{"id":"db588fb3-ab59-42ec-ba30-317250e6c4b6","arxiv_id":"2601.13236","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Conformalized quantile regression gives pixel-wise uncertainty intervals for accelerated MRI reconstructions that closely track true reconstruction error, with brain test correlations near or above 0.90 at 4–10x acceleration.","lead":"This paper adds a second neural network that marks which pixels in a fast MRI reconstruction are unreliable, using statistically calibrated uncertainty intervals that need no ground-truth image at scan time. The authors show these pixel-wise maps track actual reconstruction error better than a simpler residual-based heuristic in brain and knee scans.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported ~90% Pearson correlation may be inflated by un-masked background pixels; need a masked re-analysis.","rationale":"The reader's weakest assumption concerns calibration/deployment exchangeability, which is a valid external-validity limitation. However, the most load-bearing internal concern is that the headline correlation, the paper's primary evidence, may be inflated by the inclusion of background pixels in the correlation computation. This is concrete, testable, and directly affects the strongest claim. If background masking destroys the correlation, the claim that QR uncertainty is a faithful surrogate for reconstruction error loses its quantitative support, leaving only qualitative examples. The paper should be accepted only if the masked correlation remains high, hence CONDITIONAL. I partially agree with the reader because both concerns question the reliability of the headline, but they target different stages: the reader's concern is about applying the method to a different distribution; mine is about the validity of the reported metric on the tested distribution.","tokens_in":12344,"tokens_out":7812,"duration_ms":88177,"concrete_test":"Recompute the per-image Pearson and Spearman correlations for the 4x brain test set (Table II) after masking out background pixels: e.g., threshold at 5% of the maximum image intensity or use an anatomical brain mask. Compare the masked correlations with the reported QR values (Pearson 0.91 ± 0.10, Region-Pearson 0.96 ± 0.11). If the masked Pearson drops below ~0.7, the central claim is substantially weakened; if it remains above ~0.85, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim (Section IV: 'QR-based uncertainty show high Pearson correlations (roughly 90%)...') rests on correlations computed from full images without any explicit background mask. Section II-E defines Pearson and Spearman correlations over 'all pixel values within the image', and the pre-processing is only a Gaussian blur (sigma=2), not a foreground mask. FastMRI brain images contain large background regions where the reconstruction error and the predicted uncertainty are both near zero, forming a dominant cluster at (0,0). Including such a cluster can inflate Pearson (and Spearman) correlations substantially, even if uncertainty tracks error poorly inside the anatomy. This is not a calibration/coverage issue; it is a direct threat to the empirical evidence for the claim that calibrated interval width is a per-pixel proxy for reconstruction error in clinically relevant regions. The qualitative lesion-localization results (Figures 4–6) are suggestive, but the headline 90% correlation may be an artifact of the evaluation metric rather than a genuine property of the uncertainty maps in tissue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conformal quantile regression framework for pixelwise uncertainty quantification in accelerated MRI reconstruction. Two U-Nets parameterize lower and upper quantile bounds of the ground-truth image conditioned on the output of a fixed E2E VarNet reconstruction; the bounds are trained with the pinball loss, and a scalar conformal factor is calibrated on a held-out split. The method is evaluated on fastMRI brain and knee data at accelerations 2x to 10x, with the primary quantitative claim being high Pearson correlation (roughly 90% for brain at 4x and above) between the calibrated interval width and the true absolute reconstruction error, plus qualitative evidence that uncertainty highlights lesions and artifacts.","tokens_in":12619,"tokens_out":8267,"duration_ms":87023,"significance":"If the quantitative claims are established, the framework addresses a clinically important problem: providing per-pixel confidence estimates for accelerated MRI reconstructions without access to a fully sampled reference at inference. The study has methodological strengths: separate training/validation/calibration/test splits, comparison against a residual-magnitude baseline, evaluation across multiple acceleration factors and two anatomies, and qualitative lesion-localization evidence. The use of conformal calibration with a concentration bound is appropriate in spirit. However, the current manuscript does not report empirical test-set coverage, the correlation metric is computed on unmasked, Gaussian-blurred images, and the definition of the calibration target is internally inconsistent. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The notation for α is internally inconsistent. The text defines the upper/lower quantiles as (1−α/2) and α/2, which is standard when α is the miscoverage rate (e.g., α=0.1 for 90% coverage). However, the text then states 'We set α = 90%', and in II-D4 the calibration target is that the average fraction of pixels outside the interval 'does not exceed α'. If α=0.9, the trained interval is only the central 10% interval and the calibration would allow 90% miscoverage. If the intended coverage is 90%, the value of α and the formulas must be corrected (e.g., α=0.1, upper=1−α/2, lower=α/2). As written, the statistical target of the method is not well defined.","section":"Section II-D2 and II-D4"},{"comment":"No empirical coverage is reported on the test set. The paper claims conformal calibration yields statistically rigorous, finite-sample valid pixelwise intervals, and λ is chosen using a Hoeffding bound on a calibration set, but nowhere is the achieved coverage (e.g., the fraction of ground-truth pixels inside T_λ) reported for the held-out test data. Since coverage is the primary statistical guarantee, a new table should report pooled and per-image empirical coverage for each acceleration, anatomy, and method, ideally stratified by foreground/background. Without this, the 'statistically rigorous' claim is unsupported.","section":"Section II-D4 / Table II"},{"comment":"The headline correlations may be inflated by background pixels. Pearson/Spearman correlations are computed over 'all pixel values within the image' after a Gaussian blur with σ=2; no foreground or tissue mask is used. FastMRI brain and knee images contain large regions of near-zero background where both the reconstruction error and the predicted uncertainty are close to zero, producing a dominant cluster at (0,0) that can raise Pearson (and Spearman) correlations even if uncertainty tracks error poorly inside the anatomy. The region-based correlations are also not masked. The authors should recompute correlations using a foreground/anatomy mask and, as a sensitivity check, without Gaussian blur, and report whether the ~90% claim survives.","section":"Section II-E, Eq. (6), Table II"},{"comment":"The finite-sample guarantee relies on a Hoeffding bound applied to 'the average fraction of all pixels' in the calibration set. Pixel residuals are not independent: they are spatially correlated within an image and across slices. If the bound is computed over pooled pixels, plain Hoeffding's inequality is not valid. The risk definition, the unit of exchangeability (pixels vs. images), and the concentration inequality need to be specified precisely. A valid approach would define a per-image risk and apply the bound over the I calibration images, or use a dependence-aware bound. As written, the calibration step may not provide the claimed distribution-free coverage guarantee.","section":"Section II-D4"}],"minor_comments":[{"comment":"The statement that 'Pearson correlation coefficient was higher than 90%' applies only to brain; Table II shows knee Pearson correlations of 0.80±0.16 and 0.82±0.15. Please qualify this claim as brain-specific in the abstract.","section":"Abstract"},{"comment":"The term 'pixelwise' should be used carefully. The scalar conformal factor provides marginal coverage over the pooled pixel population, not per-pixel conditional coverage. The authors should clarify this distinction in the text to avoid overclaiming.","section":"Section II-D2/D4"},{"comment":"For the region-based correlations, specify how patches that are entirely background or constant are handled, and state whether ties in Spearman rank computation (many zero-valued pixels) are treated consistently.","section":"Section II-E"},{"comment":"Clarify which method the 'ResM (%)' columns refer to and why ResM is only reported at 4x; also explicitly define the normalization by maximum reconstruction magnitude in the table caption.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid engineering paper that extends conformalized quantile regression to pixelwise MRI UQ and evaluates it across fastMRI brain and knee at 2–10x. The main claim—calibrated QR interval width tracks reconstruction error with ~90% Pearson correlation at 4x and above—is probably inflated by including background pixels. Section II-E computes correlations over all pixels after Gaussian blur with no foreground mask. Brain MR images have large background regions where both error and uncertainty are near zero, so the correlation partly measures the trivial agreement of zeros. The region-based correlations (0.96–0.97) are even higher, but that doesn't fix the issue unless the patches are tissue-only; the paper doesn't say that. The qualitative figures do show genuine spatial alignment with lesions, which is encouraging, but those are selected examples.\n\nWhat is genuinely useful: the offset-scaling calibration is a sensible adaptation of CQR to this architecture; the training/calibration/test split is clean; the comparison against residual magnitude is fair and shows QR is better; the discussion of conformal limitations—in-distribution calibration sets and the small knee calibration set—is honest. The method is an extension rather than a new paradigm, but for clinicians wanting a per-pixel reliability map without a reference, this is a relevant step.\n\nSoft spots, in order: First, the background-mask issue threatens the headline number. Second, no empirical test-set coverage is reported. For a conformal method, coverage is the headline guarantee; the paper gives calibration lambdas and a Hoeffding bound but never says what coverage actually was on the test set. Third, no code or data release, so reproducibility is limited. Fourth, the fact that the quantile U-Nets are trained on ground truth and then evaluated by correlation with error computed from the same ground truth is not fatal—it's supervised UQ—but it means the reported correlation partly measures how well the module learned its training target. The clinically interesting question is out-of-distribution transfer, which the paper doesn't test.\n\nI'd send this to peer review rather than desk reject: the application is important, the evaluation is broad, and the weaknesses are fixable. Ask for a foreground-masked correlation analysis, test-set coverage numbers, and code or implementation details. If the masked correlation collapses, the paper's clinical claim weakens substantially; if it holds, this is a good contribution to MRI UQ.","headline":"Useful pixelwise conformal UQ for accelerated MRI, but the headline ~90% correlation likely leans on background pixels; needs a masked re-analysis before the central claim is trusted.","tokens_in":13086,"tokens_out":2437,"would_cite":true,"duration_ms":29860,"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":"Conformalized quantile regression produces pixelwise uncertainty intervals for accelerated MRI reconstruction that track the true reconstruction error with roughly 90% linear correlation at 4x acceleration and above, without any ground-trut","keywords":["conformal prediction","quantile regression","uncertainty quantification","accelerated MRI","parallel imaging","pixelwise confidence intervals","reconstruction error","anomaly detection"],"falsifier":"Take the calibrated module and evaluate it on brain or knee images from a different scanner or with a pathology distribution not represented in the calibration set; if the linear correlation between interval width and true reconstruction error falls to the level of the residual-magnitude baseline (below 70%, or below 25% with a small calibration set), the central claim that QR uncertainty is a faithful, portable surrogate for reconstruction error is refuted. A simpler internal check: use only 50 calibration volumes on brain data and see if the correlation collapses as it does for the knee.","tokens_in":12283,"feed_emoji":"🧠","tokens_out":5940,"duration_ms":60701,"temperature":0.7,"pith_summary":"The paper tries to establish that pixelwise uncertainty intervals computed by conformalized quantile regression, wrapped around an existing accelerated MRI reconstruction network, can serve as a faithful surrogate for the true reconstruction error even when fully sampled ground-truth images are unavailable. Across brain and knee scans at acceleration factors from 2 to 10, the calibrated interval width correlates with the actual error at a linear correlation above 90% for brain images at 4x and higher, whereas a simpler residual-magnitude heuristic falls below 70% (and below 25% for knees with a small calibration set). If correct, this gives clinicians and automated protocols a per-pixel confidence signal that flags unreliable regions, lesions, and artifacts without slowing the scan or needing a reference image. The framework is an add-on to any reconstruction network and carries finite-sample coverage guarantees from conformal calibration.","feed_headline":"Conformal quantile regression flags MRI errors without a reference","feed_subtitle":"Calibrated pixelwise intervals flag lesions and artifacts in accelerated scans, no ground truth needed.","key_machinery":"The engine is conformalized quantile regression (CQR). Two U-Nets produce pixelwise lower and upper quantile estimates of the reconstruction; they are trained with the pinball loss against ground-truth images. A conformal calibration step then finds the smallest global scaling factor lambda that rescales the quantile offsets so that, on a held-out calibration set, the fraction of pixels whose ground truth falls outside the interval is controlled at the target level (alpha = 0.1), using a conservative finite-sample bound on the coverage. The calibrated interval width is the uncertainty map. CQR is what converts heuristic network outputs into intervals with a statistical guarantee.","core_discovery":"The paper claims that pixelwise uncertainty intervals produced by conformalized quantile regression—two U-Nets trained with the pinball loss to predict lower and upper conditional quantiles of the reconstruction, followed by conformal calibration of a single scaling factor on a holdout set—yield a statistically valid and spatially informative uncertainty map for accelerated MRI reconstruction. After calibration, the width of these intervals correlates with the true reconstruction error at a linear correlation above 90% for brain images at 4x acceleration and higher, while the residual-magnitude heuristic stays below 70% (and below 25% for knee images with a smaller calibration set). The unce","pith_inferences":["If the ~90% correlation persists across scanner vendors and pathology mixes not seen in training, the approach could let a radiologist set a per-scan 'trust threshold' and automatically flag slices for re-acquisition; the paper does not test this transfer.","The single global scaling factor lambda rescales all pixels uniformly; a spatially adaptive calibration (per-slice or per-region lambda) might sharpen the already strong local correlations the paper reports in region-based metrics, an extension the paper leaves open.","A concrete testable extension is to run a 10x pre-scan, measure the volume of high-uncertainty regions, and compare it with the lesion volume from a 4x diagnostic scan to quantify sensitivity and specificity as an anomaly detector.","Because the uncertainty module is trained after the reconstruction network is frozen, jointly fine-tuning the reconstruction network to reduce high-uncertainty pixels could turn the uncertainty map from a passive indicator into an active training signal; the paper lists this as future work."],"forward_implications":["At 4x acceleration and above, the calibrated interval width acts as a per-pixel proxy for reconstruction error without needing a ground-truth reference at inference.","The method transfers across brain and knee anatomies, multiple contrasts, and acceleration factors up to 10x, and the uncertainty maps spatially align with lesions and artifacts.","Regions where uncertainty exceeds a threshold derived from the 4x-acceleration maximum could flag slices that should not be read diagnostically or that need re-acquisition.","Because the uncertainty module is an add-on trained with the reconstruction network frozen, it can be attached to any reconstruction method without retraining the base model.","The framework enables a fast, highly accelerated pre-scan to act as an anomaly detector, potentially deciding when a longer diagnostic-quality scan is necessary."],"fun_headline_variants":["Pixelwise MRI error flags without reference images","Conformal quantile regression maps MRI reconstruction risk","No-ground-truth MRI uncertainty beats residual heuristic","Calibrated intervals localize MRI artifacts and lesions","Adaptive MRI scanning via statistical uncertainty maps"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole guarantee rests on the calibration set being large enough and representative of the scans the method will later see; if the scanner, patient population, or undersampling pattern shifts, the calibrated intervals and the ~90% correlation need not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Pixelwise MRI error flags without reference images","Conformal quantile regression maps MRI reconstruction risk","No-ground-truth MRI uncertainty beats residual heuristic","Calibrated intervals localize MRI artifacts and lesions","Adaptive MRI scanning via statistical uncertainty maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1259,"prompt_tokens":776,"completion_tokens":483,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":413}},"tokens_in":520,"tokens_out":483,"duration_ms":5472,"temperature":1.0,"reasoning_tokens":413,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:35:41.945735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the calibrated module and evaluate it on brain or knee images from a different scanner or with a pathology distribution not represented in the calibration set; if the linear correlation between interval width and true reconstruction error falls to the level of the residual-magnitude baseline (below 70%, or below 25% with a small calibration set), the central claim that QR uncertainty is a faithful, portable surrogate for reconstruction error is refuted. A simpler internal check: use only 50 calibration volumes on brain data and see if the correlation collapses as it does for the knee.","supporting_citations":[],"review_version":1}