{"id":"6a89fe4e-c549-4875-96b8-5c3adbeb7960","arxiv_id":"2507.01326","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"S2DNets corrects MRI bias fields with two self-supervised networks trained by structural and smoothness constraints, reporting better scores than prior methods on most comparisons.","lead":"This paper presents a self-supervised deep learning method for MRI bias field correction that combines fuzzy clustering and smoothness constraints. If it holds up, it could let hospitals correct uneven MRI brightness without needing clean reference scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"As written, the smoothness-constrained objective is not well-defined: Eq. (10) defines TV as the signed Laplacian integral, and the λ schedule in §3.2 can cancel the TV term, so the reported ablation cannot be reproduced from the text.","rationale":"The strongest claim requires that the smoothness constraint is doing useful work. The reader's weakest assumption already identifies the loss inconsistency; my independent reading confirms it and adds the precise mechanism: an adaptive weight defined as the ratio of the two loss terms can cancel the TV term entirely unless the ratio is detached, a detail absent from the paper. This is not a mere typo in an unimportant term: the ablation S2DNets* versus S2DNets is the paper's only direct evidence for the contribution of the TV loss, and the reported gains (9.8–17.6%) are attributed to it. The HCP corruption-protocol and cluster-count-selection issues are also serious reproducibility concerns, but even if those were fixed, the paper would still not specify a valid smoothness objective. I therefore treat the loss/λ formulation as the single load-bearing concern. I am not claiming the implementation is wrong; the public code may well contain a proper TV loss and a detached schedule. But the text gives no way to know, and the equations as printed make the ablation impossible. Because correction is possible, the appropriate verdict remains conditional rather than rejection.","tokens_in":9349,"tokens_out":6348,"duration_ms":76215,"concrete_test":"Inspect the released repository (https://github.com/LeongDong/S2DNets) and check the exact loss_tv implementation and whether lambda is detached. Then re-run the ablation on a fixed HCP split: train the full model with (a) the printed Eq. (10) and λ = Lossbias/LossTV without detach, and (b) λ = 0. If runs (a) and (b) produce identical PSNR, the TV term is vacuous as written; if they differ, the implementation deviates from Eqs. (9)–(10) and §3.2, and the actual objective used must be stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that the TV smoothness loss substantially improves PSNR—depends entirely on the loss actually being minimized. As printed, Eq. (10) sets LossTV = ∫Ω [∂xx b(r) + ∂yy b(r)] dr, which is the integral of the signed Laplacian, not a total variation penalty; it can be zero or negative and does not regularize the total variation of b. Section 3.2 then sets λ = Lossbias / LossTV. If λ is computed from the same loss tensors without detaching, the total loss is Lossbias + (Lossbias/LossTV)·LossTV = 2·Lossbias, so the TV term vanishes from the objective and the S2DNets* and S2DNets rows in Tables 1–3 should be identical, contradicting the reported ablation. If λ is detached, as is standard in implementations, the equality is not exact, but then the manuscript still fails to state the actual schedule, and Eq. (10) remains not a form of total variation. Either way, the distinctive smoothness constraint that the abstract credits for the results is not reproducibly defined, so the central claim is not currently supported by the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes S2DNets, a self-supervised dual-network method for MR bias field correction. Two encoder–decoder networks estimate a fuzzy clustering membership map and a bias field, and are trained with losses constructed from the closed-form fixed-point updates of a unified energy functional that combines piecewise-constant structural constraints with local and global smoothness terms. Experiments on HCP and BrainWeb report improved PSNR/SSIM, lower coefficient of variation, and better downstream segmentation Dice compared with several classical and deep-learning baselines. The authors state that source code is publicly available.","tokens_in":9602,"tokens_out":8191,"duration_ms":91725,"significance":"If the method were reproducible as described, the core idea—using closed-form updates of a fuzzy-clustering bias model as self-supervised dual-network losses—would be a useful contribution, and the downstream segmentation evaluation is a strength. The paper also provides a public code link and uses a clearly stated multiplicative image formation model. However, the manuscript contains load-bearing inconsistencies in the definition and weighting of the smoothness loss, and in the selection of the cluster number, so the reported gains are not currently supported by the text.","major_comments":[{"comment":"Equation (10) defines the 'total variation loss' as ∫Ω [∂²xb(r) + ∂²yb(r)] dr, which is the integral of the signed Laplacian, not a total variation seminorm. Such an integral can be zero or negative, and it does not penalize oscillations of b in the way a true TV term would. The authors should replace Eq. (10) with a correct total variation expression (e.g., ∫Ω |∇b| dr or ∫Ω (|∂xb| + |∂yb|) dr) and ensure that all experimental results are described consistently with that corrected definition.","section":"§2.2, Eq. (10)"},{"comment":"The text states λ = Loss_bias / Loss_tv. If λ is computed from the same loss tensors without stopping gradients, the total loss becomes Loss_bias + (Loss_bias / Loss_tv)·Loss_tv = 2·Loss_bias, so the TV term cancels identically and the ablations S2DNets* and S2DNets in Tables 1–3 should coincide. If λ is detached or computed from a different schedule, that schedule is not stated. The manuscript must specify exactly how λ is computed, whether gradients flow through it, and how the reported differences between S2DNets* and S2DNets arise.","section":"§3.2, λ setting"},{"comment":"The cluster number N is selected by 'gradually increas[ing] it by 1 to find appropriate value 4 as Fig 2 shown', and Fig. 2 reports performance on HCP, BrainWeb T1, and BrainWeb T2. If these are the same test sets used for the comparisons in Tables 1–3, then N is tuned on the test data, which invalidates a fair comparison with other methods. An independent validation split must be used for choosing N, or the authors must demonstrate that no test information influenced the choice.","section":"§3.3 and Fig. 2"},{"comment":"For HCP, the paper reports PSNR/SSIM and downstream segmentation Dice against ground truth, but it never states how clean HCP images and corrupted inputs were obtained. The sentence 'training sets are further randomly splitted into 10000/5000 in HCP ... for corrupted input data and unpaired clean data' implies that synthetic bias fields were applied to HCP slices, but the protocol (e.g., bias field model, strength range, whether the same slices appear in corrupted and clean partitions) is not described. This is essential for reproducibility and for interpreting the 'Original' row in Tables 1–2.","section":"§3.1–§3.2"},{"comment":"Equation (6) is not the exact closed-form minimizer of Eq. (4) with respect to the membership variables. Minimizing Eq. (4) for fixed b and c gives u_i(s) proportional to [∫ K(r,s)|I(s) − b(r)c_i|² dr]^{-1/(p-1)}, which depends on a kernel-smoothed version of the intensity. Equation (6) instead uses the pointwise term |I(r) − c_i ∫ K(r,s)b(s)ds|². If this replacement is an approximation, it should be stated explicitly; otherwise the claimed derivation of the reconstruction loss is incorrect.","section":"§2.2, Eq. (6)"}],"minor_comments":[{"comment":"The GitHub URL in the abstract contains duplicated braces ('.../S2DNets}{https://.../S2DNets'); it should be a single clean link.","section":"Abstract and code link"},{"comment":"The paper alternates between 'unsupervised' and 'self-supervised'; since the method trains without any clean reference, the term self-supervised should be defined and used consistently.","section":"Terminology"},{"comment":"The figure caption in the text does not state which metric is plotted on the vertical axis or how the curves are aggregated; please specify this in the caption.","section":"Fig. 2"},{"comment":"Table 3 reports Dice for GM and WM using a pre-trained 2D U-Net and FSL, but the text does not specify whether the U-Net is a single multi-class model or separate binary models, nor which FSL tool and parameters were used; these details are needed for reproducibility.","section":"§3.4, downstream segmentation"},{"comment":"Reference [28] appears not to be cited in the text, and reference [29] has garbled author formatting; please check the reference list against the citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the combination of Eq. (10) and the λ schedule in §3.2: as printed, the TV term cancels, so the central ablation claim is unreproducible. This is a load-bearing error, not a cosmetic one. Because the underlying framework is interesting and the errors are local in principle, I recommend major revision rather than rejection. The authors should also be asked to move the N selection to a validation split and to disclose the HCP synthetic corruption protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely self-supervised bias-field correction pipeline, and the dual-network/fuzzy-clustering design is a real synthesis. The experiments are broad and the downstream segmentation results are a nice touch. But the smoothness loss as printed is not a total variation term—Eq. (10) is the integral of the Laplacian—and the stated λ = Loss_bias / Loss_TV makes the TV term cancel algebraically, so the reported ablation cannot be reproduced from the text. The paper needs a corrected section 2.2 and implementation details before the claims are supported.\n\nWhat is new: combining dual-regression style self-supervision (reconstruct b and u from closed-form updates) with MICO-style fuzzy clustering and explicit smoothness penalties. That specific package is not in the cited literature. The relationship between the cluster probability map and the bias field is sensible, and the closed-form updates in Eqs. (5)–(7) look plausible. Using CV and downstream segmentation as evaluation is appropriate.\n\nSoft spots, in rough order: the TV loss definition is wrong as written; the λ schedule as written nullifies the very term the ablation credits; cluster number N is selected by looking at test-set performance (Fig. 2); the HCP synthetic corruption protocol isn't described (HCP is listed as clinical but bias field simulation is only mentioned for BrainWeb—need clarity); and the abstract's 'outperforms' claim conflicts with Table 1 where IRNet wins PSNR on HCP and SSIM on BrainWeb T1. The self-referential training losses could also admit degenerate constant-bias solutions; the paper doesn't discuss what prevents that. From the text alone, one cannot tell whether the actual implementation uses detached λ or some other schedule.\n\nThis is not a desk-reject. The core idea is promising and the experiments, once the equations are fixed, could be a solid contribution. But as it stands the central smoothness result is not reproducible from the text.\n\nRecommendation: send to peer review and ask for a major revision with corrected loss definitions, an explicit λ schedule, and a description of the HCP corruption protocol. A reader who wants to build on this would need the code, which is promised but was not inspected here.","headline":"A promising self-supervised bias-field correction pipeline, but the TV loss and λ schedule as printed cancel each other, so the central ablation is not reproducible from the text.","tokens_in":10153,"tokens_out":1613,"would_cite":false,"duration_ms":18634,"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":"S2DNets is a self-supervised dual-network framework that corrects MRI bias fields by enforcing piece-wise structure and smoothness constraints, with no clean reference images, and it reports better fidelity and segmentation Dice than…","keywords":["bias field correction","MRI intensity inhomogeneity","self-supervised learning","dual networks","fuzzy clustering","smoothness constraint","structural constraint","downstream segmentation"],"falsifier":"Run the released code with $\\lambda$ set exactly as Section 3.2 prescribes and with the smoothness loss removed; if the two networks produce the same outputs and the reported ablation gap (roughly 10--18% PSNR) does not appear, then the claimed smoothness mechanism is not what drives the improvement.","tokens_in":9121,"feed_emoji":"🧠","tokens_out":10092,"duration_ms":91439,"temperature":0.7,"pith_summary":"The paper is trying to establish that MRI bias field correction can be solved self-supervised, without paired or unpaired clean images, by making two physical properties explicit: the corrected image should be piece-wise constant within tissue classes, and the bias field should be smooth. To that end it proposes S2DNets, two encoder-decoder networks that alternately reconstruct each other's outputs: one estimates a fuzzy clustering map of tissue classes, the other estimates the multiplicative bias field. The losses come from closed-form solutions of a unified objective that combines local Gaussian smoothness, a global smoothness term, and the clustering structure constraint. On HCP and BrainWeb the paper reports best or near-best fidelity and coefficient-of-variation scores among conventional and deep-learning comparators, and improved Dice for downstream segmentation with a learned U-Net and with classical segmentation software. A reader should care because if the claim holds, non-uniform intensity can be removed from clinical MRI without any clean reference data, which is often unavailable.","feed_headline":"Two self-supervised networks beat prior MRI bias-field correction","feed_subtitle":"Structure and smoothness constraints remove MRI bias without clean references and boost segmentation.","key_machinery":"The load-bearing object is the closed-form solution loop in Eqs. (5)--(7), which turns a single variational objective into two reconstruction losses. For fixed bias field and cluster centers, Eq. (6) gives the optimal membership probability map; for fixed memberships and centers, Eq. (7) gives the optimal bias field. Each subnetwork is trained to output the quantity whose closed-form reconstruction matches its prediction, so the two networks supervise each other. Local smoothness enters through the masked Gaussian kernel $K(r,s)$ in the objective, global smoothness through the added regularization term on the bias field, and the structure constraint through the clustering assumption that clean tissue intensities are piece-wise constant.","core_discovery":"The central discovery claimed is that the bias field and the piece-wise constant structure map can be learned as mutually constraining dual tasks, and that the fixed-point equations of a single objective suffice to supervise both networks. Given a corrupted image, the model alternates between Eq. (6), which reconstructs the clustering probability map from the current bias field, and Eq. (7), which reconstructs the bias field from the current clustering map; training each network to reproduce the other's reconstruction enforces consistency without clean references. The paper argues that this dual feedback, strengthened by local Gaussian smoothness and a global smoothness term on the bias field, removes non-uniform intensity while retaining edges and tissue boundaries. It reports that the complete model reaches PSNR 28.40 / SSIM 0.979 on HCP, 31.78 / 0.979 on BrainWeb T1, and 30.98 / 0.969 on BrainWeb T2, and that the global smoothness loss contributes PSNR gains of roughly 9.8%, 17.6%, and 17.3% across the three datasets.","pith_inferences":["A testable extension the authors do not report is cross-dataset transfer: train S2DNets on HCP and apply it to BrainWeb without retraining, which would show whether the learned smoothness and structure priors generalize across scanners rather than fitting dataset-specific statistics.","The closed-form dual-loss design is not obviously limited to MRI; the same multiplicative model with a piece-wise constant prior describes illumination fields in natural images, so the method could be retargeted to Retinex-style corrections, but the paper gives no evidence on that domain.","Since evaluation is on 2D slices, an implicit claim is that slicewise bias fields are consistent with the 3D volume; a volumetric implementation would need to verify that neighboring slices do not produce discontinuous bias estimates."],"forward_implications":["If S2DNets is correct, bias field correction can be done entirely without clean reference scans, removing a major data bottleneck for clinical MRI pipelines.","The reported downstream Dice improvements imply that correcting the bias field with structural and smoothness constraints before segmentation helps both learned and classical segmentation tools.","The ablation claims that the global smoothness term alone raises PSNR by roughly 10--18%, which would mean smoothness is the component doing much of the correction work rather than the clustering structure term.","Because the same closed-form dual-loop losses need no labels, the framework could be retrained cheaply on new scanners or protocols when intensity distributions change."],"supporting_citations":[{"why":"This is the N3 baseline whose high-frequency and smoothness view of bias fields motivates the paper's smoothness constraints.","marker":"[3]"},{"why":"This is the N4 baseline, one of the two conventional methods the paper must beat.","marker":"[4]"},{"why":"This supplies the fuzzy-clustering objective the paper adapts into its piece-wise structure constraint.","marker":"[5]"},{"why":"This is the MICO baseline that couples bias field estimation with tissue segmentation.","marker":"[8]"},{"why":"This supplies the multiplicative bias field model and the Legendre-polynomial simulation used to generate BrainWeb corrupted data.","marker":"[11]"},{"why":"This is the residual cycle-GAN baseline trained with unpaired clean data.","marker":"[12]"},{"why":"This is the IRNet baseline, the strongest deep-learning comparator in the fidelity and segmentation tables.","marker":"[13]"},{"why":"This is the U-Net backbone used for both S2DNets subnetworks and for the downstream segmentation model.","marker":"[23]"},{"why":"This provides the HCP clinical dataset used in training and testing.","marker":"[24]"},{"why":"This provides the BrainWeb phantom T1/T2 slices with simulated bias fields.","marker":"[25]"}],"fun_headline_variants":["Dual self-supervised networks fix MRI bias without clean data","Self-supervised dual nets remove MRI bias, boost segmentation","Structure-smoothness dual networks outperform prior MRI bias correction","S2DNets: self-supervised dual networks for bias-free MRI","Dual networks learn bias field and structure without references"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the implemented training loss differing from Eqs. (10)--(11) as written, because the printed 'total variation' term is a sum of second derivatives and the printed adaptive weight $\\lambda = \\mathrm{Loss}_{\\mathrm{bias}}/\\mathrm{Loss}_{\\mathrm{tv}}$ would cancel the smoothness term, so the claimed smoothness mechanism must come from a corrected loss that the paper does not specify.","fun_headline_variants_meta":{"raw":{"variants":["Dual self-supervised networks fix MRI bias without clean data","Self-supervised dual nets remove MRI bias, boost segmentation","Structure-smoothness dual networks outperform prior MRI bias correction","S2DNets: self-supervised dual networks for bias-free MRI","Dual networks learn bias field and structure without references"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001052,"raw_usage":{"total_tokens":4429,"prompt_tokens":967,"completion_tokens":3462,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":3378}},"tokens_in":583,"tokens_out":3462,"duration_ms":103430,"temperature":1.0,"reasoning_tokens":3378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:54:59.178003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code with $\\lambda$ set exactly as Section 3.2 prescribes and with the smoothness loss removed; if the two networks produce the same outputs and the reported ablation gap (roughly 10--18% PSNR) does not appear, then the claimed smoothness mechanism is not what drives the improvement.","supporting_citations":[{"cited_title":"G., Zijdenbos, A","cited_arxiv_id":null,"evidence_quote":"This is the N3 baseline whose high-frequency and smoothness view of bias fields motivates the paper's smoothness constraints."},{"cited_title":"J., Avants, B","cited_arxiv_id":null,"evidence_quote":"This is the N4 baseline, one of the two conventional methods the paper must beat."},{"cited_title":"K., Agrawal, S., Panda, R., Abraham, A.: Novel fuzzy clustering-based bias field correction technique for brain magnetic resonance images","cited_arxiv_id":null,"evidence_quote":"This supplies the fuzzy-clustering objective the paper adapts into its piece-wise structure constraint."},{"cited_title":"C., Davatzikos, C.: Multiplicative intrinsic component optimiza- tion (MICO) for MRI bias field estimation and tissue segmentation","cited_arxiv_id":null,"evidence_quote":"This is the MICO baseline that couples bias field estimation with tissue segmentation."},{"cited_title":"K., Lin, W., Wang, L., Shen, D., Li, G.: ABCnet: Adversarial bias correction network for infant brain MR images","cited_arxiv_id":null,"evidence_quote":"This supplies the multiplicative bias field model and the Legendre-polynomial simulation used to generate BrainWeb corrupted data."},{"cited_title":"J., Patel, P., Liu, T., Yang, X.: Intensity non-uniformity correction in MR imaging using residual cycle generative adversarial network","cited_arxiv_id":null,"evidence_quote":"This is the residual cycle-GAN baseline trained with unpaired clean data."},{"cited_title":"M., Jing, B., Zhang, G., Chen, G","cited_arxiv_id":null,"evidence_quote":"This is the IRNet baseline, the strongest deep-learning comparator in the fidelity and segmentation tables."},{"cited_title":"In: 18th International Conference on Medical Image Com- puting and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"This is the U-Net backbone used for both S2DNets subnetworks and for the downstream segmentation model."},{"cited_title":"S., Glasser, M","cited_arxiv_id":null,"evidence_quote":"This provides the HCP clinical dataset used in training and testing."},{"cited_title":"L., Zijdenbos, A","cited_arxiv_id":null,"evidence_quote":"This provides the BrainWeb phantom T1/T2 slices with simulated bias fields."}],"review_version":1}