote":"Provides the free-running self-gated acquisition and signal-extraction pipeline on which the phantom and in vivo data are modeled."},{"cited_title":"MRXCAT: Realistic numerical phantoms for cardiovascular magnetic resonance,","cited_arxiv_id":null,"evidence_quote":"MRXCAT numerical phantom supplies the ground-truth anatomies, cardiac and respiratory dynamics, and simulated k-space data for the phantom study."},{"cited_title":"Edge sharpness assessment by parametric modeling: application to magnetic resonance imaging,","cited_arxiv_id":null,"evidence_quote":"Edge-sharpness assessment by parametric modeling is the primary quantitative outcome measure used in both phantom and in vivo studies."},{"cited_title":"Verification of forecasts expressed in terms of probability,","cited_arxiv_id":null,"evidence_quote":"Brier score quantifies the bin-assignment accuracy improvement that is central to the claim that EMORe corrects misbinning."}],"review_version":1}