{"id":"0ce6321d-55e8-445c-a3fe-103fb6fbb73d","arxiv_id":"2605.31426","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Regularized MID/s2ISM reconstruction with residual-whiteness-based automatic lambda selection yields stable ISM images without empirical early stopping.","lead":"This paper adds explicit regularization and automatic parameter tuning to standard image-scanning-microscopy reconstructions, replacing manual early stopping. The method stabilizes low-photon reconstructions on simulated and real microscopy data while preserving super-resolution and optical sectioning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RWP parameter selection relies on an unproven zero-mean residual assumption; with only one simulated validation and an unspecified s2ISM high-pass region, the 'self-tuning' claim is under-supported.","rationale":"The reader's weakest assumption correctly identifies the core of the paper's contribution: the residual whiteness principle must select a good regularization parameter without ground truth. The paper provides only a single simulated validation of this principle and no quantitative real-data measure. I examined alternative concerns—such as the unspecified s2ISM high-pass region, the noisy whiteness estimator with many large lags, and MD-l1's retained semi-convergence—but these are either secondary or explicitly acknowledged by the authors. The zero-mean residual assumption (Remark 1) is the linchpin: if it fails, the mathematical justification for minimizing W disintegrates, and the claimed elimination of empirical stopping loses its foundation. My proposed test would directly probe this assumption across photon levels and model mismatch, which are precisely the low-photon and real-data conditions where the paper claims improvement. Since the paper is a proof-of-concept and the concern is about insufficient validation rather than an identified flaw, the existing CONDITIONAL verdict remains appropriate.","tokens_in":29704,"tokens_out":6449,"duration_ms":70694,"concrete_test":"On a simulated dataset with known ground truth, vary the photon flux F over {5, 10, 20, 50} and add PSF mismatch by convolving with slightly perturbed PSFs (e.g., 5% width change). For each condition and 20 noise realizations, compute lambda* by minimizing the masked RWP (with a fixed, specified high-pass region for s2ISM) and compare to the oracle lambda that maximizes PSNR. Report the PSNR at lambda* relative to oracle, and the mean of the standardized residual field at lambda*. If the mean residual magnitude exceeds 0.1 or the PSNR gap exceeds 2 dB in any low-flux or mismatch condition, the RWP self-tuning claim is not supported in those regimes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that automatic lambda selection via the residual whiteness principle (RWP) replaces empirical early stopping. This hinges on the assertion in Section 6, Remark 1, that the standardized residual field Z_lambda is approximately centered (E[Z_lambda]~0) and that minimizing W(Z_lambda) yields a reconstruction close to ground truth. The only evidence is Figure 2 (mean of Z_lambda over 50 realizations for PGD-TV on one simulated ISM dataset) and Figure 4a (agreement between lambda* from W and lambda maximizing PSNR on one dataset). No proof is given, and no real-data quantitative validation exists. Moreover, the masked RWP (Section 6.1) conditions on zero-truncated Poisson variables; the derived moments are exact only if the model is correct, but at low photon counts or under PSF/background mismatch, the standardized masked residual is not guaranteed centered. The high-pass extension for s2ISM (Section 6.1.1) introduces an unspecified region B in frequency space, which is a manual tuning parameter that undermines the 'fully automatic' claim and makes the method non-reproducible. If the zero-mean assumption fails or W is not a faithful proxy, lambda* may be over- or under-regularized, and the user would again need manual adjustment, collapsing the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an explicitly regularized variational framework for image scanning microscopy (ISM) and its sectioning variant s2ISM, replacing the implicit regularization of Richardson–Lucy-type methods. Within a MAP formulation, the multi-frame Poisson negative log-likelihood is combined with ℓ1 or smoothed total variation regularization, and the regularization parameter is selected automatically by a residual-whiteness principle adapted to Poisson statistics, including a masked version and a high-pass spectral extension for s2ISM. First-order optimization algorithms (proximal gradient, projected gradient, and mirror descent) with adaptive backtracking are analyzed and tested on one simulated tubulin dataset and several real TOMM20 acquisitions.","tokens_in":29993,"tokens_out":6370,"duration_ms":72490,"significance":"If the automatic parameter-selection claim holds, the paper would make a practical and timely contribution to computational ISM: it would remove the need for empirical early stopping and manual regularization tuning, which are known practical obstacles in RL-based MID/s2ISM reconstruction. The strengths of the paper include the clean Bayesian variational formulation, explicit Lipschitz bounds for the Poisson data term and smoothed TV, a correct derivation of the zero-truncated Poisson moments used for the masked residual, the use of the publicly available BrightEyes-ISM simulator, and transparent reporting of algorithmic complexity. The central novelty, however, is the whiteness-based 'self-tuning' mechanism, and the current evidence for that mechanism is narrow: one simulated dataset for validation, no quantitative real-data verification, and a dependence on an unspecified high-pass region in the s2ISM case. These issues need to be addressed before the main claim can be considered established.","major_comments":[{"comment":"The central claim that minimizing the whiteness functional W(Z_λ) selects a near-optimal λ relies on the assumption E[Z_λ]≈0 and, more broadly, on W being a faithful proxy for reconstruction error. The paper itself flags this in Remark 1 as an assumption, not a theorem. The supporting evidence is limited to Figure 2 (mean residual over 50 realizations for PGD-TV on one simulated ISM dataset) and Figure 4a (agreement between λ* and PSNR-optimal λ for PGD-TV on one dataset). This is not sufficient to support the general claim that the method is 'fully automatic and ground-truth-free' across algorithms, regularizers, flux levels, and model-mismatch conditions. In particular, at low photon counts or under PSF/background mismatch, the standardized residual may be biased, and the selected λ may be systematically over- or under-regularizing. The authors should provide either a theoretical argum","section":"Section 6, Remark 1 and Eq. (26)"},{"comment":"The high-pass extension for s2ISM introduces a low-frequency mask region B centered at ω=0, but the size and shape of B are never specified. This makes the method irreproducible and introduces a manual tuning parameter, which contradicts the paper's stated goal of a fully automatic parameter-selection strategy. The authors should specify how B is chosen (e.g., a fixed fraction of the frequency domain, a data-adaptive rule, or a calibration experiment) and provide a sensitivity analysis showing that the selected λ does not critically depend on the choice of B. Without this, the 'self-tuning' claim for s2ISM is incomplete.","section":"Section 6.1.1, Eq. (32)"},{"comment":"The paper acknowledges that MD-ℓ1 'closely resembles' Richardson–Lucy and that MD-based schemes 'retain part of the transient semi-convergent behavior.' Yet the abstract and Section 8.1.4 state that the framework enables stable reconstructions without empirical early stopping. These statements are in tension: one of the four proposed algorithms still exhibits semi-convergence and may require a stopping iteration. The authors should either restrict the stability claim to the algorithms for which it holds, or provide a convergence/stopping analysis for MD-ℓ1 that explains why the RWP-selected λ and the convergence criterion (39) yield acceptable reconstructions despite the oscillatory error behavior. As written, the blanket claim is stronger than the reported results.","section":"Section 8.1.4, Fig. 4b and Section 7.1.1"},{"comment":"The convergence results in Appendix C are stated for fixed stepsizes satisfying a global Lipschitz/relative-smoothness condition. However, the implemented MD algorithms use the adaptive backtracking rule of Algorithm 5, and the paper states that the MD backtracking rule is adopted 'mainly from a practical perspective.' No convergence theorem is provided for the adaptive MD scheme, and the descent test in Algorithm 5 for the MD case (Δ = −(d/α)D_h(u_cand, u_k)) is not proven to guarantee sufficient decrease or convergence. If the paper claims that the proposed algorithms converge, the adaptive MD variant needs either a proof or an explicit statement that its convergence is empirical only.","section":"Section 7.2 and Appendix C"}],"minor_comments":[{"comment":"The smoothed TV functional is defined with +ε inside the square root, but the gradient formula in Eq. (16) uses ∥∇x∥²+ε² in the denominator. The Hessian proof in Appendix B similarly uses √(∥(∇x)_i∥²+ε²). The notation should be made consistent (likely the functional should read +ε², or the gradient denominator should use +ε).","section":"Section 5.2.2, Eq. (15)-(16)"},{"comment":"The empirical normalized autocorrelation in Eq. (21) does not normalize by the number of admissible pairs for each lag. The efficient FFT formula (23) relies on periodic boundary conditions, which is not stated in the main text and may be violated by the masked residual field (zeros on non-active pixels). Please clarify the boundary treatment and the effect of masking on the FFT-based whiteness evaluation.","section":"Section 6, Eq. (21)"},{"comment":"The knee-point detection procedure introduces another heuristic decision (signed orthogonal distance to the chord). While the examples in Figure 10 illustrate its behavior, a robustness analysis with respect to the grid density and normalization would strengthen the claim that the knee-point criterion does not bias the parameter choice.","section":"Section 9.4 and Appendix E"},{"comment":"There are several typographical and formatting issues: in Figure 8, decimal commas appear (e.g., '0,52' versus '0.52'), and the text contains OCR-like artifacts such as '2 /uni03BCμ'. These should be cleaned before publication.","section":"Throughout"},{"comment":"The validation of the RWP in Figure 4a is only for PGD-TV. It would be helpful to show the same W(λ) vs PSNR comparison for the other three algorithms in the supplementary material, since the selected λ depends on the optimization scheme and the manuscript claims the criterion works for all considered methods.","section":"Section 8.1.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and addresses a timely problem in computational microscopy. The variational formulation and optimization details are competently presented, and the adaptation of whiteness-based parameter selection to multi-frame Poisson ISM/s2ISM is a reasonable extension of prior work. However, the central 'self-tuning' claim is currently supported by very limited validation: one simulated dataset, one algorithm for the residual-mean check, and an unspecified high-pass region for s2ISM. I recommend major revision rather than rejection because the concerns are addressable with additional experiments, clarity, and, if possible, a limited theoretical analysis of the residual-whiteness criterion. I would also ask the authors to reconcile the MD-ℓ1 semi-convergence behavior with the abstract's blanket stability claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper does something useful: it turns the implicit early-stopping heuristic in MID and s2ISM reconstruction into an explicit MAP problem with TV or l1 priors, and adapts the residual whiteness principle to the multi-frame Poisson case, including a zero-truncated mask for low-photon data and a high-pass spectral mask for s2ISM. The optimization side is standard but competently done — Lipschitz bounds are explicit, convergence theorems are cited appropriately, and the adaptive backtracking is a practical improvement. The one simulation where lambda* from whiteness lands near the PSNR-best lambda is the key positive evidence, and they are honest in Remark 1 that the zero-mean residual assumption is empirical, not proven.\n\nSoft spots, in order of concern. First, the automatic-selection claim rests on a single simulated dataset. The 50-realization plot in Figure 2 addresses the zero-mean assumption, but not the more important question of whether minimizing W(Z_lambda) reliably selects good lambda across different flux levels, PSF mismatch, and background errors. Second, the s2ISM high-pass region B is never specified; it is a manual tuning knob that undercuts the 'fully automatic' claim and makes the experiments hard to reproduce. Third, real-data results are qualitative; that is acceptable for a methods paper, but it means the practical benefit is asserted more than demonstrated. Fourth, the MD backtracking has no theory and the MD-l1 variant behaves like RL, which they acknowledge; that is a minor issue since PGD is the recommended workhorse. No code is provided, which compounds the reproducibility problem.\n\nNone of this is fatal. The central stabilization result — regularized methods run to convergence without the semi-convergence blowup that RL exhibits — is demonstrated clearly in both simulation and real data. The whiteness principle could fail under severe model mismatch or at very low counts, but the paper flags that and gives one piece of empirical support. I'd want to see more before betting a pipeline on it, but it is a legitimate contribution.\n\nWho should read it: anyone working on computational ISM, and people in Poisson inverse problems who care about unsupervised parameter selection. It deserves a serious referee, not a desk reject. I'd recommend conditional acceptance after the authors release code, specify B, run a broader simulation study with multiple noise realizations and error bars, and compare against at least one alternative selection rule.","headline":"A legitimate and well-executed variational extension of MID/s2ISM with automatic parameter selection, but the core whiteness heuristic needs broader validation before it can be called fully self-tuning.","tokens_in":30525,"tokens_out":3111,"would_cite":false,"duration_ms":30891,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J22","68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that explicitly regularized MID and s2ISM reconstructions, with the regularization weight chosen automatically by a masked residual-whiteness principle, produce stable, high-quality images in low-photon fluorescence micros","keywords":["Image Scanning Microscopy","regularization parameter selection","residual whiteness principle","Poisson inverse problems","Richardson-Lucy semi-convergence","total variation regularization","optical sectioning","super-resolution microscopy"],"falsifier":"Simulate ISM data with the same setup but a flux factor well below F=20 (e.g., F=2), run reconstructions over the full λ grid, and test whether the λ minimizing the masked whiteness functional coincides with the λ maximizing PSNR/MicroSSIM against the known ground truth. Any systematic divergence in this low-photon regime, where the zero-truncation correction dominates, would falsify the claim that the self-tuning rule selects the reconstruction closest to the truth.","tokens_in":29563,"feed_emoji":"🔬","tokens_out":5370,"duration_ms":53543,"temperature":0.7,"pith_summary":"The paper takes aim at the semi-convergence problem of Richardson–Lucy-based ISM reconstruction, where iterative deconvolution first improves and then progressively amplifies noise. It proposes a Bayesian MAP formulation that adds an explicit ℓ1 or smoothed total-variation prior to the multi-frame Poisson data fidelity term, and a fully automatic, ground-truth-free rule for setting the regularization weight: choose the weight that makes the standardized residual field look as white as possible. On simulated and real fluorescence data, this self-tuning rule selects parameters close to the PSNR-optimal values and yields reconstructions that no longer depend on choosing a stopping iteration. The upshot is that a practical bottleneck of MID/s2ISM—manual early stopping—can be replaced by a principled variational objective, making super-resolution and optical sectioning usable in low-photon conditions.","feed_headline":"Whiteness rule auto-tunes ISM deconvolution, no early stopping","feed_subtitle":"A masked whiteness criterion picks the regularization weight from data alone, enabling stable ISM reconstruction.","key_machinery":"The load-bearing object is the whiteness functional W(Z) = Σ ĉ_{l,m}(Z)², computed from the standardized residual field Z_λ = (y − ν_λ)/√ν_λ via an FFT-based expression. Minimizing W over a grid of λ values selects the regularization weight; for s2ISM, W is evaluated on a high-pass-filtered power spectrum to prevent unregularized background errors from dominating, and masked residuals correct for the zero-truncated Poisson statistics of dark pixels. The reconstructions themselves are minimizers of the KL data fidelity term plus λR(x), with R either ℓ1 or smoothed total variation, solved by proximal gradient or mirror descent with adaptive backtracking stepsize selection.","core_discovery":"The central claim is that the residual whiteness principle, adapted to multi-frame Poisson data with masking for zero photon counts and a high-pass spectral correction for the two-plane s2ISM model, provides a valid unsupervised selector for the regularization parameter λ in explicitly regularized MID and s2ISM reconstruction. Under this choice, the authors report that the regularized reconstruction at convergence matches or exceeds the quality of Richardson–Lucy with carefully chosen early stopping, suppresses out-of-focus background, and avoids the characteristic noise amplification of late RL iterations. The validation includes simulated tubulin data, where the whiteness minimizer closely","pith_inferences":["Inference: if the whiteness heuristic is accepted, it could serve as a general parameter-selection or stopping principle for other Poisson inverse problems in microscopy—such as single-molecule localization or light-sheet deconvolution—not just ISM.","Inference: the paper's validation of the zero-mean residual assumption is a single simulation at one flux level; a natural stress test is to check whether the W minimizer still tracks the PSNR-optimal λ as photon flux drops toward zero counts in most pixels, where the zero-truncation correction dominates.","Inference: the high-pass filtering for s2ISM may mask systematic low-frequency model error rather than correct it; if the background PSF is misspecified, the criterion could select an over-regularized foreground while still producing white-looking high-frequency residuals.","Inference: a testable practical extension—mentioned by the authors as future work—is to optimize the whiteness functional directly in a bilevel framework, which would remove the grid search over 150 candidates and make the self-tuning procedure substantially faster."],"forward_implications":["MID and s2ISM reconstruction can be run to convergence without an empirically chosen stopping iteration, removing a major reproducibility bottleneck.","Low-photon acquisitions become practical: explicit regularization plus automatic λ yields stable super-resolution and optical sectioning where Richardson–Lucy would amplify noise.","The parameter-selection machinery is not tied to TV or ℓ1; it extends to other priors, including learned or plug-and-play regularizers, so the same whiteness principle could automate them.","Because the criterion is ground-truth-free and computed from the same data, it can be deployed in routine microscopy pipelines without calibration images.","Not every algorithm variant benefits equally: mirror-descent ℓ1 retains a multiplicative dynamics close to Richardson–Lucy and still shows semi-convergent oscillations, so the early-stopping problem is solved more convincingly for proximal-gradient variants."],"fun_headline_variants":["Whiteness principle auto-selects ISM deconvolution penalty","Data-driven λ from whiteness: ISM without early stopping","Self-tuning regularization: ISM stable at convergence","Whiteness test picks ISM regularization weight blindly","No early stopping: whiteness tunes ISM deconvolution"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the standardized residual field is approximately centered (mean zero) and that minimizing its autocorrelation selects a reconstruction close to the ground truth; the paper verifies this empirically with one simulation but provides no proof and no quantitative check on real data.","fun_headline_variants_meta":{"raw":{"variants":["Whiteness principle auto-selects ISM deconvolution penalty","Data-driven λ from whiteness: ISM without early stopping","Self-tuning regularization: ISM stable at convergence","Whiteness test picks ISM regularization weight blindly","No early stopping: whiteness tunes ISM deconvolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000401,"raw_usage":{"total_tokens":1959,"prompt_tokens":804,"completion_tokens":1155,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1089}},"tokens_in":548,"tokens_out":1155,"duration_ms":8300,"temperature":1.0,"reasoning_tokens":1089,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T12:43:49.362800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate ISM data with the same setup but a flux factor well below F=20 (e.g., F=2), run reconstructions over the full λ grid, and test whether the λ minimizing the masked whiteness functional coincides with the λ maximizing PSNR/MicroSSIM against the known ground truth. Any systematic divergence in this low-photon regime, where the zero-truncation correction dominates, would falsify the claim that the self-tuning rule selects the reconstruction closest to the truth.","supporting_citations":[],"review_version":2}