{"id":"a9bc4536-ee30-4cf3-a4df-f8f74b568b48","arxiv_id":"2506.07482","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A Fisher information-based framework optimizes the phase biases in sensorless adaptive optics, yielding lower residual aberrations than heuristic designs in simulated multiphoton microscopy.","lead":"This paper uses Fisher information to calculate how much wavefront information different phase biases provide, then optimizes those biases to improve sensorless adaptive optics correction. The framework is tested in simulations of multiphoton microscopes and produces more accurate corrections than heuristic bias choices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CRLB-optimized bias designs may be overfit to the fixed object structure; the claimed generality of the framework is not tested.","rationale":"The paper presents a coherent framework and supports it with internally consistent simulations: the CRLB computation, the optimization of biases, and the ML-based evaluation all share the same forward model, and the results show that CRLB-guided biases reduce residual aberration in the tested scenarios. The reader's CONDITIONAL verdict is appropriate, but the stated weakest assumption—pixel independence—is not the most load-bearing. Given the Poisson generative model, pixel independence is a legitimate modeling assumption. The more serious gap is that the optimization and evaluation are performed with a fixed object (or at most one object per scenario). The claimed generality of the framework depends on the optimized biases being robust to object variation, which is not demonstrated. A concrete simulation with randomized, unseen object structures would settle this concern. Because the potential issue is a missing validation rather than a demonstrated internal inconsistency, the verdict should remain conditional: the paper is valuable but should be revised to address object dependence before broad claims of method-agnostic and modality-agnostic applicability are accepted.","tokens_in":11150,"tokens_out":6996,"duration_ms":99869,"concrete_test":"Re-run the Section 2 bias-mode optimization and evaluation with a varied object ensemble. For each training example, draw a new volumetric object realization (e.g., random sub-resolution bead distributions or a random texel pattern) while keeping the aberration distribution, photon budget, and ML estimator fixed. Then evaluate the CRLB-optimized biases on a validation set of unseen object realizations, comparing against single-Zernike-mode biases and the published astigmatism-like optimized bias. If the optimal bias modal composition shifts substantially, or if the optimized biases no longer outperform the heuristic biases for unseen objects, the generality claim is refuted; if the advantage persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that minimizing the CRLB loss (Eq. 8) over bias aberrations yields sensorless AO configurations that transfer beyond the synthetic settings used. In the presented simulations, the object O appears to be fixed for each analysis: the optimization dataset is generated by randomizing system aberrations only, and the Fisher information is computed from the image model of Eq. 7 while treating the object structure as a known part of the forward model. The loss is averaged over aberrations but not over object realizations. Consequently, the derived biases could be tuned to the spatial-frequency content of one particular object, which would not generalize to the realistic situation where the sample is unknown and varies. The paper's own conclusion recognizes this class of problem when it notes that defocus can produce high Fisher information from sample-structure variation rather than from aberration information, and then excludes defocus for volumetric objects; however, the same confound may affect the other optimized modes. There is no evidence that the optimized biases remain optimal for unseen object structures. Notably, the reader's pixel-independence concern is not the weakest point: under the stated Poisson observation model (Eq. 7), conditional on the object and aberration, pixel counts are independent, so that assumption is exact within the model. The more consequential and unverified assumption is the fixed, known object used in the CRLB computation and in the learning-based evaluations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a Fisher information-based framework for optimizing bias aberrations in image-based sensorless adaptive optics (AO). The authors model the microscope imaging process with a scalar Fourier optics forward model and a Poisson noise distribution, then compute the Cramér–Rao lower bound (CRLB) from the Fisher information matrix. They define a loss function (Eq. 8) that sums the square-root CRLBs for the aberration coefficients, and minimize it over the bias aberrations using gradient descent. The framework is demonstrated in three simulation studies: (1) optimal bias amplitude for a 2N+1 parabolic fitting method in two- and three-photon microscopes; (2) optimal bias mode shapes for point-like versus volumetric objects, with validation using a neural-network estimator; and (3) optimal number of bias measurements under a fixed total photon budget. The simulations show that the CRLB-optimized biases generally lead to lower residual errors than heuristic bias choices, supporting the central claim that information-guided design can improve sensorless AO accuracy and efficiency.","tokens_in":11427,"tokens_out":7655,"duration_ms":90323,"significance":"If the framework holds, it provides a general, method-agnostic procedure for designing sensorless AO correction strategies, replacing heuristic choices with a principled optimization based on statistical information theory. The paper's strengths include carefully constructed simulations, a reasonable forward model for multiphoton microscopy, and the use of an independent neural-network estimator as a validation tool. The results are encouraging in most scenarios, and the observation that optimized bias modes resemble defocus for point-like objects and astigmatism for volumetric objects is consistent with earlier findings. However, the generality of the framework is not fully established because all validations use the same fixed object structures and the same simulation model, and there is no experimental confirmation.","major_comments":[{"comment":"The text states that random aberrations were generated following a uniform distribution in an n-sphere as defined by Ref. [26], and that 'any random combination of these five modes within the selected aberration RMS range had an equal opportunity to be generated.' Ref. [26] (Marsaglia 1972) describes sampling from the surface of a sphere, not the solid ball. Uniform sampling from the interior of a ball requires multiplying the surface directions by U^{1/K} (with K the number of modes). The supplementary information repeats this for the neural-network training data, saying 'randomly generated from a n-sphere distribution [3] and its RMS value was between 0 and 3 rad.' If the code implemented only the surface method, then the aberration distribution would be concentrated on the sphere of maximum RMS, and low-aberration cases would be underrepresented. Please clarify the exact sampling procedure; if the interior was intended, correct the reference or add the radial scaling, and if the simulations were run with surface-only sampling, re-run the analyses to confirm that the conclusions are unchanged.","section":"Method, Optimization process; Supplementary Sec. 4"},{"comment":"The Fisher information in Eq. (7) is computed for a fixed object O (point-like or a specific volumetric object). The loss in Eq. (8) is averaged over random system aberrations but not over object realizations, so the optimized bias aberrations could be tuned to the spatial-frequency content of that particular object. The paper's own conclusion identifies this class of problem when it notes that defocus can produce high Fisher information from sample-structure variation rather than from aberration information, and for this reason defocus is excluded for volumetric objects. The same confound may affect the other optimized modes, but no test is reported. To support the claimed generality, please repeat the optimization of Section 2 (and, if feasible, Section 3) for several structurally different objects, or average the CRLB over an ensemble of objects, and show that the optimized bias modes and the residual errors are robust to object choice.","section":"Method, Adaptive microscope numerical simulation and Optimization process; Results, Section 2; Conclusion"},{"comment":"The independent check of the CRLB optimization is the residual error of a machine-learning estimator. However, that estimator is trained and evaluated on images synthesized from the same forward model (Eq. 7) used for the CRLB computation, with the same object and the same aberration distribution. The shared simulation model means the comparison does not test sensitivity to model mismatch, such as detector pixel correlations, PSF model errors, or out-of-focus background that does not follow the assumed Poisson statistics. A demonstration with an independent forward model (e.g., a vectorial PSF model or experimental data) would materially strengthen the claim that the framework is method-agnostic and generally applicable.","section":"Results, Section 3, Fig. 4; Method, Eq. (7)"}],"minor_comments":[{"comment":"The phrase 'Results suggested that' should be 'Results suggest that' for grammatical consistency with the present-tense summary.","section":"Abstract"},{"comment":"The vertical dashed lines marking the optimized bias amplitudes are not included in the legend or described in the caption; please label them or add a sentence in the caption for clarity.","section":"Fig. 2"},{"comment":"The term 'n-sphere' is ambiguous; if the uniform distribution is over the interior (the ball), use 'n-ball' consistently throughout the paper and the supplementary information.","section":"Method and Supplementary"},{"comment":"The normalization factor for the Zernike polynomials is not displayed explicitly; please include the full expression for the radial polynomial normalization to aid reproducibility.","section":"Supplementary, Eq. (S1)"},{"comment":"The statement that 'the pixel size was set such that two pixel-widths matched the full-width-half-maximum of the PSF' is a useful sampling rule, but the exact pixel size in physical units would help readers compare with their own systems.","section":"Method, after Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a simulation-only study with broad claims of generality. The most significant concerns are the ambiguous aberration sampling method (surface vs. ball) and the lack of object-structure variation in the optimization and validation. These are addressable with additional simulations, but they are load-bearing for the central claim. The work is well-aligned with the journal's scope in adaptive optics, and the topic is timely. I would encourage the authors to strengthen the generalization evidence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is new and worth taking seriously: instead of picking bias aberrations for sensorless AO by heuristic, the authors minimize a CRLB loss derived from Fisher information. The three simulation studies are carefully built, and the results make sense—defocus for point-like objects, astigmatism for volumetric objects, and more biases under a fixed photon budget. The agreement with previously reported heuristics is a good sanity check, and the framework is presented as method-agnostic, which is fair.\n\nWhat the paper does well: the imaging model is explicit, and the Fisher information computation is exact under the stated Poisson observation model, conditional on the object and aberration. So the pixel-independence worry the reader raised is not actually a flaw within the model; that assumption is correct there. The authors also show good judgment in excluding defocus for volumetric objects, since defocus-induced changes contain sample-structure information rather than aberration information. In Section 2, using the neural network's residual error as an independent test is the right kind of evaluation, because it tests a real estimator rather than just the bound.\n\nThe soft spots are real but not fatal. The biggest one is that the optimization is done for a fixed object structure. The loss is averaged over random system aberrations but not over object realizations, so the optimized biases could be locked onto the spatial-frequency content of one particular sample. The paper does not test whether they transfer to unseen object structures. This is not a theoretical quibble—the stated goal is a practical framework for unknown samples. The authors' own discussion of the defocus confound shows they are aware of this category of problem, but they don't address it for the other modes. A second issue is that the CRLB comparison in Fig. 3b is partly circular: the optimized modes were designed by minimizing that exact loss, so their lower CRLB is guaranteed by construction. The neural-network residual error is the non-circular evidence, and it supports the claim, but the paper should separate the two more clearly. Third, there is no experimental validation and no code or data release; “available upon request” is weak for a methods paper that is entirely simulation-based. The Fig. 2d mismatch is minor because the residual curve is flat there, and the authors explain it.\n\nOverall, this is a solid simulation-based methods paper that will be useful to anyone designing sensorless AO strategies, especially for machine-learning-based approaches. It deserves a serious referee. A good review should push on object generalization and reproducibility. I would engage with it.","headline":"Fisher-information optimization of bias aberrations is a genuinely new and useful idea; the simulations support it, but the fixed-object overfitting risk and missing experimental validation are the real caveats.","tokens_in":11930,"tokens_out":2064,"would_cite":true,"duration_ms":26425,"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":"Fisher information analysis can systematically improve sensorless adaptive optics by optimizing the bias aberrations used during image acquisition, as demonstrated in simulated multiphoton microscopy.","keywords":["adaptive optics","sensorless adaptive optics","Fisher information","Cramér-Rao lower bound","bias aberration optimization","multiphoton microscopy","phase diversity","machine learning"],"falsifier":"In a simulation where the observed image is passed through a small smoothing kernel after Poisson noise generation—making pixel noise spatially correlated—compute the CRLB-optimal biases under the independence assumption and compare the residual aberration of the resulting sensorless AO correction with a heuristic bias set; if the optimized set does not outperform or underperforms, the independence assumption is responsible.","tokens_in":10961,"feed_emoji":"🔬","tokens_out":4340,"duration_ms":48416,"temperature":0.7,"pith_summary":"The paper claims that the performance of image-based sensorless adaptive optics (AO) can be systematically improved by treating the choice of bias aberrations, the known phase modulations applied during image acquisition, as an information-optimization problem. Using Fisher information and the Cramér–Rao lower bound (CRLB) to quantify how much each bias setting reveals about unknown aberration coefficients, the authors minimize a CRLB-based loss over possible bias configurations. In three simulated multiphoton microscopy scenarios, the optimized bias amplitudes, bias modes, and bias counts yield lower residual aberration than heuristic designs. If correct, the framework offers a general, method-agnostic route to designing sensorless AO correction that currently relies on empirical tuning.","feed_headline":"Fisher information finds better sensorless AO settings","feed_subtitle":"Simulations show CRLB-optimized bias aberrations cut residual error versus heuristic designs in multiphoton microscopy.","key_machinery":"The key machinery is the Fisher information matrix for the image-formation model, computed under the assumption that every pixel is an independent Poisson observation, and the resulting CRLB, which bounds the variance of any unbiased aberration estimator. The optimization pipeline builds a digital twin of the microscope, generates simulated images for a dataset of random aberrations, and minimizes the loss $\\sqrt{\\sum_k \\mathrm{CRLB}_{\\hat{\\theta}_k}}$ over the bias coefficients $a_k$ using gradient descent. The CRLB thereby acts as an objective proxy for estimation accuracy that does not depend on the particular estimator, while the paper extends the analysis to specific estimators (parabolic fitting and a ResNet-based network) to check that CRLB-optimized settings remain effective in practice.","core_discovery":"The central claim is that bias aberrations can be optimized by minimizing the sum of CRLBs for the aberration coefficients to be estimated, and that the resulting settings are better conditioned for sensorless AO than heuristic choices. The paper demonstrates this for a 2N+1 parabolic fitting intensity method (optimized bias amplitude per mode), for machine-learning-based methods with two or more input images (optimized bias modal shapes), and for fixed photon-budget comparisons where the number of measurements varies. In each case, the CRLB-optimized biases lead to lower simulated residual aberration than the conventional choices, and the CRLB loss curve tracks the simulated residual error in most tested regimes. The paper also notes a qualitative match with earlier findings that astigmatism is a good bias for volumetric samples and defocus for point-like samples, while highlighting cases where the parabolic approximation breaks down, such as in three-photon microscopy with large aberrations.","pith_inferences":["The same CRLB-optimization idea could be extended to optimize other free parameters of an AO loop, such as the set of Zernike modes to correct, the order of image acquisition, or adaptive (closed-loop) bias updates, beyond the static bias sets considered here.","In practice, sample structure and background fluorescence vary across fields of view, so an interesting extension would be to compute object-adaptive bias settings from a quick pre-scan, using the Fisher information framework conditioned on the measured object.","The paper's independence assumption could be relaxed by including a noise correlation model in the Fisher information computation; the framework would still apply, but the optimal biases would likely shift, and the comparison against heuristic designs would be a useful test of robustness.","The result that astigmatism-like biases suit volumetric samples while defocus-like biases suit point-like samples suggests a general rule of thumb for sensorless AO: match the bias shape to the axial extent of the sample, which could be verified experimentally on biological specimens."],"forward_implications":["Sensorless AO methods that currently choose bias amplitudes and modes empirically can be re-designed using the CRLB optimization, potentially reducing the number of images needed for a given correction accuracy.","Machine-learning-based sensorless AO, which often uses two or a few biased images, can adopt CRLB-optimized bias modes to condition the input data better than standard Zernike modes.","Under a fixed photon budget, the framework provides a principled way to balance the number of measurements against per-image exposure time, favouring more measurements in the regimes tested.","The framework is not tied to a particular AO method or imaging modality, so it can be transferred to other sensorless AO systems such as confocal, light-sheet, or ophthalmoscopic setups.","The CRLB loss can serve as a diagnostic to compare different sensorless AO algorithms without running the full estimation loop."],"supporting_citations":[{"why":"Supplies the 2N+1 parabolic fitting sensorless AO method used as the first testbed.","marker":"[5]"},{"why":"Defines Fisher information, the core statistical measure used in the framework.","marker":"[23]"},{"why":"Provides the Cramér–Rao lower bound that links Fisher information to estimator variance.","marker":"[24]"},{"why":"Introduces a deep-learning sensorless AO method using two biased images, serving as a baseline for the bias-count comparison.","marker":"[31]"},{"why":"Earlier neural-network control work that suggested astigmatism as a good bias mode; the paper extends this with a more fundamental analysis.","marker":"[34]"},{"why":"Establishes the minimum number of biases (N+1) needed to span N aberration coefficients, used as one of the compared schemes.","marker":"[35]"}],"fun_headline_variants":["CRLB-optimized bias sharpens sensorless AO","Fisher info guides better sensorless AO choices","Information theory improves adaptive optics correction","Optimal bias aberrations from Fisher information","Sensorless AO gains from information-guided bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The calculation treats every pixel in every acquired image as an independent Poisson random variable, so the Fisher information and CRLB reflect only that idealized noise model; if real detectors or optical systems introduce pixel correlations, the optimized biases may not be the best for the actual imaging system.","fun_headline_variants_meta":{"raw":{"variants":["CRLB-optimized bias sharpens sensorless AO","Fisher info guides better sensorless AO choices","Information theory improves adaptive optics correction","Optimal bias aberrations from Fisher information","Sensorless AO gains from information-guided bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1245,"prompt_tokens":860,"completion_tokens":385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":476,"tokens_out":385,"duration_ms":4790,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:32:39.820634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a simulation where the observed image is passed through a small smoothing kernel after Poisson noise generation—making pixel noise spatially correlated—compute the CRLB-optimal biases under the independence assumption and compare the residual aberration of the resulting sensorless AO correction with a heuristic bias set; if the optimized set does not outperform or underperforms, the independence assumption is responsible.","supporting_citations":[],"review_version":1}