{"id":"001618c3-0a01-46ea-ad10-b903376ad610","arxiv_id":"2501.05583","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A reconstruction method that learns MPI scanner noise with an invertible neural network improves simulated image quality and ships with the MPI-MNIST benchmark dataset.","lead":"This paper introduces a learned noise model for Magnetic Particle Imaging reconstruction and a new public simulated benchmark built from MNIST digits. It reports better image similarity than classical methods, especially in low-concentration, high-noise settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (2.10) has a sign error before R(x): the stated objective is not the MAP estimate derived in Eqs. (2.6)–(2.9), so the evaluated method is not the method described.","rationale":"The reader's verdict identified the sign inconsistency in Eq. (2.10) as one of several issues but chose the empty-scanner-noise representativeness as the weakest load-bearing assumption. My read agrees that the noise-transfer concern is real for real-world applicability, yet the paper's central claim is explicitly validated on the MPI-MNIST benchmark, where noise is simulated from the same population used for training. On that benchmark, the most immediate threat to the claim is internal inconsistency: the published objective is not the MAP objective derived in the same section. If the minus sign is a typographical error, the method is recoverable and the numerical results may still be valid, which supports keeping the reader's CONDITIONAL verdict rather than escalating to REJECT. However, the sign error is not cosmetic: it changes the optimization landscape and potentially the well-posedness of the variational problem. Because the full text includes an explicit limitation statement that real-data evaluation and regularization theory remain for future work, the conditionality of the verdict is appropriate. My proposed check directly settles whether the sign error is merely textual or reflects a substantive mismatch between the method and its specification, which is a prerequisite for any meaningful validation of the central claim.","tokens_in":21210,"tokens_out":5184,"duration_ms":52201,"concrete_test":"Obtain the authors' code or re-implement the LDA objective exactly as printed in Eq. (2.10) using the same α grid, RK initialization, and iteration counts; if the objective is unbounded below or the optimization diverges for the reported α values, Eq. (2.10) cannot be the objective that produced Table 1. Then rerun the same experiment with the sign of R(x) flipped to plus; if Table 1 SSIM values are reproduced, the published equation is a typo that must be corrected before the method is well-defined. If neither sign reproduces the reported table, the results are not attributable to the method as specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (2.10) reads xLDA = arg min_x Dθ(Ax,yδ) − R(x). This contradicts the derivation: Eq. (2.6) minimizes −log p(yδ|x) − log pX(x), and the text directly below Eq. (2.7) defines R(x) = −log pX(x). Substituting the flow likelihood into Eq. (2.6) yields arg min_x Dθ(Ax,yδ) + R(x), not minus. If Eq. (2.10) is implemented literally with R(x) = α/2‖x‖², the term −R(x) rewards large norm; for α larger than the squared smallest singular value of the discretized forward operator, Dθ − R is unbounded below and gradient descent cannot produce the stable, high-SSIM reconstructions reported in Table 1. Since the reported LDA results are stable and good, the implementation almost certainly uses the plus sign, meaning the published equation does not describe the evaluated algorithm. This is a correctness and reproducibility defect at the center of the method definition: a reader implementing Eq. (2.10) as printed would either fail to reproduce the results or silently correct the sign and thus not test the published method. It also leaves open whether the reported SSIM gains depend on the sign-corrected objective or on auxiliary implementation choices such as RK initialization and the 2000-iteration budget.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Learned Discrepancy Approach (LDA) for linear inverse problems, in which the data discrepancy term of a Tikhonov-type objective is replaced by a learned log-likelihood of the noise, modeled by an invertible neural network trained on measured empty-scanner noise samples. It also introduces MPI-MNIST, a simulated benchmark of MNIST-derived phantoms with model-based MPI system matrices at several resolutions and additive noise frames from a preclinical Bruker scanner. In numerical experiments on this dataset, LDA initialized by regularized Kaczmarz reconstructions and minimized by gradient descent is reported to achieve the highest mean SSIM across all tested concentrations, with the largest gains at low concentration. The paper claims the method is applicable beyond MPI and that the dataset helps close the data gap in MPI.","tokens_in":21526,"tokens_out":6663,"duration_ms":66788,"significance":"The paper makes two useful contributions: a publicly accessible MPI benchmark with ground truth and realistic simulated measurements (Zenodo DOI), and a principled way of injecting a learned noise distribution into MAP estimation via a normalizing flow. The statistical derivation in Section 2.2 is clear, and the use of independent empty-scanner noise samples for training the discrepancy network avoids the most blatant form of circularity. The value is conditional, however: the experimental evidence is entirely simulated, with test noise drawn from the same scanner-noise population used for training, and hyperparameters are selected on test data. If the sign error in Eq. (2.10) is corrected and the evaluation is tightened, the method is a credible contribution to learned discrepancies for MPI and other inverse problems.","major_comments":[{"comment":"There is a sign inconsistency in the printed objective. Eq. (2.6) gives the MAP objective as argmin_x [ -log p(y_delta|x) - log p_X(x) ]. Substituting the flow-based noise density and using R(x) = -log p_X(x) as defined below Eq. (2.7) yields argmin_x [ D_theta(Ax,y_delta) + R(x) ], where D_theta(Ax,y_delta) = (1/2)||phi_theta(y_delta - Ax)||^2 - log|det J_phi_theta(y_delta - Ax)|. The printed Eq. (2.10), x_LDA = argmin_x D_theta(Ax,y_delta) - R(x), therefore has the wrong sign in front of R. For the implementation R(x) = (alpha/2)||x||^2 used in Section 6.1, the printed objective can be unbounded below when alpha is large or when the discretized forward operator is rank-deficient, so the stable and high-SSIM results in Table 1 cannot come from the objective as printed. The authors should correct the sign and state unambiguously which objective was used in the experiments: with the plus sign the method is the MAP estimator derived in the text; with the minus sign it is not, and the reported results would be unexplained.","section":"Section 2.2, Eq. (2.10)"},{"comment":"The empirical validation is in-distribution with respect to the noise. The discrepancy network is trained on roughly 300,000 noise samples from large_NoiseMeas.mdf, while the test measurements y_delta are generated by adding noise samples from the same scanner acquisition campaign (H^D_phantom) to simulated noiseless data. This demonstrates that LDA can exploit the specific noise population used during training, but it does not test performance under real forward measurements, operator error, drift, or non-additive disturbances. The authors should either temper the abstract and conclusion or add out-of-distribution experiments, for example noise from a different acquisition session, system-matrix mismatch with A_rec different from A_coarse_opt, or real MPI data. The current evidence supports a claim about the MPI-MNIST benchmark, not a claim about MPI scanner noise in general.","section":"Section 6.2 and Section 5.4.3"},{"comment":"The regularization parameters are selected on the test set: the grid search uses the first 100 images of the test dataset for Tikhonov, regularized Kaczmarz, and whitened regularized Kaczmarz, and the first 5 images for LDA, with SSIM as the selection criterion. The reported means over the full test set are therefore not independent generalization results, and the imbalance in tuning budget between LDA and the classical methods could bias the comparison. The authors should use a separate validation set for parameter selection, or explicitly present the current numbers as test-set-tuned performance rather than predictive performance on unseen data.","section":"Section 6.1"}],"minor_comments":[{"comment":"The text contains two typos: 'noise sampes' in Section 4 should be 'noise samples', and 'seperate' in Section 5.4.2 should be 'separate'.","section":"Section 4 and Section 5.4.2"},{"comment":"The caption in Figure 2 reads 'PNSR' but should read 'PSNR'.","section":"Figure 2 caption"},{"comment":"Please specify the gradient descent step size and stopping criterion used for the LDA minimization. With 2000 iterations and RK initialization, the optimization protocol is part of the method and is not fully reproducible from the text.","section":"Section 6.1"},{"comment":"The sentence defining the entries w_j,j uses std(n_j), which should be std(eta_j) for consistency with the notation for the noise components introduced in the same equation.","section":"Section 2.1, Eq. (2.3)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the dataset is a genuine community resource and the topic fits the journal. I would be comfortable with major revision rather than rejection, provided the sign error in Eq. (2.10) is corrected and the in-distribution nature of the experiments is stated clearly. No concerns about attribution or citation behavior."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper before reading further: the MPI-MNIST dataset is a real community resource, and the central equation in the method definition is wrong as printed. The sign error is not cosmetic.\n\nThe LDA idea is sensible. Deriving the data discrepancy from a MAP estimate with a normalizing-flow noise model is a clean way to escape Gaussian assumptions. The change-of-variables derivation in Eqs. (2.6)–(2.9) is standard, and putting the learned log-likelihood into the objective is a legitimate move. The dataset itself is the more tangible contribution: 70,000 simulated MPI measurements from MNIST phantoms, with model-based system matrices at multiple resolutions and roughly 470,000 real empty-scanner noise samples. That is useful for benchmarking and for training learned methods.\n\nNow the soft spots, in decreasing order of severity.\n\nEquation (2.10) reads xLDA = arg min Dθ(Ax,yδ) − R(x). But (2.6) gives arg min[−log p(yδ|x) − log pX(x)], and the text defines R(x) = −log pX(x). Substituting the flow likelihood yields Dθ + R(x), not Dθ − R(x). The printed objective is not the MAP estimate. Worse, with R(x)=α/2||x||², the minus sign makes the objective unbounded below for typical α; gradient descent could not produce the stable reconstructions in Table 1. So the implementation almost certainly used the plus sign, and the paper as written does not describe the evaluated method. A reader implementing Eq. (2.10) literally would either fail to reproduce the results or silently correct the sign and not test the published method. This is a correctness defect at the center of the method definition.\n\nThe rest is more standard. The regularization parameter for every method is chosen by grid search on the test set, with LDA tuned on just five test images. The discrepancy network is trained and tested on noise samples from the same scanner population, so the reported gains are in-sample. The paper acknowledges the lack of real-data validation in the conclusion. There is no code released. The computational cost is substantial: 2000 iterations and ~90 seconds per image at low concentration, versus fractions of a second for Kaczmarz. These are not deal-breakers, but they temper the headline claims.\n\nBottom line: the dataset and the flow-based discrepancy idea deserve a serious referee. The sign error has to be fixed and the authors should be asked to confirm the implemented objective, ideally by releasing code. If corrected, this is a solid contribution for the MPI and learned-inverse-problems communities.","headline":"Real dataset, clean idea, but the central equation has a sign error that makes the printed method different from the one actually evaluated.","tokens_in":22050,"tokens_out":7362,"would_cite":true,"duration_ms":62230,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J22","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing the Gaussian noise assumption with a learned flow-based discrepancy improves MPI reconstruction similarity at every tested tracer concentration.","keywords":["magnetic particle imaging","learned discrepancy","invertible neural network","normalizing flow","image reconstruction","inverse problems","MPI-MNIST","noise modeling"],"falsifier":"Apply LDA to real MPI measurements of a physical phantom with known concentration, using the same system matrix calibration and noise sampling procedure: if its structural similarity advantage over regularized Kaczmarz does not reproduce there, the learned noise distribution does not transfer. A controlled version would add a slow scanner drift or non-additive perturbation to the test data and check whether the reported gains disappear.","tokens_in":21022,"feed_emoji":"🧲","tokens_out":10910,"duration_ms":94663,"temperature":0.7,"pith_summary":"Magnetic particle imaging (MPI) reconstructs images from voltage measurements whose noise is not well described by the Gaussian model assumed in standard Tikhonov and Kaczmarz reconstructions. The paper introduces the Learned Discrepancy Approach (LDA), which replaces the usual squared-error discrepancy with one built from an invertible neural network trained on real scanner noise, and derives this objective from Bayesian maximum-a-posteriori estimation. It also introduces MPI-MNIST, a dataset of 70,000 simulated MPI measurements of handwritten digits with realistic measured scanner noise. Using that dataset, the paper reports that LDA attains the highest mean structural similarity at every tested tracer concentration, with the largest advantage at the lowest concentration (SSIM 0.9314 versus 0.8527 for regularized Kaczmarz at 2 mgFe/mL). This matters because the method provides a principled way to reconstruct when the noise model is unknown, and the new dataset gives other researchers the means to test it.","feed_headline":"Learned noise model lifts MPI image quality at every tested dose","feed_subtitle":"Replacing Gaussian noise assumptions with a flow-based model lifts reconstruction similarity where noise is highest.","key_machinery":"The discrepancy network $\\varphi_\\theta$ is the load-bearing object: a multi-scale invertible neural network built from affine coupling layers, trained by maximum likelihood on approximately 300,000 measured noise samples. Through the change-of-variables formula, it turns the unknown noise distribution into a Gaussian latent distribution, so the negative log-likelihood of the residual $y^\\delta - Ax$ becomes $\\frac{1}{2}\\|\\varphi_\\theta(y^\\delta - Ax)\\|^2 - \\log|\\det J_{\\varphi_\\theta}(y^\\delta - Ax)|$. The log-determinant term is what distinguishes this from a mere learned whitening: it keeps the likelihood exact for non-Gaussian noise, and it is what carries the claimed reconstruction improvements.","core_discovery":"Starting from the linear inverse problem $y^\\delta = Ax + \\eta$, the paper treats the noise density $p_H$ as unknown and represents it through a diffeomorphism $g$ using the change-of-variables identity $p_H(\\eta) = p_Z(g(\\eta))\\,|\\det J_g(\\eta)|$, where $p_Z$ is a known latent density. An invertible neural network $\\varphi_\\theta$ is trained on empty-scanner noise samples to approximate $g$, and substituting the learned density into the maximum-a-posteriori objective yields the LDA minimizer $x_{\\mathrm{LDA}} = \\arg\\min_x \\frac{1}{2}\\|\\varphi_\\theta(y^\\delta - Ax)\\|^2 - \\log|\\det J_{\\varphi_\\theta}(y^\\delta - Ax)| - \\log p_X(x)$. On the MPI-MNIST test set, the paper reports that this objective gives the highest mean structural similarity at all tested concentrations, with the largest gap at the highest noise level (SSIM 0.9314 for LDA versus 0.8527 for regularized Kaczmarz at 2 mgFe/mL). The method is presented as a post-processing refinement of a Kaczmarz reconstruction and as a general variational framework for linear inverse problems beyond MPI.","pith_inferences":["If the measured noise is actually non-additive or time-varying, the learned discrepancy would be misspecified; testing LDA on real phantom scans, rather than simulated data, is the direct next experiment.","The same learned-discrepancy idea could be paired with a learned prior, relaxing the Gaussian prior assumption in addition to the noise assumption.","The high iteration count makes LDA unsuitable for real-time imaging without acceleration, so a faster optimizer or a distilled approximation would be a natural engineering follow-up.","The dataset's multiple system matrices make it reusable for benchmarking operator-correction methods independently of noise-model learning."],"forward_implications":["MPI reconstruction no longer has to assume Gaussian noise; the discrepancy term can be fit to the scanner's measured noise.","Because LDA is formulated for general linear inverse problems, the same construction can be applied to other imaging modalities with structured noise.","MPI-MNIST gives the community 70,000 phantom-measurement pairs with several system-matrix models and noise levels, enabling reproducible algorithm comparison.","The reported gains are largest at low tracer concentration, where noise dominates, so LDA is most valuable in precisely the regime where classical methods struggle.","The cost is substantially more iterations and longer reconstruction times, which the paper identifies as a current limitation."],"supporting_citations":[{"why":"supplies the regularized Kaczmarz algorithm used as the classical baseline and as the initialization for LDA.","marker":"[51]"},{"why":"provides the whitened regularized Kaczmarz baseline and the data-based whitening idea that LDA generalizes.","marker":"[29]"},{"why":"underlies the Fokker-Planck-based simulation of model-based system matrices used for MPI-MNIST data generation.","marker":"[24]"},{"why":"supplies the numerical method for solving the magnetization dynamics in each pixel of the simulated system matrices.","marker":"[52]"},{"why":"provides the MNIST handwritten-digit images that are processed into the MPI-MNIST phantoms.","marker":"[50]"},{"why":"motivates the upsampling used to avoid an inverse crime when generating measurements from the phantoms.","marker":"[54]"},{"why":"defines the structural similarity index used for the quantitative comparison across methods.","marker":"[57]"},{"why":"supplies the multi-scale affine-coupling architecture used to build the discrepancy network.","marker":"[59]"}],"fun_headline_variants":["Learned noise model lifts MPI image quality at every dose","Invertible network learns MPI noise, boosts reconstruction","Noise-aware reconstruction beats Gaussian assumptions in MPI","Learned discrepancy reconstruction improves MPI image similarity","Flow-based noise model in MPI yields sharper image reconstructions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the empty-scanner noise samples used to train the discrepancy network faithfully represent the additive noise present during real MPI measurements, and that this noise is independent of the phantom.","fun_headline_variants_meta":{"raw":{"variants":["Learned noise model lifts MPI image quality at every dose","Invertible network learns MPI noise, boosts reconstruction","Noise-aware reconstruction beats Gaussian assumptions in MPI","Learned discrepancy reconstruction improves MPI image similarity","Flow-based noise model in MPI yields sharper image reconstructions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3244,"prompt_tokens":1031,"completion_tokens":2213,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":2138}},"tokens_in":647,"tokens_out":2213,"duration_ms":15479,"temperature":1.0,"reasoning_tokens":2138,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:00.956853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply LDA to real MPI measurements of a physical phantom with known concentration, using the same system matrix calibration and noise sampling procedure: if its structural similarity advantage over regularized Kaczmarz does not reproduce there, the learned noise distribution does not transfer. A controlled version would add a slow scanner drift or non-additive perturbation to the test data and check whether the reported gains disappear.","supporting_citations":[{"cited_title":"On row relaxation methods for large constrained least squares problems","cited_arxiv_id":null,"evidence_quote":"supplies the regularized Kaczmarz algorithm used as the classical baseline and as the initialization for LDA."},{"cited_title":"Enhanced reconstruction in magnetic particle imaging by whitening and random- ized SVD approximation","cited_arxiv_id":null,"evidence_quote":"provides the whitened regularized Kaczmarz baseline and the data-based whitening idea that LDA generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the numerical method for solving the magnetization dynamics in each pixel of the simulated system matrices."},{"cited_title":"The MNIST database of handwritten digits","cited_arxiv_id":null,"evidence_quote":"provides the MNIST handwritten-digit images that are processed into the MPI-MNIST phantoms."},{"cited_title":"Statistical and Computational Inverse Problems","cited_arxiv_id":null,"evidence_quote":"motivates the upsampling used to avoid an inverse crime when generating measurements from the phantoms."},{"cited_title":"Density estimation using Real NVP","cited_arxiv_id":null,"evidence_quote":"supplies the multi-scale affine-coupling architecture used to build the discrepancy network."}],"review_version":1}