{"id":"f3c3947e-5a96-45a2-8e70-d5be7524be6a","arxiv_id":"2504.14157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DeepPD jointly recovers object and phase aberration from five phase-diversity images, replacing the linear mirror calibration with a learned neural mirror model that improves reconstruction quality.","lead":"DeepPD reconstructs the blur-causing wavefront and the true sample from just five microscope images, using a neural network that learns how the deformable mirror actually bends light. It produced sharper reconstructions than earlier phase-diversity methods on calibration slides and stained cells, especially under large aberrations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The mirror model's training labels come from the same phase-diversity reconstruction it is meant to improve; without quantitative independent wavefront validation, the reported gain over the linear-mirror variant may be self-consistency rather than physical accuracy.","rationale":"We read the paper in good faith. The method is clearly described, the implementation is plausible, and the qualitative comparisons on the Argo-HM slide and myosin filaments show visible improvements consistent with the central claim. The runtime is reasonable, and the use of only five images is an improvement in acquisition burden. However, the load-bearing premise is that the learned mirror model accurately predicts the wavefront produced by the deformable mirror for the voltages used as diversities. The paper itself flags that training relies on phase estimates from the same phase-diversity reconstruction (Sec. 4.3.1) and that the only independent validation, a Shack-Hartmann sensor, is qualitative (Fig. 2d). This is precisely the weakest link: if the training labels carry algorithmic bias, the neural mirror model can become a learned version of that bias, and the improved reconstruction may merely be the algorithm agreeing with itself. The proposed simulation test is decisive because it provides ground truth for the mirror response and applies the same training loop; it can determine whether the iterative scheme is capable of recovering a genuine nonlinear mapping or only a self-consistent one. We also note the absence of code and quantitative error bars, but those are secondary to the validation issue. We therefore agree with the reader's assessment that the paper should be accepted only with additional validation, and we do not change the verdict.","tokens_in":13309,"tokens_out":7425,"duration_ms":61231,"concrete_test":"Simulate a ground-truth nonlinear deformable-mirror model (e.g., nonlinear influence functions with cross-coupling). Generate the same number and type of training phase-diversity stacks as in Sec. 4.3.1 (3,930 sets) with known phase and object, then run the exact iterative training protocol of Sec. 4.3.4 (initial linear model, three cycles of reanalysis and retraining). Compare the learned voltage-to-phase network to the ground-truth map on a held-out set of voltages using RMS wavefront error. If the learned map matches the true map within the tolerance implied by the qualitative SH agreement, the self-referential loop is validated; if it instead converges to a map reproducing the linear model or the reconstruction's systematic errors, the claimed improvement is confounded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DeepPD accurately reconstructs phase and object from five images and improves over prior methods depends on the learned mirror model (Sec. 2; Sec. 4.3) predicting the true wavefront for voltages used as phase diversities. Training labels for the voltage-to-phase network are phase estimates obtained by the same neural-representation phase-diversity analysis that DeepPD is designed to improve (Sec. 4.3.1; Sec. 4.3.4). Any systematic bias in those estimates, whether from the initially linear mirror assumption, from noise, or from algorithmic regularizers, can be learned by the mirror model. The three-cycle iterative retraining can reach a self-consistent but physically incorrect fixed point. The only external check is qualitative comparison to a Shack-Hartmann sensor for two voltage sets (Fig. 2d); the authors state that quantitative comparison 'provides limited value' and that pupil diameters cannot be matched directly. The Discussion further notes that calibration is reliable only within the range of phase estimates that the reconstruction can retrieve, but does not establish that those estimates are accurate absolutely. If the mirror model is not physically accurate, the reported advantage of DeepPD over NNRs (Figs. 3-5) could reflect fitting to the reconstruction algorithm's biases rather than genuine optics modeling. This is the most load-bearing uncertainty in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DeepPD, a phase-diversity reconstruction framework that jointly estimates the object and phase aberration using neural representations of both quantities, with a learned deformable-mirror model replacing the usual linear influence-function calibration. The method uses five images: one aberrated image and four images with astigmatism phase diversities. The mirror model consists of voltage-to-phase and phase-to-voltage networks trained on 3,930 phase-diversity stacks recorded on bead samples, using phase estimates obtained from the same neural-representation phase-diversity analysis, refined over three cycles. The authors evaluate DeepPD on an Argo-HM calibration slide and on immunolabeled myosin in fixed PtK2 cells, comparing against Gauss-Newton, Poisson, and neural-representation baselines with a linear mirror model (NNRs). They report improved object-reconstruction metrics (SSIM, PCC, PSNR, DCT norm), particularly for RMS wavefront errors above roughly 100 nm, and qualitatively compare the learned mirror model to Shack-Hartmann measurements.","tokens_in":13588,"tokens_out":4270,"duration_ms":38035,"significance":"If the learned mirror model is physically accurate, DeepPD is a valuable contribution: it reduces the number of diversity images required for guidestar-free joint phase/object estimation, removes the Zernike-basis limitation of earlier methods, and explicitly addresses nonlinear deformable-mirror response. The writing is clear, the method is described in sufficient detail to be reproduced, and the authors compare against several relevant baselines. The central empirical claim, however, depends on the mirror model's physical accuracy, which is not quantitatively validated; the training labels for the mirror model come from the same reconstruction algorithm that DeepPD is meant to improve. Hence the current evidence supports a conditional assessment rather than a definitive one.","major_comments":[{"comment":"The mirror model is trained on phase estimates produced by the same phase-diversity reconstruction that DeepPD is designed to improve, and the three-cycle iterative retraining only re-analyzes the same training datasets with the previous mirror model. This can converge to a self-consistent but physically incorrect solution. The only external validation, Fig. 2d, is qualitative; the manuscript states that a quantitative SH comparison 'provides limited value' and that pupil diameters cannot be matched directly. Because the claimed advantage of DeepPD over the NNRs variant (linear mirror model) is the learned mirror model, this circularity is load-bearing. Please provide quantitative validation of the mirror model on held-out voltages against SH measurements (or a synthetic ground-truth phase) and report the phase error; this is needed to establish that the improvement over the linear model reflects physical accuracy rather than algorithmic self-consistency.","section":"Sec. 4.3.1, Sec. 4.3.4"},{"comment":"The quantitative comparison in Fig. 5 uses the recorded unaberrated image as the reference for SSIM, PCC, and PSNR, after convolving the object estimates with the microscope's PSF. The unaberrated image is not ground truth; it contains noise and possibly residual aberration, and the assumed PSF may not match the actual imaging system. This makes the reported differences in metrics difficult to interpret and could bias the comparison. Please evaluate on synthetic data with known ground truth and report error bars or confidence intervals for the metrics across the 121 datasets.","section":"Sec. 5 (Metrics) and Fig. 5"}],"minor_comments":[{"comment":"The sentence beginning 'whereas NeuWS requires around 100 images' should start with 'Whereas'.","section":"Sec. 2"},{"comment":"Equation (12) has an unbalanced parenthesis in the DCT-norm definition; please check the formula.","section":"Sec. 4.5, Eq. (12)"},{"comment":"The caption says the SH measurements use a zonal representation, but no scale or colorbar is provided for the phase maps, making the qualitative comparison hard to assess.","section":"Fig. 2d"},{"comment":"The paper does not state the number of random trials or seeds used for the neural-representation reconstructions; adding this would help assess the stability of the reported metrics.","section":"Sec. 4.2.6"},{"comment":"In the Discussion, 'that that leverages' has a duplicated word.","section":"Sec. 3"},{"comment":"The code availability statement says code will be released upon publication; for reproducibility, consider providing a version at submission.","section":"Declarations"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the circular validation of the mirror model. The authors may be able to address this with a few additional experiments (quantitative SH comparison or synthetic ground truth). The paper is otherwise well written and within scope for the journal. No concerns about authorship or citation integrity; note that a patent application is disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: DeepPD does something real—it replaces the standard linear influence-function model of the deformable mirror with a learned voltage-to-phase network inside a joint phase-and-object reconstruction from five phase-diversity images. The idea is straightforward, and the empirical comparison against the linear-mirror variant (NNRs) is the right control: if the learned mirror model didn't help, the comparison would show it. The gains at high aberration magnitudes are credible, and the paper is honest that quantitative Shack-Hartmann validation is limited and that the mirror calibration is only reliable within the range the phase-diversity analysis can retrieve.\n\nThe genuinely new piece is the mirror model training with cycle consistency. The rest—neural representations from NeuWS and the four astigmatism diversities from Johnson et al.—is a sensible combination, not a leap. The methods are described clearly enough to reproduce, and the 121-dataset comparison gives some weight to the robustness claim.\n\nThe soft spots are the ones you'd expect. Most load-bearing: the mirror model's training labels are phase estimates produced by the same phase-diversity analysis that DeepPD is meant to improve. The three-cycle retraining can converge to a self-consistent but physically wrong fixed point. The Shack-Hartmann comparison in Fig. 2d is qualitative, and the authors explicitly say quantitative comparison provides limited value. That's the weakest link. The metrics in Fig. 5 are reported as distributions without error bars or statistical tests, and the reference image for SSIM/PCC/PSNR is the recorded unaberrated image, which is not ground truth. These are addressable: independent wavefront measurements on a subset of mirror training voltages, confidence intervals, and code release would firm it up.\n\nThat said, the circularity concern doesn't sink the paper. The comparison against NNRs is still informative: both use the same reconstruction machinery, and the only difference is the mirror model. If the learned model merely encoded the reconstruction's biases, it's not obvious those biases would consistently improve the DCT norm and high-frequency content. The improvement is credible, just not fully nailed down.\n\nWho it's for: anyone working on phase diversity, adaptive optics, or computational microscopy. It's a solid methods paper that deserves serious peer review. I'd send it to a referee with a request to focus on the mirror model validation, not desk reject.\n\nRecommendation: accept for peer review, conditional on the authors addressing the circularity concern with independent wavefront data or at least a clear statement of the limitation.","headline":"DeepPD's learned deformable-mirror model is a genuine step forward for phase-diversity microscopy, but the self-referential training of that model is the load-bearing soft spot; the paper deserves peer review with a focus on independent wavefront validation.","tokens_in":14174,"tokens_out":1754,"would_cite":true,"duration_ms":15472,"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":"DeepPD recovers both the unknown optical aberration and the object from only five phase-diversity images, by replacing the usual linear deformable-mirror calibration with a learned neural mirror model.","keywords":["phase diversity","deformable mirror","adaptive optics","fluorescence microscopy","neural representations","wavefront sensing","joint estimation"],"falsifier":"Measure the mirror's actual wavefront with an external sensor, such as a pupil-matched interferometer or Shack-Hartmann sensor, across the same random voltage range used in training, and compare those measurements with the voltage-to-phase model's predictions. If the learned model's error is not smaller than the linear influence-function model's error by roughly the margin claimed for the reconstruction improvements, the central claim that the learned calibration drives the gains is not supported.","tokens_in":13058,"feed_emoji":"🔬","tokens_out":5578,"duration_ms":49368,"temperature":0.7,"pith_summary":"The paper presents DeepPD, a method that jointly estimates the unknown optical aberration (phase) and the fluorescent object from only five images: one aberrated image plus four deliberately aberrated diversity images. Its central move is to replace the standard linear calibration of a deformable mirror, which assumes the mirror's wavefront response is a sum of fixed influence functions, with a trained neural network that predicts the actual wavefront from the applied actuator voltages. Because the diversity phases enter the reconstruction model directly, a more faithful mirror model keeps the optimization from being misled, and the paper reports improved object and phase estimates compared with earlier approaches. A sympathetic reader would care because this moves phase-diversity imaging closer to practical biological use: fewer exposures, less phototoxicity, and joint aberration correction and deconvolution without a guidestar.","feed_headline":"Five images recover object and wavefront in fluorescence microscopy","feed_subtitle":"A learned deformable-mirror model replaces linear calibration, letting DeepPD beat earlier methods under severe aberrations.","key_machinery":"The load-bearing mechanism is the learned deformable-mirror model, specifically a voltage-to-phase neural network that maps the 52 actuator voltages to a predicted phase map over the pupil. It is trained together with a reversed phase-to-voltage network using a cycle-consistency loss, so that voltage-to-phase-to-voltage and phase-to-voltage-to-phase round trips return the original inputs. This learned model replaces the linear influence-function calibration that prior phase-diversity methods rely on, and it is what lets the forward model capture nonlinear mirror effects and higher-order aberrations beyond the first 21 Zernike modes. The mirror model is bootstrapped: initial phase estimates come from the same neural-representation reconstruction, then the model is retrained for three cycles using its own predicted diversity phases.","core_discovery":"DeepPD's central claim is that a single framework can jointly estimate the unknown phase aberration $\\phi$ and the fluorescent object $O$ from exactly five images—one aberrated image plus four astigmatism phase-diversity images—by simulating the imaging equation $I_k = O * |\\mathcal{F}^{-1}(P e^{i(\\phi+\\psi_k)})|^2$ and minimizing the mismatch between simulated and acquired images. The object and phase are represented by neural networks rather than pixel arrays, so the phase estimate is not confined to a low-order Zernike basis. The decisive addition is a learned mirror model: a neural network that predicts the actual wavefront produced by the deformable mirror from the 52 applied actuator voltages, replacing the assumption that the mirror response is a linear superposition of influence functions. Trained jointly with an inverse phase-to-voltage network and cycle-consistency losses, this model yields diversity phases that match the real mirror, and the paper reports that the resulting object and phase estimates are more accurate than those of Gauss-Newton, Poisson, and linear-mirror neural-representation baselines, with reliable wavefront recovery up to roughly 350 nm RMS wavefront error.","pith_inferences":["Because the mirror model is trained in an iterative self-bootstrap from its own reconstructions, the same procedure could in principle be applied to any repeatable optical element whose wavefront response drifts, such as a spatial light modulator, without requiring an external wavefront sensor.","A closed-loop use of the trained phase-to-voltage network, applying the estimated aberration correction at acquisition time on moving samples, is a natural next step that the paper mentions but does not demonstrate.","The 121-dataset comparison suggests a threshold near 100 nm RMS below which the learned mirror model adds little, and above which its advantage grows; a targeted study varying aberration composition rather than just magnitude would clarify whether high-order modes drive the gain.","If the phase estimates used as training labels carry systematic biases from the reconstruction algorithm itself, the learned mirror model may be self-consistent with DeepPD rather than physically accurate; an independent wavefront comparison over the full voltage range would settle this directly."],"forward_implications":["Only five acquired images (one aberrated plus four astigmatism-diversity images) are needed for joint object and phase recovery, a substantial reduction from roughly 100 images used by earlier neural-representation phase-diversity schemes.","Wavefront recovery is reported to be reliable up to about 350 nm RMS error, extending the previously reported range of about 200 to 250 nm for phase-diversity approaches.","Because the object is estimated jointly with the phase, the output is an aberration-corrected, deconvolved image even when no guidestar or unaberrated reference is available.","The trained phase-to-voltage network can later be used to apply desired phase diversities, and ultimately phase corrections, at acquisition time.","At high aberration magnitudes (roughly above 100 nm RMS wavefront error), the recovered-object quality metrics exceed those of the Gauss-Newton and Poisson baselines, while remaining comparable at low aberration magnitudes."],"supporting_citations":[{"why":"Supplies the five-image phase-diversity acquisition scheme, the four astigmatism diversities, the linear mirror calibration that DeepPD replaces, the Gauss-Newton baseline, and the biological myosin dataset used for comparison.","marker":"[7]"},{"why":"Introduces the neural representations of object and phase that DeepPD adapts to the phase-diversity setting.","marker":"[6]"},{"why":"Provides the Poisson-noise iterative phase-diversity algorithm used as a baseline.","marker":"[11]"},{"why":"Documents nonlinear deformable-mirror response, the premise motivating the learned mirror model.","marker":"[9]"},{"why":"Underlies the analytic-derivative Gauss-Newton phase-diversity algorithm used as a baseline.","marker":"[14]"},{"why":"Defines the DCT-norm metric used to score high-frequency content of the object estimates.","marker":"[12]"},{"why":"Provides the K-planes representation used for the object estimate.","marker":"[15]"}],"fun_headline_variants":["Joint phase and object from five microscopy images","Neural mirror calibration enables five-image phase recovery","Learned mirror model boosts phase diversity reconstruction","Five images beat Zernike limits with neural mirror model","DeepPD: wavefront and object from five images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned mirror model is trained using phase estimates produced by the same phase-diversity reconstruction approach it is meant to improve, so if those estimates carry systematic errors, the calibrated mirror model is only self-consistent with the algorithm rather than physically accurate.","fun_headline_variants_meta":{"raw":{"variants":["Joint phase and object from five microscopy images","Neural mirror calibration enables five-image phase recovery","Learned mirror model boosts phase diversity reconstruction","Five images beat Zernike limits with neural mirror model","DeepPD: wavefront and object from five images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2363,"prompt_tokens":948,"completion_tokens":1415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1343}},"tokens_in":564,"tokens_out":1415,"duration_ms":9484,"temperature":1.0,"reasoning_tokens":1343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:54:59.654843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the mirror's actual wavefront with an external sensor, such as a pupil-matched interferometer or Shack-Hartmann sensor, across the same random voltage range used in training, and compare those measurements with the voltage-to-phase model's predictions. If the learned model's error is not smaller than the linear influence-function model's error by roughly the margin claimed for the reconstruction improvements, the central claim that the learned calibration drives the gains is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the five-image phase-diversity acquisition scheme, the four astigmatism diversities, the linear mirror calibration that DeepPD replaces, the Gauss-Newton baseline, and the biological myosin dataset used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the neural representations of object and phase that DeepPD adapts to the phase-diversity setting."},{"cited_title":"& La Riviere, P","cited_arxiv_id":null,"evidence_quote":"Provides the Poisson-noise iterative phase-diversity algorithm used as a baseline."},{"cited_title":"& Bifano, T","cited_arxiv_id":null,"evidence_quote":"Documents nonlinear deformable-mirror response, the premise motivating the learned mirror model."},{"cited_title":"R., Chan, T","cited_arxiv_id":null,"evidence_quote":"Underlies the analytic-derivative Gauss-Newton phase-diversity algorithm used as a baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the DCT-norm metric used to score high-frequency content of the object estimates."},{"cited_title":"& Kanazawa, A","cited_arxiv_id":null,"evidence_quote":"Provides the K-planes representation used for the object estimate."}],"review_version":1}