{"id":"9b2401cf-20cc-4153-915b-8c2d9994efb9","arxiv_id":"2608.08693","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CUPA-T2* propagates inter-echo covariance from accelerated MRI reconstruction into T2* fitting, improving white-matter accuracy and uncertainty alignment, while trading off calibration quality.","lead":"This paper proposes a framework that passes uncertainty from accelerated MRI reconstruction into T2* fitting, using covariance-aware sampling and a correlation regularizer. It reports modest white-matter accuracy gains at higher acceleration rates and much stronger alignment between reconstruction uncertainty and predicted T2* variance, at the cost of worse calibration and selective prediction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 24×24 inter-echo covariance in §2.2 is estimated from only 100 MC samples per voxel with no shrinkage or conditioning; if ill-conditioned, Cholesky sampling injects estimation noise and the reported WM benefit may not reflect true covariance structure.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the 100-sample per-voxel covariance is the foundation of the covariance-aware sampling mechanism, and its reliability is not independently validated. I considered whether the circularity of the alignment metric (ρ is the training objective) is more fundamental, but the paper honestly reports the resulting AURC/ECE trade-off, and the alignment increase is a design objective rather than an unsupported empirical discovery. The covariance reliability issue is more load-bearing because the central novel mechanism, inter-echo covariance propagation, collapses if Σ_v is dominated by estimation noise. The proposed split-half and conditioning checks would settle this directly, and the shrinkage-covariance re-ablation would show whether the off-diagonal benefit in Table 4 is robust. Since the reader's CONDITIONAL verdict already requires additional validation, this concern does not move the verdict; it strengthens the case for keeping the conditionality and adding the specific covariance diagnostics to the revision requirements.","tokens_in":7693,"tokens_out":7352,"duration_ms":86884,"concrete_test":"On the test slices, split the S=100 MC reconstructions into two independent 50-sample halves. For each voxel compute Σ_v^(1) and Σ_v^(2), and report the median relative Frobenius difference ||Σ_v^(1)-Σ_v^(2)||_F / ||(Σ_v^(1)+Σ_v^(2))/2||_F, together with the median condition number and minimum eigenvalue of the full-sample Σ_v. If the half-split difference is large relative to the off-diagonal signal (e.g., >50%) or the condition number exceeds 1e4, S=100 is insufficient for stable Cholesky propagation. In addition, rerun the Table 4 full-covariance vs ΣI ablation at R=4 using a shrinkage covariance estimator (e.g., Ledoit-Wolf) to check whether the WM NRMSE gain (0.168 vs 0.173) persists when the covariance estimate is regularized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core novelty of CUPA-T2* is covariance-aware propagation in Eq. (1), and the only source of the inter-echo covariance is Σ_v, computed per voxel from S=100 MC dropout reconstructions (§2.2). Each voxel is described by a 2E=24-dimensional real vector, so Σ_v is a 24×24 sample covariance from 100 draws. With no shrinkage, eigenvalue clipping, or spatial pooling described, the off-diagonal entries that distinguish Σ_v from a diagonal covariance carry sampling noise of order 1/sqrt(100), and the matrix is likely ill-conditioned because MC dropout reconstructions are highly correlated across echoes. Cholesky factorization of a nearly singular empirical covariance produces an L_v that amplifies directions with small estimated variance, so the samples in Eq. (1) can be dominated by estimation error rather than true inter-echo structure. The paper itself asserts in §3.3 that WM benefits arise because white matter yields 'more reliable inter-echo covariance estimates,' making the reliability of Σ_v load-bearing for the central claim. If the covariance is unreliable, the R=4 ablation difference between full covariance and diagonal ΣI (Table 4) may reflect dataset-specific noise rather than a principled off-diagonal effect, and the claimed covariance-aware advantage would not generalize.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CUPA-T2*, a two-stage framework for T2* mapping from accelerated multi-echo MRI. It estimates a per-voxel inter-echo covariance matrix from S=100 Monte Carlo dropout reconstructions, factorizes it via Cholesky decomposition, and generates uncertainty-aware complex-valued samples whose magnitudes feed a heteroscedastic MLP. The MLP is trained with a beta-NLL loss and an RMS-based Pearson correlation regularizer that aligns predicted log-variance with reconstruction-derived RMS uncertainty. Experiments on the public T2*-MOVE dataset show statistically significant white-matter NRMSE improvements at R=3 and R=4 relative to two baselines, along with a large increase in the correlation between reconstruction uncertainty and predicted variance, at the cost of worse ECE and AURC. An ablation at R=4 isolates the contributions of covariance-aware sampling, full versus diagonal covariance, and the alignment weight.","tokens_in":8041,"tokens_out":6267,"duration_ms":63184,"significance":"If the technical claims hold, CUPA-T2* is a meaningful step beyond scalar uncertainty propagation in accelerated quantitative MRI, explicitly carrying inter-echo covariance structure into downstream parameter fitting. The paper's strengths include public code and dataset, a clear ablation design, and statistical tests on tissue-specific metrics. The alignment regularizer and the WM improvements are potentially useful for biomarker studies if the covariance estimation is reliable. However, the central headline result—the increased alignment ρ—is to a large extent an optimization artifact because ρ is exactly the objective being maximized, and the reliability of the per-voxel covariance estimate remains a load-bearing but unexamined assumption.","major_comments":[{"comment":"The reported increase in ρ is expected by construction, because ρ is the exact objective being optimized through L = L_β-NLL + λ(1−ρ). Reporting this as a finding is circular; it should be framed as a sanity check that the regularizer works as intended. The independent evidence for the method lies in the WM NRMSE improvements and the AURC/ECE trade-off, and these should be foregrounded in the abstract and conclusions.","section":"§2.3, Table 3"},{"comment":"The per-voxel 24×24 empirical covariance Σ_v is estimated from only S=100 Monte Carlo samples, with no shrinkage, eigenvalue clipping, or spatial pooling described. The paper's own explanation for the WM benefit is that white matter yields 'more reliable inter-echo covariance estimates,' which makes the conditioning of Σ_v load-bearing for the central claim. Please report the condition number or eigenvalue spectrum of Σ_v, or introduce covariance regularization, and demonstrate that the Cholesky-based sampling is not dominated by estimation noise rather than true inter-echo structure.","section":"§2.2, §3.3"},{"comment":"The test-time protocol is ambiguous. The method section states that the magnitude of the sampled representation x̃_v serves as input to the fitting network, but later states 'During validation and testing, the MC mean and standard deviation are used.' If sampling is applied only during training as a data-augmentation strategy and the test-time input is deterministic, then the claim of explicitly propagating covariance to downstream fitting at inference is not supported. Please clarify whether test-time inputs are stochastic, and, if not, explain how the covariance structure is preserved at inference.","section":"§2.3"},{"comment":"The sampling coefficient α_v is a function of u_v, which is itself derived from the trace of the same covariance matrix that produces L_v. This double-counts the uncertainty magnitude and may confound the effect of the off-diagonal covariance structure with the noise schedule. An ablation with a constant α (or with α based on an independent uncertainty estimate) would help isolate the contribution of the covariance structure per se.","section":"§2.2, Eq. (1)"},{"comment":"The reported WM improvements are small (0.135 vs 0.137 at R=3; 0.168 vs 0.174 at R=4). Since the paper performs many tissue × acceleration Wilcoxon signed-rank tests at p=0.001, please state whether any multiple-comparison correction was applied and report the number of independent units (subjects or slices) entering each test. Without this, the statistical superiority claims are not fully supported.","section":"§3.3, Table 2"}],"minor_comments":[{"comment":"The definition of the β-NLL loss in Eq. (2) is difficult to parse because of the nested exp/log structure; please add a short derivation or reformulate it to make the stop-gradient and β-exponent roles clearer.","section":"§2.3, Eq. (2)"},{"comment":"The abbreviation PUQ is used for the baseline from Sun et al. but the text calls it a 'non-heteroscedastic MLP (PUQ)'; please define the acronym explicitly in the experiments section to avoid confusion with the phase-wise uncertainty method in the cited work.","section":"§3.2"},{"comment":"In the ablation, λ=0.0 achieves the lowest WM NRMSE (0.167) while the full method (λ=0.05) yields 0.168; the discussion should explicitly address why λ=0.05 was chosen despite this, beyond the statement that it gave the best validation performance in most settings.","section":"Table 4"},{"comment":"The text says 'a covariance factor L_v is calculated such that Σ_v ≈ L_v L_v^T' but does not state how non-positive-definite empirical covariances are handled; please mention whether eigenvalues are clamped or a diagonal jitter is added.","section":"§2.2"},{"comment":"The paper does not report the computational overhead of generating S=100 MC reconstructions and computing per-voxel covariance matrices; a brief runtime or complexity note would help readers assess practical feasibility.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the code release is a positive. The main risk is not the novelty but the gap between the claimed 'covariance-aware propagation' and the actual test-time mechanism, plus the unexamined conditioning of the empirical covariance. The circularity of the ρ metric is a framing issue that can be fixed by repositioning the claim. I would encourage the editors to request the clarifications and diagnostics outlined in the major comments before accepting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a real methodological step: nobody else, as far as I can tell, has propagated voxel-wise inter-echo covariance from stochastic reconstructions into downstream T2* fitting, and the tissue-stratified RMS correlation regularizer is a reasonable way to avoid variance-dominated alignment. The authors release code and are honest about the calibration/selective prediction trade-off. Second, the paper's strongest-sounding result—that CUPA-T2* substantially increases alignment ρ between reconstruction uncertainty and predicted variance—is largely circular, because the training loss directly maximizes that Pearson correlation. The ablation shows λ=0 gives ρ=0.250, λ=0.01 gives 0.370, so it's doing what it's told. That's not a discovery; it's a check on the optimizer.\n\nThe independently interesting results are the tissue-specific ones: WM NRMSE improves at R=3 and R=4 (0.135 vs 0.137, 0.168 vs 0.174), while overall NRMSE is slightly worse than the simpler PUQ baseline. The effect sizes are small, but the authors use proper statistical tests and don't oversell. The covariance ablation (ΣI vs full Σ) also suggests the off-diagonal structure matters, which is the core claim.\n\nNow the soft spots. The inter-echo covariance is estimated per voxel from only S=100 MC dropout reconstructions, giving a 24×24 sample covariance with no shrinkage, eigenvalue clipping, or spatial pooling. Hundred samples in 24 dimensions from a highly correlated process will produce a poorly conditioned matrix, and the Cholesky factor will amplify directions of small estimated variance. The paper even attributes the WM benefit to 'more reliable inter-echo covariance estimates' in white matter, which makes this assumption load-bearing. If the covariance is noisy, the ΣI vs full Σ comparison could reflect estimation noise rather than true inter-echo structure. This needs a sensitivity analysis—e.g., varying S, adding shrinkage, or comparing against a larger-sample reference.\n\nAlso missing: values for α_min, α_max, and γ in the sampling schedule are never given, which makes the method hard to reproduce. The λ and β are reported, so the omission is fixable.\n\nBottom line: worth a serious referee. The idea is new, the writing is clear, the limitations are acknowledged, and the code is out. But the circular alignment metric and the unvalidated covariance estimate need to be addressed before I'd trust the quantitative claims. Recommend major revision, not rejection.","headline":"Covariance-aware uncertainty propagation for T2* mapping is a genuine idea, but the reported alignment gain is partly circular and the 100-sample covariance estimate needs validation.","tokens_in":8518,"tokens_out":2873,"would_cite":false,"duration_ms":30056,"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":"CUPA-T2* propagates the full voxel-wise inter-echo covariance from accelerated MRI reconstructions into downstream T2* fitting, improving white-matter error at R=3 and R=4 while aligning predicted variance with reconstruction uncertainty.","keywords":["T2* mapping","quantitative MRI","uncertainty propagation","Monte Carlo dropout","covariance-aware sampling","heteroscedastic regression","accelerated MRI","uncertainty alignment"],"falsifier":"On the T2*-MOVE test set at R=4, randomly permute the off-diagonal entries of every voxel's inter-echo covariance matrix before Cholesky factorization; if white-matter NRMSE stays at 0.168 rather than moving toward the 0.173 of the diagonal baseline, the specific covariance structure is not carrying the reported benefit.","tokens_in":7524,"feed_emoji":"🧠","tokens_out":10447,"duration_ms":98274,"temperature":0.7,"pith_summary":"Accelerated T2* mapping is clinically useful only if reconstruction artifacts do not corrupt the fitted decay parameter, and this paper's claim is that the fix is to carry the full inter-echo covariance forward rather than a per-voxel scalar variance. CUPA-T2* does this in three connected stages: a Monte Carlo dropout reconstruction produces many stochastic versions of the images, a per-voxel covariance matrix over the 12 echoes is factored by Cholesky decomposition and used to sample uncertainty-aware inputs, and a heteroscedastic network trained with a tissue-stratified correlation regularizer predicts both T2* and its variance. The experiments show the framework stays competitive overall on brain data while achieving the lowest white-matter error at accelerations R=3 and R=4, and it makes predicted T2* variance track reconstruction uncertainty far more closely than a heteroscedastic baseline (Pearson correlation 0.65-0.77 versus 0.19-0.28). The same results expose a trade-off: calibration error and area under the risk-coverage curve worsen as alignment increases. If the result holds, accelerated T2* protocols could ship with voxel-wise uncertainty maps that point at the regions most likely to be wrong.","feed_headline":"Inter-echo covariance propagation improves white-matter T2* maps","feed_subtitle":"CUPA-T2* carries correlated echo uncertainty into fitting, improving white-matter T2* maps at R=3 and R=4.","key_machinery":"The central mechanism is a per-voxel inter-echo covariance matrix $\\Sigma_v$, estimated from $S=100$ Monte Carlo dropout reconstructions over the $2E$ real and imaginary components of $E=12$ echoes and factorized as $\\Sigma_v \\approx L_v L_v^T$ via Cholesky decomposition. It enters the regression through the sampling rule $\\mathbf{x}_v^{(s)} = \\mu_v + \\alpha_v L_v \\mathbf{z}^{(s)}$ with $\\mathbf{z}^{(s)} \\sim \\mathcal{N}(0, I)$, where the scalar $\\alpha_v$ grows with the RMS marginal standard deviation $u_v = \\sqrt{\\frac{1}{2E}\\mathrm{tr}(\\Sigma_v)}$, so clean voxels receive almost no injected noise while noisy voxels receive correlated echo-to-echo variation. The second half of the machinery is a correlation-based regularizer: the Pearson correlation between predicted log-variance and $\\log u_v$, averaged with equal weights over white matter, gray matter, and CSF, is added to a $\\beta$-NLL fitting loss as $\\mathcal{L} = \\mathcal{L}_{\\beta\\text{-NLL}} + \\lambda(1-\\rho)$. This regularizer is what forces the predicted aleatoric variance to line up with the upstream reconstruction uncertainty, and the covariance-aware samples are what feed that alignment meaningful structure.","core_discovery":"On its own terms, the paper establishes that structured inter-echo uncertainty from undersampled MRI reconstruction can be propagated into T2* regression without collapsing it to a scalar, and that this changes where the fitted parameter is accurate. The load-bearing result is that covariance-aware sampling improves white-matter T2* accuracy at R=3 and R=4, while the RMS correlation regularizer raises the Pearson correlation between reconstruction-derived uncertainty and predicted T2* variance from roughly 0.19-0.28 with a heteroscedastic baseline to 0.65-0.77 with CUPA-T2*. The paper attributes the tissue-dependence to the reliability of the estimated covariance: white matter is homogeneous with fewer physiological fluctuations, whereas high-uncertainty CSF has less coherent inter-echo structure. It also reports an explicit trade-off, since the alignment objective that maximizes correlation degrades ECE and AURC, and the ablation at R=4 shows that the covariance structure and the regularizer each contribute separately.","pith_inferences":["A testable extension the paper does not run: if shared undersampling is what creates the inter-echo covariance, then independently drawn sampling masks per echo should shrink or eliminate CUPA-T2*'s white-matter advantage; this can be checked by simulation before any clinical study.","Because the regularizer uses per-tissue Pearson correlation on log-variances, it rewards monotone association but not calibrated scale; a rank-based or quantile-matching alignment might retain most of the rho gain with less ECE penalty.","The dependence on $S=100$ Monte Carlo samples suggests a sensitivity check: repeating the R=4 experiment with $S=20$ and $S=200$ would reveal whether the benefit comes from covariance structure itself or simply from more ensemble averaging.","CSF performance degrades at R=4 under the current uncertainty-based alpha schedule, so explicitly conditioning $\\alpha_v$ on tissue type in addition to $u_v$ is a natural way to recover high-uncertainty regions without losing the white-matter gain."],"forward_implications":["White-matter T2* mapping, which matters for iron-related lesion assessment, becomes accurate at R=3 and R=4 while total scan time falls to roughly 55 seconds at R=4.","Predicted T2* variance now highlights the same voxels where reconstruction artifacts concentrate, so the uncertainty map can serve as a per-voxel trust signal when reading accelerated maps.","Deployment must treat calibration and selective prediction as competing objectives: with the alignment weight at 0.05, the correlation gain comes at the cost of higher ECE and AURC, so applications that need calibrated confidence should lower $\\lambda$.","The same covariance-aware propagation and alignment recipe should transfer to other quantitative parameters whose echoes or contrasts share undersampling-induced correlations, as the paper states as future work."],"supporting_citations":[{"why":"Supplies the public T2*-MOVE brain k-space dataset and fully sampled reference reconstructions used for training, validation, and testing.","marker":"[4]"},{"why":"Provides the unrolled multi-echo reconstruction architecture with CNN denoising and data consistency that CUPA-T2* adopts.","marker":"[5]"},{"why":"Grounds Monte Carlo dropout as a Bayesian approximation, the mechanism that generates the S=100 stochastic reconstructions.","marker":"[6]"},{"why":"Defines the PUQ baseline and the earlier phase-wise uncertainty guidance that CUPA-T2* extends with covariance and alignment.","marker":"[7]"},{"why":"Supplies the beta-NLL loss with beta=0.5 that trains the heteroscedastic regressor and stabilizes variance learning.","marker":"[10]"},{"why":"Defines the heteroscedastic baseline and the aleatoric/epistemic decomposition used for predictive uncertainty.","marker":"[2]"},{"why":"Supports MC-dropout-based epistemic uncertainty estimation inside the heteroscedastic MLP and the total uncertainty decomposition.","marker":"[15]"}],"fun_headline_variants":["Covariance-aware uncertainty boosts white-matter T2* maps","Propagating echo covariance improves T2* in white matter","CUPA-T2* sharpens white-matter T2* via uncertainty alignment","T2* mapping gains from covariance propagation, with a trade-off"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a 24-by-24 per-voxel covariance matrix estimated from only 100 stochastic reconstructions reflects genuine inter-echo correlation structure rather than estimation noise, so Cholesky sampling injects useful variability instead of corrupting the fit.","fun_headline_variants_meta":{"raw":{"variants":["Covariance-aware uncertainty boosts white-matter T2* maps","Propagating echo covariance improves T2* in white matter","CUPA-T2* sharpens white-matter T2* via uncertainty alignment","T2* mapping gains from covariance propagation, with a trade-off"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1273,"prompt_tokens":978,"completion_tokens":295,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":218}},"tokens_in":594,"tokens_out":295,"duration_ms":3676,"temperature":1.0,"reasoning_tokens":218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:26:32.660040+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the T2*-MOVE test set at R=4, randomly permute the off-diagonal entries of every voxel's inter-echo covariance matrix before Cholesky factorization; if white-matter NRMSE stays at 0.168 rather than moving toward the 0.173 of the diagonal baseline, the specific covariance structure is not carrying the reported benefit.","supporting_citations":[{"cited_title":"et al.: T2*-MOVE: An Open k-Space Dataset to Address Motion in T2* Quantification from Gradient Echo MRI","cited_arxiv_id":null,"evidence_quote":"Supplies the public T2*-MOVE brain k-space dataset and fully sampled reference reconstructions used for training, validation, and testing."},{"cited_title":"Magnetic Resonance in Medicine95, 346–362 (2026)","cited_arxiv_id":null,"evidence_quote":"Provides the unrolled multi-echo reconstruction architecture with CNN denoising and data consistency that CUPA-T2* adopts."},{"cited_title":"In: Proceedings of the 33rd International Conference on Machine Learning (ICML), PMLR 48, 1050–1059 (2016) 10 G","cited_arxiv_id":null,"evidence_quote":"Grounds Monte Carlo dropout as a Bayesian approximation, the mechanism that generates the S=100 stochastic reconstructions."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp","cited_arxiv_id":null,"evidence_quote":"Defines the PUQ baseline and the earlier phase-wise uncertainty guidance that CUPA-T2* extends with covariance and alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the heteroscedastic baseline and the aleatoric/epistemic decomposition used for predictive uncertainty."},{"cited_title":"Uncertainty Quantification in Deep Learning for Safer Neuroimage Enhancement","cited_arxiv_id":"1907.13418","evidence_quote":"Supports MC-dropout-based epistemic uncertainty estimation inside the heteroscedastic MLP and the total uncertainty decomposition."}],"review_version":1}