{"id":"fb9fd8eb-1b89-4d6d-af87-a6a4c66e5b7b","arxiv_id":"2506.06043","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding TV regularization to coil sensitivity maps in implicit neural representation MRI reconstruction improves image quality, especially with limited calibration lines.","lead":"MRI scanning is slow, so researchers reconstruct images from fewer collected measurements. This paper adds a smoothness penalty on coil sensitivity maps inside an implicit neural representation method, improving reconstruction quality and robustness when calibration data is scarce.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sensitivity-map accuracy is asserted but never measured; the TV term's benefit may act through a different mechanism than claimed.","rationale":"The reader's weakest assumption correctly identifies the scale ambiguity and the missing quantitative validation of sensitivity maps. I agree with that assessment. My stress-test adds a concrete consequence of the ambiguity: because the data term is invariant to the scaling (S_j, X) → (S_j/c, cX) and the TV penalty on S is not scale-invariant, the learned Sβ is only determined up to a gauge that depends on λ1/λ2. Without a normalization step or a sensitivity error metric, the claimed mechanism (more accurate sensitivities) is unverified. However, this does not invalidate the paper's primary reconstruction-quality claims, which are supported by the ablation and comparative experiments. The missing sensitivity metric should be added as a condition for full acceptance, but it does not change the conditional verdict. No further objection is raised. The proposed concrete test directly targets the uncertain link between the regularization and the sensitivity accuracy and would settle whether the central mechanism actually lands.","tokens_in":13101,"tokens_out":5383,"duration_ms":61397,"concrete_test":"On Dataset III at ACS=8, R=5 and ACS=24, R=5, compute the relative L2 error (or mean absolute error) between the estimated sensitivity maps and the paper's own ground-truth sensitivities (per-coil division by SOS), after removing the global multiplicative ambiguity by normalizing each estimated S_j to the same RMS (or max-abs) as the corresponding ground-truth S_j over the object support. Compare INR-CRISTAL with and without sensitivity TV, and include ESPIRiT as a reference. If sensitivity error does not improve with the TV term while PSNR improves, the mechanism claimed in Eqs. (6)/(9) is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Eqs. 6 and 9) is that adding the sensitivity-map regularization λ2·R(Sβ) yields more accurate coil sensitivity estimates and thereby better reconstruction. Yet the paper never reports a quantitative sensitivity error; Figs. 1 and 3 are visual only. This matters because the forward model is invariant under the rescaling (S_j, X) → (S_j/c, cX). The TV penalty on Sβ is scale-sensitive (TV(cS)=|c|·TV(S)), so the relative strengths λ1 and λ2 select a gauge rather than necessarily the physically correct sensitivities. Without normalization or a sensitivity error metric, the observed PSNR gains could arise from the TV term acting as an indirect regularizer on the image through the gauge, not from more accurate sensitivity estimates. The paper also does not report the values of λ1 and λ2 (or their selection), and the ablation is limited to one dataset (Dataset III) at two ACS/R settings, making it difficult to verify robustness of the mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes INR-CRISTAL, a scan-specific implicit neural representation method for parallel MRI reconstruction that jointly estimates the image and coil sensitivity maps and adds an explicit regularization term on the sensitivity maps, with total variation (TV) recommended. The method is evaluated against GRAPPA, pFISTA-SENSE, AC-LORAKS, NERP, and IMJENSE on four datasets, under varying ACS lines and acceleration factors, with ablations comparing L1, low-rank, and TV sensitivity regularization. The central claim is that adding the sensitivity-map regularization, especially with TV, yields more accurate coil sensitivity estimates and better reconstruction quality, particularly when ACS data are limited.","tokens_in":13317,"tokens_out":3072,"duration_ms":34492,"significance":"If the central claim holds, the contribution is a simple and potentially useful extension of INR-based parallel MRI: imposing a physically motivated smoothness constraint on the jointly estimated sensitivity maps improves reconstruction robustness to limited ACS and high acceleration. The paper provides consistent comparative results across four datasets, a Wilcoxon signed-rank analysis on Dataset III, and an ablation table showing consistent PSNR/SSIM gains from the sensitivity constraint. The novelty is incremental—sensitivity-map regularization has been used in optimization-based parallel imaging—but applying it inside a joint INR framework with a full comparison is a reasonable contribution. However, the claimed mechanism (more accurate sensitivity maps) is not directly measured, and several implementation details are missing, which limits the strength of the conclusions.","major_comments":[{"comment":"The load-bearing claim that adding λ2·R(Sβ) leads to 'more accurate coil sensitivity estimates' is not quantitatively verified. Sensitivity maps are only inspected visually, and no sensitivity error metric (e.g., relative L2 error against reference maps after gauge alignment) is reported. This matters because the forward model is invariant under the rescaling (S_j, X) → (S_j/c, cX), while TV(cS) = |c|·TV(S), so the penalty strength selects a gauge rather than necessarily selecting the physically correct sensitivities. The observed PSNR gains could therefore arise from indirect image regularization through the gauge, not from improved sensitivity estimates. Please report a quantitative sensitivity-map error, describe how the gauge ambiguity is handled, and provide the values of λ1 and λ2 used in the reported experiments.","section":"Section III-C, Eq. (6), Figs. 1 and 3"},{"comment":"The regularization type (TV) is chosen after comparing L1, low-rank, and TV on the same datasets that are later used for the final performance evaluation. This creates a selection-on-test-data risk: the reported gains of INR-CRISTAL over the ablated variants may partly reflect fitting the regularization choice to the evaluation set. The authors should either use a held-out validation set for the regularization selection or present results for all three regularization types on all datasets so the reader can assess the variance of the comparison.","section":"Section IV-D"},{"comment":"The ablation evidence for the central mechanism is thin: Table II reports only single reconstruction values on Dataset III at two ACS settings and two acceleration factors, with no error bars, no multiple slices, and no statistical test. The claim that the sensitivity constraint is 'crucial' under limited ACS rests on essentially two numbers (ACS=8, R=5 and R=6). Please extend the ablation to at least one additional dataset and report mean±std over slices, or otherwise temper the strength of the conclusion.","section":"Section V-C, Table II"},{"comment":"Key hyperparameters and their selection procedure are not reported. The paper states that Bayesian optimization was used to tune parameters for all methods, but does not give the ranges or final values for λ1, λ2, the Fourier feature variance σ², or the SIREN frequency parameter w0. Without these values, the comparison is not reproducible, and it is unclear whether the advantage of INR-CRISTAL is robust to reasonable hyperparameter variation. Please report these settings or provide the code/configuration used.","section":"Section IV-B and IV-C"}],"minor_comments":[{"comment":"The data consistency term is defined with a squared l2 norm in Eq. (2) but as an l1 norm in Eq. (10). Please unify the notation and clarify which loss is actually minimized in Eq. (9).","section":"Eq. (2) vs Eq. (10)"},{"comment":"The operator U is described in a confusing and tautological way: 'performing zero-filling at the positions where U samples U, and performing zero-filling at the positions where U samples U'. Please rewrite this definition clearly.","section":"Eq. (13)"},{"comment":"The labels R_L1, R_LR, and R_TV are used in the figure but not defined in the caption; please define them explicitly.","section":"Fig. 3 caption"},{"comment":"There are several typos and grammatical errors, e.g., 'with much few data' in the abstract, 'Faulty of Applied Science' in the affiliations, and 'supported by supported in part' in the funding footnote. A language edit is recommended.","section":"Abstract and affiliations"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a straightforward extension of IMJENSE/CineJENSE with an additional sensitivity-map penalty. The main technical risk is that the claimed mechanism—improved sensitivity-map accuracy—is not measured, and the gauge ambiguity is not addressed. The missing hyperparameter reporting and test-set-based regularization selection also need attention. I would be open to accepting after a revision that addresses these points quantitatively."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper adds a total-variation penalty on coil sensitivity maps to the joint INR image/sensitivity reconstruction objective, and shows that this helps: the ablation in Table II is consistent across ACS and R settings, the gains are real if modest (roughly 0.2 to 2.5 dB), and the comparisons against GRAPPA, IMJENSE, pFISTA-SENSE, and others on four datasets are thorough. The robustness curves in Figs. 8 and 9 and the ACS-free experiments in Fig. 10 are a nice addition. This is an incremental but honestly executed empirical extension of IMJENSE/CineJENSE, and the claim that it is the first to apply TV directly to sensitivity maps in this setting seems plausible.\n\nThe soft spots are real but not fatal. The biggest one: the paper says the regularization gives 'more accurate coil sensitivity estimates,' but it never measures sensitivity error. Figures 1 and 3 are visual only. Given the gauge ambiguity in the forward model — you can rescale S and X inversely and leave the data term unchanged — the TV penalty could be acting partly by selecting a particular gauge rather than by driving S toward the true physical sensitivities. The image TV term also shifts under that rescaling, so the two regularizers jointly set a scale. Without a sensitivity error metric or at least a normalization strategy, the claimed mechanism is unverified. The reconstruction gains are measured, so the method stands on that, but the explanation is not proven.\n\nSecond, the values of λ1 and λ2 are not reported anywhere, which makes exact reproduction difficult. Third, the choice of TV over L1 and low-rank is made after looking at results on Dataset III, and the ablation in Table II is only on Dataset III. That is a mild selection-on-evaluation concern, not a fatal one, but it should be stated. No code release makes it harder to check the sensitivity maps directly.\n\nOverall, the central empirical claim holds up: adding the sensitivity regularizer improves reconstruction quality, especially at low ACS. The mechanism is under-supported, but the paper is clear about what it did and the results are consistent. It deserves a serious referee; the right response is a major revision asking for sensitivity error metrics, reported λ values, and ideally an analysis of the gauge issue. I would bring it to a reading group and would cite it if I worked on INR-based MRI, mainly as a reference for sensitivity regularization.\n\nRecommendation: send to peer review. The flaws are fixable and the core finding is useful.","headline":"A simple, sensible extension — TV regularization on coil sensitivities in joint INR reconstruction — with consistent but modest gains; the mechanism is asserted rather than measured.","tokens_in":13820,"tokens_out":1291,"would_cite":true,"duration_ms":15940,"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":"Adding a smoothness penalty on coil sensitivity maps, not just on the image, is what keeps implicit-neural-representation MRI reconstruction accurate when calibration data is scarce.","keywords":["implicit neural representation","MRI reconstruction","coil sensitivity map","total variation regularization","parallel imaging","SIREN","scan-specific reconstruction","undersampled k-space"],"falsifier":"Compute ground-truth sensitivities from fully sampled multi-coil data (coil image divided by the sum-of-squares image) and, under a fixed mask such as ACS=8, R=5, run INR-CRISTAL with and without the sensitivity TV term, reporting the relative $\\ell^2$ error of the estimated sensitivity maps. If the regularized run does not reduce sensitivity error relative to the unregularized run — while still improving image metrics — the paper's stated mechanism fails; if the sensitivity error falls together with reconstruction error, the mechanism is confirmed. A phantom with a known sharp sensitivity edge would also test whether the TV prior discards genuine sensitivity structure.","tokens_in":1961,"feed_emoji":"🧲","tokens_out":4475,"duration_ms":120495,"temperature":0.7,"pith_summary":"The paper claims that the limiting factor in highly accelerated parallel MRI reconstruction with implicit neural representations is not the image model but the estimated coil sensitivity maps, and that this can be fixed by regularizing the sensitivities themselves. It proposes INR-CRISTAL, a scan-specific method with two SIREN networks that jointly estimate a composite image and per-coil sensitivity maps, trained by data consistency plus total-variation penalties on both the image and the sensitivity maps. Adding the sensitivity penalty, the paper argues, removes undersampling artifacts that otherwise accumulate in INR-generated sensitivities and propagate into the reconstruction, which matters most when auto-calibration signal (ACS) lines are scarce. On brain and knee datasets the method reports the lowest reconstruction error against GRAPPA, pFISTA-SENSE, AC-LORAKS, NERP, and IMJENSE, and keeps its advantage as ACS shrinks and the acceleration factor rises.","feed_headline":"TV-smooth coil maps sharpen undersampled MRI","feed_subtitle":"A smoothness prior on coil sensitivities removes artifacts when calibration lines are scarce, brain and knee tests show.","key_machinery":"The machinery is a pair of SIREN (sine-activation) multilayer perceptrons, both fed with Fourier-feature-embedded image coordinates $\\theta$: one produces the composite image $X_\\alpha(\\theta)$, the other the per-coil sensitivity maps $S_\\beta(\\theta)$. The training objective (Eq. 9) is $L_{\\mathrm{total}} = L_{DC} + \\lambda_1 L_I + \\lambda_2 L_S$, where $L_{DC}$ enforces consistency between the forward model $\\mathcal{U}\\mathcal{F} S_{\\beta j} \\odot X_\\alpha$ and the sampled k-space data $Y_j$, $L_I$ is the image total variation, and $L_S = \\|\\mathbf{G}(S_\\beta)\\|_1$ is the sensitivity total variation — the load-bearing new term. At inference, a combined k-space step (Eq. 13) replaces predicted samples with the actually acquired ones before the inverse Fourier transform. The TV penalty carries the argument: it encodes the physical prior that coil sensitivities are spatially smooth, and the paper motivates its choice by comparing against Fourier-sparsity and low-rank sensitivity priors.","core_discovery":"INR-CRISTAL's central claim is that explicit regularization of the coil sensitivity maps — specifically a total-variation penalty on the sensitivity network's output — is what separates a working scan-specific parallel-imaging reconstruction from a degraded one under aggressive undersampling. The paper formulates the task as a joint optimization over two implicit networks, the image network $X_\\alpha$ and the sensitivity network $S_\\beta$ (Eqs. 6, 8, 9), adding the term $\\lambda_2 R(S_\\beta)$ on top of data consistency and image total variation. The authors state the conclusion directly: 'INR-CRISTAL provides more accurate coil sensitivity estimates with fewer artifacts, and delivers superior reconstruction performance in terms of artifact removal and structure preservation.' Among candidate constraints — Fourier-domain $\\ell^1$ sparsity, nuclear-norm low-rankness, and total variation — the paper finds TV gives the best trade-off of artifact suppression, fidelity, and runtime, and reports this as the first application of a TV constraint to sensitivity maps in this setting. The same formulation also reconstructs without any auto-calibration signal, beating SAKE, P-LORAKS, and IMJENSE under Gaussian undersampling.","pith_inferences":["The paper ties image-quality gains implicitly to sensitivity accuracy, yet reports no direct sensitivity error metric; a quantitative test of whether the TV-regularized sensitivities are literally closer to ground truth would confirm or break the stated mechanism.","Because image-times-sensitivity decompositions are ambiguous up to rescaling, the TV prior selects one member of that equivalence class; normalizing sensitivities to unit sum-of-squares and checking whether TV still helps would separate the smoothness prior from magnitude calibration.","The paper's own stated bottleneck is reconstruction time (two SIREN networks, 1000 iterations); a practical follow-up would warm-start from a low-resolution run or distill the TV-regularized sensitivity maps into a lightweight network to make the method clinically usable."],"forward_implications":["With limited ACS lines, the sensitivity TV penalty buys nearly 1.8 dB of PSNR at R=5, ACS=8 on the brain ablation and more than 2 dB at R=6, turning a failed reconstruction into a usable one.","INR-CRISTAL reports the lowest RLNE across all tested ACS counts and acceleration factors (R=4–6), so the method's value grows exactly where fast imaging is hardest.","The constraint choice matters: TV beats Fourier-domain $\\ell^1$ sparsity and nuclear-norm low-rank on reconstruction metrics and costs far less than the SVD-per-iteration low-rank option.","ACS-free reconstruction is possible: under Gaussian undersampling without calibration lines, INR-CRISTAL outperforms SAKE, P-LORAKS, and IMJENSE.","Image and sensitivity TV are separate tuning knobs ($\\lambda_1$, $\\lambda_2$); the ablation shows the sensitivity term still contributes roughly 0.5 dB even when ACS is plentiful."],"supporting_citations":[{"why":"IMJENSE supplies the joint image-sensitivity INR objective (Eq. 1) that INR-CRISTAL extends with a sensitivity regularizer.","marker":"[38]"},{"why":"CineJENSE is the INR-based sensitivity estimation that fails with limited ACS; the paper's 'OURS w.o. Reg' ablation is a 2D version of it.","marker":"[39]"},{"why":"SIREN provides the periodic-activation architecture used for both the image and sensitivity networks, credited with capturing high-frequency detail.","marker":"[47]"},{"why":"iSENSE establishes that sensitivity-map regularization improves optimization-based parallel imaging, motivating the added $R(S_\\beta)$ term.","marker":"[44]"},{"why":"JSENSE-Pro is the source of the pre-learned-subspace sensitivity priors; it motivates the Fourier-sparsity and low-rank regularizers compared in Section III-B.","marker":"[45]"},{"why":"JSENSE introduced joint image-sensitivity estimation, the formulation INR-CRISTAL builds on.","marker":"[7]"},{"why":"Fourier feature embedding is the input encoding for both networks, included to help them learn high-frequency image content.","marker":"[48]"},{"why":"fastMRI knee dataset supplies Dataset IV used for the robustness comparisons across ACS and R.","marker":"[51]"},{"why":"fastMRI brain datasets supply Datasets I and II for the main reconstruction comparisons.","marker":"[52]"}],"fun_headline_variants":["TV-regularized coil maps sharpen undersampled MRI","INR-CRISTAL: Smooth sensitivity maps beat auto-calibration","Coil sensitivity prior improves fast parallel MRI","Smoothing coil maps lifts MRI artifact removal","New INR method refines MRI with coil map constraint"],"cache_read_input_tokens":16128,"weakest_assumption_plain":"The load-bearing assumption is that the total-variation penalty steers the jointly estimated sensitivities toward the true physical coil sensitivities, rather than toward some other smooth image-times-sensitivity split that also fits the measured data; since the paper never quantifies sensitivity error — the maps are only inspected visually (Figs. 1 and 3) — the claimed mechanism that 'better sensitivities drive better images' is not directly measured.","fun_headline_variants_meta":{"raw":{"variants":["TV-regularized coil maps sharpen undersampled MRI","INR-CRISTAL: Smooth sensitivity maps beat auto-calibration","Coil sensitivity prior improves fast parallel MRI","Smoothing coil maps lifts MRI artifact removal","New INR method refines MRI with coil map constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1233,"prompt_tokens":995,"completion_tokens":238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":161}},"tokens_in":611,"tokens_out":238,"duration_ms":3240,"temperature":1.0,"reasoning_tokens":161,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T06:01:07.588625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute ground-truth sensitivities from fully sampled multi-coil data (coil image divided by the sum-of-squares image) and, under a fixed mask such as ACS=8, R=5, run INR-CRISTAL with and without the sensitivity TV term, reporting the relative $\\ell^2$ error of the estimated sensitivity maps. If the regularized run does not reduce sensitivity error relative to the unregularized run — while still improving image metrics — the paper's stated mechanism fails; if the sensitivity error falls together with reconstruction error, the mechanism is confirmed. A phantom with a known sharp sensitivity edge would also test whether the TV prior discards genuine sensitivity structure.","supporting_citations":[{"cited_title":"IMJENSE: scan-specific implicit representation for joint coil sensitivity and image estimation in parallel MRI,","cited_arxiv_id":null,"evidence_quote":"IMJENSE supplies the joint image-sensitivity INR objective (Eq. 1) that INR-CRISTAL extends with a sensitivity regularizer."},{"cited_title":"CineJENSE: Simultaneous cine MRI image reconstruction and sensitivity map estimation using neural representa- tions,","cited_arxiv_id":null,"evidence_quote":"CineJENSE is the INR-based sensitivity estimation that fails with limited ACS; the paper's 'OURS w.o. Reg' ablation is a 2D version of it."},{"cited_title":"Implicit neural representations with periodic acti- vation functions,","cited_arxiv_id":null,"evidence_quote":"SIREN provides the periodic-activation architecture used for both the image and sensitivity networks, credited with capturing high-frequency detail."},{"cited_title":"Iterative estimation of MRI sensitivity maps and image based on sense reconstruction method (iSENSE),","cited_arxiv_id":null,"evidence_quote":"iSENSE establishes that sensitivity-map regularization improves optimization-based parallel imaging, motivating the added $R(S_\\beta)$ term."},{"cited_title":"JSENSE-Pro: Joint sensitivity estimation and image reconstruction in parallel imaging using pre-learned subspaces of coil sensitivity functions,","cited_arxiv_id":null,"evidence_quote":"JSENSE-Pro is the source of the pre-learned-subspace sensitivity priors; it motivates the Fourier-sparsity and low-rank regularizers compared in Section III-B."},{"cited_title":"Joint image reconstruction and sensitivity estimation in SENSE (JSENSE),","cited_arxiv_id":null,"evidence_quote":"JSENSE introduced joint image-sensitivity estimation, the formulation INR-CRISTAL builds on."},{"cited_title":"Fourier features let networks learn high frequency functions in low dimensional domains,","cited_arxiv_id":null,"evidence_quote":"Fourier feature embedding is the input encoding for both networks, included to help them learn high-frequency image content."},{"cited_title":"fastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning,","cited_arxiv_id":null,"evidence_quote":"fastMRI knee dataset supplies Dataset IV used for the robustness comparisons across ACS and R."}],"review_version":1}