{"id":"2dcc83b6-7768-4ac4-a024-42339ae8a4b2","arxiv_id":"2501.18074","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Concatenating a Tikhonov-regularized version of the noisy MRI signal to the network input improves myelin water fraction estimation in biexponential analysis.","lead":"This paper tests a neural network trick called input layer regularization: before estimating myelin water fraction from MRI signals, the noisy signal is concatenated with a smoothed, regularized version of itself. On synthetic and brain data, the authors report small but consistent accuracy gains over plain networks, with classical GCV tuning slightly better than a learned lambda selector.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic and in vivo echo-time grids may differ ([8, 256] ms vs [11.3, 361.6] ms), so the claimed in vivo ILR benefit may be an out-of-distribution artifact.","rationale":"The paper's central claim is that ILR improves MWF estimation both on synthetic data and in vivo brain data. The synthetic evidence (Table 2) is plausible, but the in vivo demonstration is essential for the practical claim of improved MWF mapping. My review of the acquisition details reveals a likely inconsistency: synthetic training uses echo times 8–256 ms (Table A.3), while in vivo acquisition uses 11.3–361.6 ms (Appendix E.1). Deep learning estimators for decay signals are not invariant to the time grid; they learn features specific to the sampled decay curve. Without resampling to a common grid or retraining on the in vivo grid, the networks are applied out-of-distribution. This would not affect the synthetic comparison but would invalidate the in vivo claim that (ND, Reg) GCV outperforms (ND, ND). This is a more concrete and testable concern than the general interpolation limitation: it identifies a specific distribution shift, and it can be settled by inspecting the released code. The Discussion (Section 5) acknowledges the lack of a gold standard for in vivo data but does not mention any echo-time grid mismatch; if present, this is an unacknowledged out-of-distribution application. I therefore recommend keeping the verdict CONDITIONAL, pending verification of the echo-time grid consistency.","tokens_in":20906,"tokens_out":15358,"duration_ms":154813,"concrete_test":"Check the repository https://github.com/ShashankSule/ILR-for-MWF to confirm the exact echo-time vector used to generate synthetic training data in Appendix A and the echo-time vector used for in vivo voxel signals in Appendix E.1. If the vectors are identical (e.g., both use the same 32 time points), this concern is resolved. If they differ, rerun the in vivo comparison with networks trained on the in vivo time grid (or resample signals to the training grid) and see whether the (ND, Reg) GCV advantage over (ND, ND) persists; if it disappears, the in vivo demonstration of ILR benefit is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ILR improves MWF estimation on brain data rests on applying networks trained on synthetic signals to in vivo signals. Appendix A (Table A.3) states that synthetic training signals use Nt=32 acquisition times spanning [8.0, 256.0] ms, while Appendix E.1 states that the in vivo GRASE data acquire 32 echoes at times tn = n×11.3 ms, spanning [11.3, 361.6] ms. Since the network input is the raw signal vector with no explicit time-axis encoding, a change in the echo-time grid changes the input distribution; the text does not state that the in vivo signals were resampled to the training grid or that the networks were retrained on the in vivo grid. If the grids differ, the in vivo comparison in Figure 7 (and the conclusion that (ND, Reg) GCV outperforms (ND, ND)) is an out-of-distribution evaluation, and the reported advantage could be an artifact of input distribution shift rather than a genuine ILR effect. The synthetic results in Table 2 are unaffected, but the practical claim of improved MWF mapping in human brain would be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes input layer regularization (ILR) for estimating the myelin water fraction c1 and the associated decay constants from biexponential MRI signals. The pipeline selects a Tikhonov regularization parameter per signal either by a dedicated neural network (λNN) or by generalized cross-validation (GCV), constructs a regularized signal G(p*_λ(s)), concatenates it with the noisy signal to form the network input, and trains a multilayer perceptron to estimate (c1, T2,1, T2,2). Synthetic experiments compare (ND,Reg)NN, (ND,Reg)GCV, (ND,ND), TR-NLLS, and NLLS at three SNR levels; in vivo brain data are analyzed after AIC-based selection of biexponential pixels, with a NESMA-denoised NLLS map as the reference standard. The central claim is that ILR significantly improves c1 estimation accuracy relative to plain networks and classical methods.","tokens_in":21130,"tokens_out":5995,"duration_ms":58113,"significance":"If the results hold, ILR is a simple and appealing hybrid method: it combines classical Tikhonov regularization with deep learning, removes the need for a manually fixed regularization parameter, and extends prior two-parameter work [36] to the practically relevant three-parameter problem including MWF. The manuscript's strengths include a reproducible synthetic design with known ground truth, comparison against several baselines, per-signal λ selection, and a detailed region-wise analysis in Appendix D. The main limitation is that the in vivo evaluation, which is essential for the practical claim, currently rests on a distribution-shifted input grid and a reference standard from the same model family as the method; the synthetic claims are also reported without uncertainty quantification.","major_comments":[{"comment":"The synthetic training signals use Nt=32 acquisition times spanning [8.0, 256.0] ms, while the in vivo GRASE data acquire 32 echoes at t_n = n×11.3 ms, spanning [11.3, 361.6] ms. Because the network input is the raw signal vector with no explicit time-axis encoding, a change in the echo-time grid changes the input distribution. The text nowhere states that the in vivo signals were resampled to the training grid or that the networks were retrained on the in vivo grid. Consequently, the in vivo comparisons in Figure 7 and the conclusion that (ND,Reg) GCV outperforms (ND,ND) on brain data are out-of-distribution evaluations, and the claimed practical benefit is not yet supported. Please either demonstrate that the training grid covers the in vivo acquisition (for example, by resampling or retraining) or restrict the claims to the synthetic setting.","section":"Appendix A, Table A.3; Appendix E.1"},{"comment":"The central claim of a 'significant improvement' is based on single RMSE values per method and SNR, with no error bars, confidence intervals, or multiple-seed variability. At SNR=5 the difference between (ND,Reg)NN (0.1652) and (ND,ND) (0.1671) is about 1%, and at SNR=50 the difference between the two ILR variants is 0.0931 versus 0.0945. Without repeated training runs or a statistical test, the aggregate improvement at low SNR and the claimed difference between GCV and NN selection cannot be distinguished from noise. Please add uncertainty quantification or soften the significance claim accordingly.","section":"Table 2; Section 5"},{"comment":"The validation and testing parameter ranges (c1 in [0, 0.5], T2,1 in [5, 45], T2,2 in [45, 200]) are strictly inside the training ranges ([0, 0.6], [1, 50], [40, 225]), and no out-of-distribution or extrapolation experiments are reported. The reported accuracy gains are therefore interpolation results only. Since the in vivo data involve a different echo-time grid and likely different noise conditions, the paper should either add out-of-distribution tests (for example, parameters outside the training box, or synthetic signals on the in vivo grid) or explicitly limit the generalization claim.","section":"Table A.3; Appendix D"},{"comment":"The in vivo reference standard is obtained by applying the NESMA denoising filter followed by NLLS, and the ILR input in Eq. (13) contains a TR-NLLS-regularized curve. Both the reference and one input component therefore come from the same biexponential NLLS family, which could systematically favor (ND,Reg) over (ND,ND) in the in vivo comparison independent of any genuine ILR effect. The manuscript acknowledges that the reference is 'necessarily flawed' but does not address this shared-model concern. Please discuss this potential bias and, if possible, provide a validation on data whose ground truth is not NLLS-based (for example, synthetic data on the in vivo grid with known parameters).","section":"Section E.2; Eq. (13)"}],"minor_comments":[{"comment":"In the low-SNR discussion, the sentence 'when 0.25 ≤ c1 ≤ 0.32, (ND, Reg) NN outperforms (ND, ND) for more values of (T21, T22) than (ND, Reg) NN' appears to compare (ND,Reg)NN with itself; it should likely compare (ND,Reg)NN with (ND,Reg)GCV.","section":"Appendix D"},{"comment":"The right panel caption contains 'GCG', which should be 'GCV'.","section":"Figure 7 caption"},{"comment":"In the text after Eq. (1), 'the corresponding spin-spin decay constants T2,1 and T2,1' should read 'T2,1 and T2,2'.","section":"Section 2.1"},{"comment":"Equations (B.1) and (B.2) contain unbalanced parentheses, and the loss multiplies the c1 error by 100 relative to the time constants; this weighting should be stated in the main text because it directly affects the reported RMSE comparisons.","section":"Appendix B.2, Eq. (B.1)-(B.2)"},{"comment":"Section 4.2 states that 'a multi-spin-echo sequence with 64 values of TE was applied', whereas Appendix E.1 reports that echo data were acquired for 32 echoes with t_n = n×11.3 ms; these statements should be reconciled.","section":"Section 4.2 versus Appendix E.1"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the echo-time grid mismatch between synthetic training and in vivo acquisition, which undermines the practical claim as currently written. If the authors can show that the in vivo signals are effectively within the training distribution (for example, by resampling to the same grid) or retrain on the in vivo grid, the synthetic results alone may justify a revised paper. The manuscript is otherwise within scope for q-bio.QM, and the code release is a positive feature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The synthetic core of this paper holds up: combining input layer regularization with per-signal lambda selection (GCV or a small NN) does improve c1 estimation over a plain network at mid and high SNR, and the code is available. That is a genuine, if incremental, extension of your earlier ILR work, and the synthetic experiments use known ground truth. Credit where due: the paper is honest about its limitations, the network comparisons are controlled for input length, and the disaggregated analysis in Appendix D is a nice touch.\n\nThe soft spots are real, but not all equally load-bearing. The biggest one is the stress-test concern, and it is correct. Training signals are generated on a 32-point grid spanning [8, 256] ms, while the in vivo GRASE data are on a different 32-point grid, [11.3, 361.6] ms, with no stated resampling or retraining. Since the network input is just the raw signal vector, Figure 7 is an out-of-distribution evaluation. That does not touch the synthetic Table 2, but it means the practical in vivo claim is unsupported as written. Second, the test parameters are an interpolation subset of the training ranges, so the reported gains are only demonstrated inside the training support. Third, Table 2 has no error bars or significance tests; at SNR=5 the ILR gain over (ND, ND) is 0.0019 RMSE, which is noise-level. The GCV-over-NN advantage is also only shown on the single in vivo dataset, not in synthetic data. Fourth, the in vivo reference standard is a NESMA-filtered NLLS surrogate, so it shares the same model family as the method. Minor issues: Algorithm 1 uses an undefined 'popt', and references [36] and [67] are the same paper.\n\nNone of this breaks the central synthetic claim. The paper deserves a serious referee: the method is sensible, the code is public, and the core comparison is not circular. The authors should fix the echo-time grid issue, add error bars or repeated-run statistics, and soften the in vivo conclusions unless the grid mismatch is resolved. I would send it out, with the expectation of a major revision.\n\nI would bring it to a reading group only if the focus is on practical regularization for relaxometry; the methodological novelty is modest. I would cite it for the synthetic ILR result if I worked on MWF estimation, but I would not cite the in vivo comparison without seeing the grid issue addressed.","headline":"Solid synthetic evidence for ILR-based MWF estimation, but the in vivo comparison is undermined by an unaddressed echo-time grid mismatch.","tokens_in":21720,"tokens_out":2138,"would_cite":true,"duration_ms":23801,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E24","92B20","45Q05","62P10","92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Input layer regularization—concatenating a noisy MRI decay signal with a Tikhonov-regularized version of itself—lowers myelin water fraction estimation error relative to plain networks and classical NLLS in simulation and in vivo brain…","keywords":["Input layer regularization","Myelin water fraction","Biexponential analysis","Tikhonov regularization","Generalized cross validation","Bilevel optimization","Magnetic resonance relaxometry","Deep learning"],"falsifier":"Train both (ND, Reg) and (ND, ND) on the published grid, then evaluate them on known test signals with parameters outside that grid, for example $c_1>0.6$ or $T_{2,1}<1$ ms; if the ILR advantage disappears or reverses there, the improvement has only been shown inside the training distribution.","tokens_in":20692,"feed_emoji":"🧠","tokens_out":9010,"duration_ms":89768,"temperature":0.7,"pith_summary":"The paper proposes a hybrid method, input layer regularization (ILR), that combines classical Tikhonov regularization with a neural network for estimating myelin water fraction from MRI biexponential decay signals. Instead of feeding the network only the measured noisy signal, ILR appends a regularized reconstruction of that same signal, with the per-signal regularization strength chosen either by generalized cross validation or by a trained network. On simulated data, the augmented networks estimate the fast-decay fraction $c_1$ with lower root mean squared error than a plain network, Tikhonov-regularized NLLS, and conventional NLLS at all tested SNR levels. On in vivo brain data, the ILR networks again outperform the plain network, and GCV-based selection is somewhat superior to network-based selection. This matters because myelin water fraction is a candidate biomarker for demyelinating disease, so more accurate pixel-wise estimation could improve clinical mapping of white matter.","feed_headline":"Regularized input layer cuts myelin water estimation error","feed_subtitle":"Classical Tikhonov regularization combined with a neural network beats plain deep learning on simulations and human brain MRI.","key_machinery":"The load-bearing object is the augmented input vector $x=(s,\\, G(p^*_{\\lambda(s)}(s)))$, in which the noisy decay vector $s$ is concatenated with a regularized noiseless signal generated from a Tikhonov-regularized NLLS fit at a per-signal $\\lambda(s)$. Two selectors produce $\\lambda(s)$: a convolutional network $\\lambda_{\\mathrm{NN}}$ trained with $L^1$ loss to match an oracle $\\lambda$ computed by grid search on the bilevel problem, and generalized cross validation $\\lambda_{\\mathrm{GCV}}$ for nonlinear inverse problems. The parameter-estimation network then maps this concatenated vector to $(c_1,T_{2,1},T_{2,2})$; a control network (ND, ND) receives the signal concatenated with itself to keep input length identical. The regularized component injects a smooth, denoised version of the signal so that the estimator can exploit both the raw measurements and a stabilized reconstruction.","core_discovery":"The central claim is that appending a Tikhonov-regularized version of the noisy decay signal to the network input improves estimation of $c_1$, the myelin water fraction, in the three-parameter biexponential model $s(t)=c_1 e^{-t/T_{2,1}}+(1-c_1)e^{-t/T_{2,2}}$ under Rician noise. The regularized component is formed as $G(p^*_{\\lambda(s)}(s))$, where $p^*_{\\lambda(s)}$ is the Tikhonov-regularized NLLS estimate with a signal-dependent regularization parameter $\\lambda(s)$ selected either by a convolutional network trained against an oracle $\\lambda$ or by generalized cross validation. On synthetic testing data, both ILR variants reduce $c_1$ RMSE relative to the plain (ND, ND) network at SNR 5, 50, and 100, and both are far below NLLS and TR-NLLS; on human brain data, (ND, Reg) GCV outperforms (ND, Reg) NN and (ND, ND) across most pixels. The paper thereby claims that classical regularization-parameter selection can be productively embedded in a deep-learning parameter-estimation pipeline.","pith_inferences":["Beyond the paper's experiments, ILR should transfer to other multiexponential signal models and to Gaussian noise, since the augmentation mechanism does not depend on the Rician likelihood or on the specific two-decay model.","This suggests an SNR-aware implementation: use $\\lambda_{\\mathrm{GCV}}$ when the estimated SNR is low and $\\lambda_{\\mathrm{NN}}$ when it is high, since their relative fidelity to the oracle distribution reverses with SNR.","A testable extension is to replace per-pixel independent processing with spatial context, for example by including neighboring voxels or a spatial penalty in the loss, which the paper notes as an open direction.","The parameter regions where the plain network wins—small $c_1$ with nearly identical decay times—imply ILR should be switched off when the fit approaches monoexponential behavior, rather than applied uniformly."],"forward_implications":["The same ILR network outputs improved estimates of $T_{2,1}$ and $T_{2,2}$ along with $c_1$, so the method addresses the full biexponential parameter vector, not just the myelin water fraction.","Because $\\lambda_{\\mathrm{NN}}$ approximates the oracle distribution better by earth mover's distance while GCV better captures the low-SNR, small-$\\lambda$ modes, the choice of regularizer selector can be guided by the signal's SNR regime.","Restricting analysis to AIC-selected biexponential pixels avoids the underdetermined monoexponential case, so ILR is intended for white-matter-like voxels rather than arbitrary tissue.","At medium and high SNR, ILR's advantage concentrates in the low-$c_1$ region with intermediate $T_{2,2}-T_{2,1}$ separation and in the roughly equal-weight region $c_1 \\ge 0.4$ where the biexponential problem is most ill-posed."],"supporting_citations":[{"why":"Establishes input layer regularization and its reported ~10% improvement in the two-parameter biexponential problem that this work extends to three parameters and MWF.","marker":"[36]"},{"why":"Supplies the neural-network approach for learning the optimal regularization parameter lambda from noisy signals.","marker":"[29]"},{"why":"Provides generalized cross validation as the classical method for choosing the ridge/Tikhonov parameter.","marker":"[25]"},{"why":"Extends GCV to nonlinear inverse problems, yielding the lambda_GCV criterion used to build the regularized input.","marker":"[28]"},{"why":"Supplies the biexponential decay model and the exponential-analysis framework underlying parameter estimation.","marker":"[10]"},{"why":"Provides the NESMA denoising filter used to form the in vivo comparison standard for myelin water fraction.","marker":"[39]"},{"why":"Supplies the Akaike information criterion used to classify pixels as biexponential before applying ILR.","marker":"[38]"}],"fun_headline_variants":["Deep learning plus Tikhonov improves myelin water maps","Automated tuning sharpens myelin water estimation","Regularized inputs sharpen myelin water estimates","Neural network tunes regularization for myelin water"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed accuracy gain assumes interpolation: every validation and test signal lies inside the parameter ranges used for training, and no out-of-distribution or extrapolation experiment is performed.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning plus Tikhonov improves myelin water maps","Automated tuning sharpens myelin water estimation","Regularized inputs sharpen myelin water estimates","Neural network tunes regularization for myelin water"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1555,"prompt_tokens":1024,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":640,"tokens_out":531,"duration_ms":5514,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:45:36.531379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train both (ND, Reg) and (ND, ND) on the published grid, then evaluate them on known test signals with parameters outside that grid, for example $c_1>0.6$ or $T_{2,1}<1$ ms; if the ILR advantage disappears or reverses there, the improvement has only been shown inside the training distribution.","supporting_citations":[{"cited_title":"Rozowski, J","cited_arxiv_id":null,"evidence_quote":"Establishes input layer regularization and its reported ~10% improvement in the two-parameter biexponential problem that this work extends to three parameters and MWF."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides generalized cross validation as the classical method for choosing the ridge/Tikhonov parameter."},{"cited_title":"Haber, D","cited_arxiv_id":null,"evidence_quote":"Extends GCV to nonlinear inverse problems, yielding the lambda_GCV criterion used to build the regularized input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the biexponential decay model and the exponential-analysis framework underlying parameter estimation."},{"cited_title":"Bouhrara, D","cited_arxiv_id":null,"evidence_quote":"Provides the NESMA denoising filter used to form the in vivo comparison standard for myelin water fraction."},{"cited_title":"James, D","cited_arxiv_id":null,"evidence_quote":"Supplies the Akaike information criterion used to classify pixels as biexponential before applying ILR."}],"review_version":1}