{"id":"5e8ec569-5c81-463f-ae78-d6629fee65eb","arxiv_id":"2411.12919","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Training MRI reconstruction networks on GSURE-denoised images improves accuracy and speed in low-SNR settings, with only marginal differences at native SNR.","lead":"This paper tests whether cleaning noisy MRI training images with a self-supervised denoiser makes deep-learning reconstruction models perform better. It finds clear gains at low signal-to-noise ratios on brain and knee scans, for both diffusion-based and unrolled reconstruction models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All gains are measured against noisy native-SNR references, and no clean-reference oracle is tested; the reported improvements may partly reflect reference-noise retention rather than true signal fidelity, and the 'bypassing noise-free references' claim lacks its necessary control.","rationale":"The reader's verdict of CONDITIONAL is appropriate. My stress-test identifies a load-bearing gap that is related to, but not identical with, the reader's designated weakest assumption. The reader emphasizes the Gaussian-noise and low-field transfer assumption, while my primary concern is the absence of any clean-reference evaluation and clean-trained oracle, which makes the observed quantitative gains and the 'bypassing noise-free references' conclusion ambiguous. The authors themselves flag the biased reference and hidden-noise issues, so this is a conceded limitation rather than an external attack. However, the concern is about evidence strength, not internal inconsistency: the experiments are internally coherent, the GSURE methodology is principled, and the paired comparisons across anatomies, accelerations, SNR levels, and model families provide substantial support for the narrow claim that denoised targets beat noisy targets at low SNR. For that reason I do not recommend REJECT; a conditional acceptance with a clean-reference validation step is the correct posture. The agreement_with_reader is 'partial' because the reader's rationale does mention the noisy-reference bias as a weakness, though their chosen weakest_assumption is the noise-model/low-field transfer assumption rather than the missing clean-reference oracle.","tokens_in":21295,"tokens_out":14758,"duration_ms":164941,"concrete_test":"Build a synthetic multi-coil MRI benchmark with known clean complex images (e.g., BrainWeb-style phantoms or high-SNR multi-average acquisitions) and simulate the same brain and knee SNR levels (32/22/12 dB and 24/14/4 dB) with R=4 and R=8 sampling. Train GSURE-DPS, Naive-DPS, and an oracle Clean-DPS trained on the true clean references, and likewise for MoDL, using the released code. Recompute NRMSE, SSIM, and PSNR against the known clean reference. If GSURE-DPS and GSURE-MoDL remain significantly better than Naive-DPS and Naive-MoDL at 12/4 dB and approach the clean-oracle performance, the concern is resolved; if the advantage shrinks, reverses, or falls far short of the oracle, the paper's central and 'bypassing clean references' claims are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests entirely on error metrics computed against the native-SNR fully sampled fastMRI image, which the authors concede is noisy (Evaluation Metrics; also Discussion's hidden-noise discussion). With reference r = c + n0, the expected squared error is E||e-r||^2 = E||e-c||^2 + sigma0^2 - 2E[(e-c)^T n0]. An estimate that reproduces a component of the reference noise n0 is rewarded, while an estimate that suppresses n0 is penalized. Because the low-SNR test measurements are the native fully sampled data plus added Gaussian noise, the same native noise realization n0 appears in both the measurements and the reference; data-consistency steps in DPS and MoDL can therefore create exactly this covariance. Thus the reported NRMSE/SSIM/PSNR gaps at 12 dB and 4 dB are not directly interpretable as gains in true signal fidelity. The missing control is more than a metric caveat: the abstract's stronger claim that GSURE denoising can 'potentially bypass the need for noise-free reference MRI scans' requires showing that training on GSURE-denoised targets approximates training on clean references. The experiments compare GSURE-denoised training only against noisy-target training, never against an oracle trained on clean or near-clean references, so the surrogate-quality claim is unsupported. The authors' own limitation statements about hidden noise and the need for radiologist evaluation reinforce, rather than resolve, this gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for low-SNR multi-coil MRI reconstruction: pre-whiten and normalize the k-space data, train a UNet denoiser with the GSURE self-supervised loss, and then use the denoised coil-combined images as training targets for two reconstruction methods, diffusion posterior sampling (DPS) with an EDM prior and Model-Based Deep Learning (MoDL). The experiments compare these GSURE-trained models against the same architectures trained on the original noisy targets, across T2-weighted brain and fat-suppressed knee fastMRI data, at three SNR levels per anatomy and acceleration factors R=4 and R=8. The authors report that GSURE training targets yield lower NRMSE and higher SSIM/PSNR at low SNR, with comparable performance at native SNR, and conclude that self-supervised denoising may bypass the need for noise-free reference scans. The paper includes 72 paired Wilcoxon tests with Bonferroni correction, a pathology example, a training-set-size study, and open-source code.","tokens_in":21588,"tokens_out":9730,"duration_ms":98212,"significance":"If the reported effects are genuine, the pipeline is a practically useful preprocessing recipe for low-SNR MRI, particularly for low-field systems, and the head-to-head comparison of DPS and MoDL under identical preprocessing is a useful reference for the community. The paper is strong in its breadth: two reconstruction paradigms, three SNR levels per anatomy, two acceleration factors, statistical testing, qualitative pathology assessment, and a concrete data-efficiency study. The authors are also transparent about the main limitation, namely that all quantitative metrics are computed against inherently noisy native-SNR references, and they explicitly call for NCE-based and radiologist evaluation in future work. The central quantitative claim, however, rests on metrics whose interpretation is confounded by the presence of the same native noise realization in both the measurements and the reference; the missing clean-reference oracle control leaves the strongest conclusion, the bypass of noise-free references, unsupported.","major_comments":[{"comment":"The primary quantitative evidence is computed against native-SNR fully sampled images, which the paper itself calls inherently noisy (Evaluation Metrics, Discussion). Writing the reference as r = c + n0 and the estimator as e, the expected squared error is E||e-r||^2 = E||e-c||^2 + sigma0^2 - 2E[(e-c)^T n0]. The low-SNR test measurements are generated by adding Gaussian noise to the same fully sampled k-space that contains n0, and the data-consistency steps in DPS (Eq. 12) and MoDL (Eq. 14) can create dependence between e and n0. Two estimators equally close to c can therefore receive very different NRMSE/PSNR/SSIM scores depending on whether they reproduce components of n0. The cross term has no definite sign, so the reported GSURE-DPS and GSURE-MoDL gains at 12 dB and 4 dB are not directly interpretable as gains in signal fidelity. This is not merely a caveat: the abstract's quantitative claims are stated in these metrics. The paper should add an experiment with known-clean references (e.g., synthetic ground truth with controlled noise) or at least report the NCE metric and show that the conclusion is unchanged.","section":"Evaluation Metrics; Data; Discussion"},{"comment":"The conclusion that GSURE preprocessing may 'potentially bypass the need for noise-free reference MRI scans' is not supported by the experimental design. The comparisons are exclusively between networks trained on GSURE-denoised targets and networks trained on noisy targets; there is no oracle condition trained on clean or near-clean references. Consequently, the results show only that GSURE-denoised targets are better than noisy targets under the chosen metric, not that they approximate clean targets well enough to substitute for them. A direct control, training the same DPS and MoDL pipelines on simulated clean references from known ground truth or on multi-average low-noise data, would make the surrogate-quality claim testable. Without such a control, the 'bypass' claim should be removed or explicitly labeled as a hypothesis.","section":"Proposed Approach; Conclusion; Abstract"},{"comment":"The theoretical guarantee of GSURE and the validity of the simulated SNR conditioning both rely on the assumption that, after pre-whitening and ACS normalization, the noise is i.i.d. Gaussian with known scalar variance sigma^2_eta* (Eq. 16). The experiments add zero-mean Gaussian noise to already noisy fastMRI k-space, so this assumption is satisfied by construction in the simulated low-SNR conditions. However, the Discussion motivates the method by low-field and real-world acquisitions, where noise can be non-Gaussian, spatially varying, or correlated across coils, and where sensitivity estimation may itself be noise-contaminated. The paper should either restrict the claims to the Gaussian-noise setting or include a robustness experiment with a non-Gaussian or correlated noise model; as written, the extrapolation to real low-SNR MRI is a conjecture.","section":"Theory; Eq. (16); Discussion"}],"minor_comments":[{"comment":"'Proof-of-principal' should be 'proof-of-principle'.","section":"Proposed Approach"},{"comment":"'Eluded another limitation' should read 'alluded to another limitation.'","section":"Discussion"},{"comment":"'These experimentes' is a typo for 'these experiments.'","section":"Experiments"},{"comment":"The captions state 'GSURE-MoDL outperforms Naive-DPS'; the comparison is against Naive-MoDL, so the captions should refer to Naive-MoDL.","section":"Figures S7 and S8"},{"comment":"'Wehighlight' in the table captions should be 'we highlight.'","section":"Tables 1 and 2"},{"comment":"The SNR definition should explicitly state that the normalizing constant is the 99th percentile of the ACS root-sum-of-squares reconstruction, since that normalization sets the signal scale used in the SNR expression.","section":"Eq. (17)"},{"comment":"The statement that GSURE 'can handle ... each coil can have different noise levels' is not demonstrated by the implementation, which pre-whitens the data to a scalar noise variance; please rephrase or provide a per-coil experiment.","section":"Self-Supervised Denoising"},{"comment":"The training durations '200 iterations' for the GSURE network and '3,000 iterations' for the EDM network should be clarified as epochs or iterations with the stated batch size; 200 iterations seems surprisingly short for a 65M-parameter UNet.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is the arXiv version of a paper already published in Magnetic Resonance in Medicine, so the recommendation is relative to that context. The main scientific weakness is the absence of a clean-reference oracle control, which leaves the 'bypass noise-free references' claim unsupported; the noisy-reference bias compounds this. The authors are unusually transparent about the hidden-noise limitation and have released code, and the within-comparison results are extensive. I do not see grounds for rejection, but I would ask for the oracle control or a softening of the bypass claim before considering the central conclusion fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know upfront: this is a genuinely useful empirical paper, but its own stated headline promise—that GSURE denoising can bypass the need for noise-free references—is not actually tested. What is tested is whether training on GSURE-denoised targets beats training on noisy targets for two reconstruction families across a range of SNRs. On that more modest claim, the evidence is fairly strong.\n\nThe systematic comparison across brain and knee anatomies, DPS and MoDL, three SNR levels, two acceleration factors, pathology, training-set size and reconstruction speed, plus 72 paired tests and released code, is a real extension beyond the authors' earlier conference abstracts. Including an ENSURE comparison is useful.\n\nThe control is right: same architecture and training procedure, only training targets differ. Low-SNR gains are consistent across anatomies and model families, and the pathology example is convincing at face value. The authors are explicit that native-SNR references are noisy and call for better metrics and radiologist reads. That honesty is worth something.\n\nBut the missing oracle is more than a caveat. Because the reference images are the native fully sampled scans, and the same native noise realization also lives in the low-SNR test measurements, data consistency in DPS and MoDL can reproduce a component of that noise; NRMSE, PSNR, and SSIM can reward that. So the reported gaps at 12 dB and 4 dB may partly reflect reference-noise retention rather than true signal fidelity. The claim that GSURE-denoising could bypass the need for clean references requires a model trained on clean or near-clean references as the oracle comparison; no such experiment is present. The word \"essential\" in the conclusion also overstates the case—at native SNR the differences are small and sometimes favor the naive model.\n\nMinor: Tables 1 and 2 give means without variance, though box plots in the supplement cover that. The i.i.d. Gaussian noise assumption after pre-whitening is reasonable for fastMRI, but transfer to real low-field data remains speculative, which the authors acknowledge.\n\nWho this is for: anyone working on self-supervised MRI reconstruction or low-field MRI. They will want to know the empirical result even while noting the evaluation gap. The paper deserves a serious referee; the overall design is careful, the contribution is incremental but real, and code is available. A revision should add an oracle-trained reference and NCE metric.\n\nRecommendation: engage with it; treat the low-SNR improvements as promising but not fully established, and press for a clean-reference control before relying on the stronger claim.","headline":"Empirically careful study showing GSURE-denoised training targets help at low SNR across two model families, but the 'bypass clean references' claim is untested because no clean-reference oracle is included.","tokens_in":22197,"tokens_out":3755,"would_cite":true,"duration_ms":37792,"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 argues that GSURE-based denoising of MRI training data improves low-SNR diffusion and MoDL reconstruction.","keywords":["MRI reconstruction","self-supervised denoising","GSURE","diffusion probabilistic models","MoDL","fastMRI","low-SNR imaging","parallel imaging"],"falsifier":"Acquire real low-field MRI k-space with a noise pre-scan, run the same two-stage pipeline with its measured non-Gaussian or spatially correlated noise, and check whether reconstructions trained on GSURE-denoised data still beat reconstructions trained on noisy data; if they do not, or if the gains vanish when sensitivity maps are estimated from the noisy data itself, the paper's central claim would not transfer beyond the simulated setting.","tokens_in":21015,"feed_emoji":"🩻","tokens_out":6992,"duration_ms":66219,"temperature":0.7,"pith_summary":"The paper asks whether adding a self-supervised denoising step before training makes deep-learning MRI reconstruction better, and argues that it does. Using GSURE, an unbiased loss that estimates the mean-squared error of a denoiser from noisy measurements alone, the authors train a denoiser on noisy multi-coil k-space and use its output as the training target for two reconstruction families: diffusion probabilistic models sampled with diffusion posterior sampling, and the unrolled model-based deep learning network MoDL. On T2-weighted brain and fat-suppressed knee scans from the fastMRI dataset at three SNR levels each and acceleration factors 4 and 8, denoised training targets give lower NRMSE, higher SSIM, and higher PSNR at low SNR, and performance comparable to noisy-data training at native SNR. The authors conclude that denoising can replace the need for noise-free reference scans when training reconstruction networks for low-SNR settings such as low-field MRI.","feed_headline":"Denoising MRI training data boosts low-SNR reconstruction","feed_subtitle":"Self-supervised denoising improves diffusion and MoDL reconstructions without needing clean reference MRI scans.","key_machinery":"The central object is GSURE (Generalized Stein's Unbiased Risk Estimate), a loss that computes an unbiased estimate of the supervised mean-squared-error loss from noisy measurements alone. The authors implement it by adding the divergence of the denoiser network, approximated with a Monte Carlo perturbation, to a data-consistency term, and applying it to pre-whitened, normalized adjoint images scaled by the estimated noise variance. This GSURE loss trains a UNet denoiser to produce MMSE-denoised coil-combined images, which then replace the raw noisy images as the training targets for an EDM diffusion prior and for MoDL.","core_discovery":"On the paper's own terms, the discovery is that GSURE-denoised training targets improve both generative and end-to-end accelerated MRI reconstruction without any clean reference images. Denoised targets yield better priors: unconditional samples from the diffusion model are cleaner, and DPS reconstructions degrade more gracefully as inference SNR drops, including on a knee case with a meniscus tear. The same pattern holds for MoDL, with the largest margins at the lowest simulated SNRs (12 dB brain, 4 dB knee), and GSURE-trained models reach target NRMSE with fewer posterior averages and fewer training examples. The authors interpret the gains as coming from the denoising operation itself, not from the noise level of the training data, because low-SNR GSURE models sometimes beat higher-SNR GSURE models while the equivalent noisy-data models do not.","pith_inferences":["If real low-field MRI noise behaves like the simulated Gaussian noise used here, this two-stage recipe could make clinical training pipelines practical at 0.55 T and lower, where clean references are unavailable; the paper only demonstrates the simulated case.","Because the gains concentrate at low SNR and are small or negative for some high-SNR SSIM comparisons, the practical value likely depends on how far a deployment is from native SNR rather than on uniform improvement.","A single-stage formulation that learns the noisy distribution directly while incorporating denoising is theoretically possible and the paper leaves it open; the reported stability of the two-stage approach is an empirical argument for keeping the steps separate.","The core premise about known Gaussian noise could be tested by estimating the noise covariance from a pre-scan on real low-field systems and checking whether GSURE-trained networks still beat noisy-data-trained ones."],"forward_implications":["Noise-free reference scans are not required for training low-SNR reconstruction models; noisy fully sampled data plus GSURE denoising can serve as the target.","At low inference SNR, both diffusion-based and unrolled reconstructions become more accurate, with the largest gains at the worst noise levels tested.","Training and inference become more efficient: fewer posterior samples and fewer training examples reach a given NRMSE after denoising.","The two-stage preprocessing pipeline is architecture-agnostic, so new denoisers or new reconstruction networks could be substituted without changing the core idea.","In the unrolled case, GSURE-denoised training also outperforms an alternative self-supervised reconstruction loss when the training SNR is low."],"supporting_citations":[{"why":"Supplies GSURE, the unbiased risk estimator that defines the self-supervised denoising loss.","marker":"[58]"},{"why":"Provides the Monte Carlo divergence approximation that makes GSURE training tractable for large MRI images.","marker":"[41]"},{"why":"Defines the EDM diffusion architecture and training loss used for the generative priors.","marker":"[26]"},{"why":"Provides the diffusion posterior sampling algorithm used for reconstruction from the trained priors.","marker":"[60]"},{"why":"Defines the MoDL unrolled architecture trained on noisy versus GSURE-denoised targets.","marker":"[14]"},{"why":"Supplies the fully sampled T2 brain and fat-suppressed knee k-space data for denoising, training, and evaluation.","marker":"[59]"},{"why":"Serves as the alternative self-supervised reconstruction baseline that GSURE-MoDL is compared against at low training SNR.","marker":"[53]"}],"fun_headline_variants":["GSURE denoising boosts MRI reconstruction without clean data","Denoise training data, improve MRI at low SNR","No clean scans needed: denoised training lifts MRI","Self-supervised denoising sharpens MRI reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline is unbiased only if, after pre-whitening and normalization, the measurement noise is independent Gaussian noise with known variance; if real low-SNR acquisitions violate that, the reported gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["GSURE denoising boosts MRI reconstruction without clean data","Denoise training data, improve MRI at low SNR","No clean scans needed: denoised training lifts MRI","Self-supervised denoising sharpens MRI reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000667,"raw_usage":{"total_tokens":3084,"prompt_tokens":1027,"completion_tokens":2057,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1991}},"tokens_in":643,"tokens_out":2057,"duration_ms":16062,"temperature":1.0,"reasoning_tokens":1991,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:03:28.887595+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire real low-field MRI k-space with a noise pre-scan, run the same two-stage pipeline with its measured non-Gaussian or spatially correlated noise, and check whether reconstructions trained on GSURE-denoised data still beat reconstructions trained on noisy data; if they do not, or if the gains vanish when sensitivity maps are estimated from the noisy data itself, the paper's central claim would not transfer beyond the simulated setting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GSURE, the unbiased risk estimator that defines the self-supervised denoising loss."},{"cited_title":"Monte-Carlo SURE: A black-box optimization of regular- ization parameters for general denoising algorithms.IEEE Transactions on image processing 17.(9) (2008), 1540–1554","cited_arxiv_id":null,"evidence_quote":"Provides the Monte Carlo divergence approximation that makes GSURE training tractable for large MRI images."},{"cited_title":"Elucidating the design space of diffusion-based generativemodels","cited_arxiv_id":null,"evidence_quote":"Defines the EDM diffusion architecture and training loss used for the generative priors."},{"cited_title":"T., Klasky, M","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion posterior sampling algorithm used for reconstruction from the trained priors."},{"cited_title":"K., Mani, M","cited_arxiv_id":null,"evidence_quote":"Defines the MoDL unrolled architecture trained on noisy versus GSURE-denoised targets."},{"cited_title":"fastMRI: A publicly available raw k-space and DI- COM dataset of knee images for accelerated MR image reconstruction using machine learning","cited_arxiv_id":null,"evidence_quote":"Supplies the fully sampled T2 brain and fat-suppressed knee k-space data for denoising, training, and evaluation."},{"cited_title":"K., Pramanik, A., John, M., and Jacob, M","cited_arxiv_id":null,"evidence_quote":"Serves as the alternative self-supervised reconstruction baseline that GSURE-MoDL is compared against at low training SNR."}],"review_version":1}