Phase retrieval via Zernike phase contrast microscopy with an untrained neural network

Keiichiro Toda; Kohki Horie; Rikimaru Kurata; Ryoichi Horisaki; Takuro Ideguchi; Zinan Zhou

REVIEW 3 major objections 5 minor 20 references

Phase retrieval via Zernike phase contrast microscopy with an untrained neural network

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An untrained neural network can turn ordinary Zernike phase-contrast microscope images into quantitative phase maps without manual tuning of regularization parameters.

desk verdict A useful, honest method paper that convincingly shows an untrained decoder prior can replace manual regularization in Zernike PCM phase retrieval—but the experimental validation is thinner than the breadth claims, and forward-model fidelity is the real risk. read the letter →

arxiv 2501.14290 v1 pith:WX6HLOYR submitted 2025-01-24 physics.optics

Zinan Zhou , Keiichiro Toda , Rikimaru Kurata , Kohki Horie , Ryoichi Horisaki , Takuro Ideguchi This is my paper · ORCID

classification physics.optics PACS 42.30.Rx

keywords phaseretrievalZernikecontrastmicroscopyuntrainedneuralnetworkdeepimagepriorcompressivepropagationquantitativeimagingcomputational

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that phase retrieval from a single Zernike phase-contrast microscopy (PCM) image can be made quantitative by replacing hand-tuned regularizers with an untrained deep decoder used as a structural prior. The method minimizes the error between the observed PCM intensity and the intensity predicted by a compressive-propagation forward model applied to the network's phase output. With fixed network architecture and optimizer settings, the authors report lower RMS phase error than a regularization-based baseline on simulated beads, cells, and resolution targets, and on experimental beads and separated cells. If correct, this means off-the-shelf PCM hardware can deliver quantitative phase imaging without per-sample parameter tuning, making QPI practical for routine biological use.

What carries the argument

The load-bearing object is the pair (D, H): D is a deep decoder—a non-convolutional network with channel-wise linear layers, bilinear upsampling, ReLU and ChannelNorm, and a final LeakyReLU that encodes the prior that the phase image has near-zero background—and H is the compressive-propagation model of PCM, which approximates partially coherent illumination by averaging M random wavefronts propagated through a condenser annulus pupil c and phase-ring pupil p. The optimization runs Adam on the network weights W with 200 random wavefronts per step, 4000 epochs, and a learning rate of 3×$10^{-4}$. The deep decoder acts as a self-learned regularizer that restricts the solution space, while H links the phase estimate to the observed intensity.

What would settle it

A direct calibration test: acquire PCM images of a well-characterized phase target (e.g., microfabricated pillars with known heights from profilometry) on the same microscope, run the proposed fixed-network pipeline, and compare recovered phase to the profilometry ground truth. If the recovered phase shows a systematic, sample-size-dependent bias that persists across random restarts, or if deliberately misestimating the pupil parameters c and p in H changes the reconstruction, then the forward model rather than the network prior is the limiting factor.

Watch

Extended reading notes

Core claim

The central claim is that the estimated phase image θ = D(W; B0), produced by a fixed deep decoder with randomly initialized weights W and a fixed input tensor B0, can be optimized by minimizing ||g − H(D(W;B0))||$2^{2}$, where H is the compressive-propagation PCM forward model, to recover quantitative phase from a single PCM intensity image. The authors establish this by comparing RMS errors against the regularization-based method of Kurata et al. under three manually tuned regularization strengths, across four simulated sample classes and two experimental classes. They report the untrained-network method matches or beats the best-tuned regularization result in nearly every class, and is the only method that does not require sample-dependent hyperparameter selection. They further map the recoverable sample size-phase range, finding diameters up to about 40 µm at phases ≲1.5 rad and phases up to 2.88 rad for small samples.

Load-bearing premise

The forward model H in Eq. (5)—with known condenser and phase-ring pupil filters and randomly drawn wavefronts—accurately represents the actual microscope's optical behavior; the numerical experiments use the same H for both simulation and inversion, and only two experimental sample types test the model, so any mismatch would bias the recovered phase regardless of the network prior.

Editorial extensions

If this is right

If correct, any existing Zernike phase-contrast microscope can provide quantitative phase maps with no hardware modification and no per-image tuning.
The fixed-network recipe removes the user-dependent choice of regularization strength, making phase retrieval practical for non-specialists.
The recovered phase range and object-size limits are now characterized (about 40 µm diameter at phase ≲1.5 rad; up to 2.88 rad for small objects), giving users a concrete applicability envelope.
The method demonstrates that deep priors are compatible with incoherent illumination models, extending untrained-network phase retrieval beyond coherent setups.
Because the network is untrained, the approach requires no dataset and can be applied immediately to new sample types without retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the forward model H is accurate only for thin, weakly scattering samples, the same fixed network will inherit that limitation; a testable extension is to calibrate H against a known standard and verify retrieval accuracy, quantifying model mismatch separately from network behavior.
The observed 'hole artifacts' suggest the deep decoder's initialization can trap optimization in a local basin; pretraining on a segmentation image helps partially, implying that a better initialization scheme or a convergence-aware optimizer could remove the residual artifacts entirely.
The authors' comparison with three fixed regularization strengths shows no single setting generalizes; this implies the advantage of the untrained network is not raw accuracy but adaptability—an inference testable by running the same benchmark with oracle-tuned per-sample regularization.
Since the network uses no convolutions, the role of the prior is purely statistical (smoothness and background sparsity); one could test whether an explicit hand-crafted prior with the same properties, such as a learned dictionary, achieves comparable results without a network.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuine new combination, not a big idea. Untrained deep priors and compressive propagation both exist; putting them into Zernike PCM phase retrieval is new in the literature, and it works at least as well as the best of three fixed-regularization baselines across four sample classes in simulation and two in experiment. The authors are also unusually honest: they state plainly that the simulation uses the exact forward model, they show the hole artifacts, they discuss the low-frequency loss, and they give the computational cost.

Where I part company with the stress-test note: I don't think the absence of grouped-cell and resolution-target experiments is fatal. The simulation evidence is still informative, and the two experimental classes, while small, are real tests with DH ground truth. The small sample size does bother me: 6 separated-cell images, 19 beads, no standard deviations. Given that the central selling point is 'improved accuracy and robustness across various samples,' the robustness half of that claim rests mostly on exact-model simulations.

The bigger issue is the forward model H in Eq. (5). The paper cites [7] for the parameter settings but does not say how the pupil filters c and p were calibrated for the particular microscope used. If those filters are off, the recovered phase will be biased no matter what prior you use. For two simple sample types the experimental agreement suggests the model is roughly right, but 'roughly right' may not be enough for the quantitative claims. This is the part I would want a referee to push on.

One more thing: the no-tuning selling point is real but modest. The network architecture, learning rate, epoch count, and M are all fixed, and they presumably selected those once on some validation set. So they eliminated per-sample regularization tuning, not all tuning. That is an honest claim and worth having, but it should not be oversold.

Recommendation: let it into peer review with the expectation of heavy-ish revision. Add more experimental samples, give error bars, and describe the pupil calibration. I would not want this desk-rejected; it is a legitimate method contribution and the limitations are openly discussed.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a computational phase retrieval method for Zernike phase contrast microscopy (PCM) using an untrained deep decoder as an image prior. The estimated phase image is represented as the output of a deep decoder applied to a fixed random tensor, and the network weights are optimized to minimize the L2 distance between the observed PCM intensity image and the intensity predicted by a compressive-propagation forward model (Eq. 5). The method is tested on simulated PCM images of microbeads, separated cells, grouped cells, and resolution targets, and on experimental PCM images of microbeads and separated cells. The authors compare their results with a previously published regularization-based ADMM approach using three representative regularization strengths. They report lower root-mean-square phase errors for the untrained-network method in most cases and conclude that it eliminates manual regularization tuning while improving accuracy and robustness.

Significance. If the reported results hold, the method would provide a practical path to quantitative phase imaging from off-the-shelf PCM without per-sample regularization tuning, which would be useful for biological applications. The paper's strengths are its simple and reproducible formulation, the use of a fixed network architecture and optimizer across all experiments, and the inclusion of both simulated and experimental data. The authors also explicitly acknowledge that the numerical experiments use an exact forward model. However, the breadth of the central claim is not yet fully supported: the experimental validation covers only two simple sample types with a small number of images and no error bars, and the robustness of the method to forward-model mismatch is not assessed. The significance is therefore moderate and depends on whether the forward-model fidelity and the experimental evidence can be strengthened.

major comments (3)

[III, Tables II and III; Eq. (5)] The simulation protocol uses the same forward model H for data generation and inversion (the authors state that 'the model used to solve the inverse problem is exact'), so the numerical comparisons in Table II cannot detect errors in the PCM model. The experimental validation is limited to 19 microbead images and 6 separated-cell images; grouped cells and resolution targets, which are central to the claimed breadth, are tested only in simulation. Since the forward model in Eq. (5) depends on the pupil filters c and p, the manuscript should explain how these filters are calibrated or validated for the experimental microscope, and should either add experimental data for a more complex sample type or clearly temper the 'various samples' claim in the conclusion. A perturbation or noise-sensitivity test of H would strengthen the robustness argument.
[II and V] The claim that the approach eliminates empirical hyperparameter tuning is broader than what is demonstrated. The method fixes the deep decoder channel count (k_i=128), the input canvas size, the upsampling schedule, the optimizer learning rate (3e-4), the number of epochs (4000), and the number of random wavefronts per step (M=200). No sensitivity study is reported for these choices, so it is not established that they transfer to other experimental setups without adjustment. The authors should clarify that the elimination refers specifically to the regularization parameters (rho, epsilon_TV, epsilon_l1) of the earlier ADMM method, and should provide a small sensitivity analysis for at least the learning rate and epoch count to support the robustness claim.
[III, Table III] The experimental improvement over the best regularization baseline is modest for separated cells (UNN RMS 0.172 rad vs 0.199 rad for R-L) and is based on only six images, with no standard deviation or per-sample values reported. Tables II and III also average five restarts per sample without reporting the spread across restarts. Given the acknowledged hole artifacts (Section IV, Appendix B), the claim that the untrained network outperforms the regularization method in experiment should be supported by error bars or a statistical test; otherwise the reader cannot assess whether the difference is meaningful.

minor comments (5)

[II, Eq. (3)] The slope of the LeakyReLU activation is not specified; please state the value (e.g., 0.01) used in the experiments.
[II, Eq. (5)] The symbol M is used both for the number of wavefronts in the forward model (4000 in simulation) and for the number used per optimization step (200); using M_sim and M_opt would avoid ambiguity.
[III, Table I] For the 'High' regularization setting, the text says it was 'determined through empirically tuning to optimize performance,' but it does not say on which sample or metric. Please specify this so the comparison baseline is transparent.
[Fig. 2 caption] The caption lists 'PCM images, ground truth phase images, and phase restorations' but the figure panels are not labeled with these categories; adding labels and a common phase colorbar would improve readability.
[IV, reference [8]] The suggestion for convergence analysis cites reference [8]; it would be helpful to state what specific result from that work transfers to the present setting, or to cite a reference that directly addresses convergence of deep-decoder optimization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the optimization is a genuine inverse problem against fixed data; model-fidelity concerns are correctness risks, not circularity.

full rationale

The derivation chain is self-contained as an inverse problem: the observed PCM intensity g is fixed input data, the phase estimate is generated by an untrained deep decoder theta = D(W; B0), and the estimated intensity is computed through the forward PCM model H in Eq. (5). The optimization in Eq. (4) minimizes ||g - H(D(W;B0))||^2 over network weights W, with no ground-truth phase entering the objective. Ground-truth phase is used only after recovery to compute RMS errors, so no fitted parameter is relabeled as a prediction. The forward model and compressive-propagation approximation are adopted from the authors' prior works [7] and [15], but those are published, externally checkable models rather than results derived in this paper, and no uniqueness theorem or ansatz is smuggled in via self-citation. The numerical experiments do use the same H for both data generation and inversion, which limits their ability to detect model mismatch, but this is an 'inverse crime' concern about external validity, not circularity: the optimization still must invert a nontrivial nonlinear map, and the deep decoder prior materially constrains the solution. Experimental results on beads and separated cells provide independent, if limited, support. Therefore the central claim does not reduce to its inputs by construction, and no circular step is identified.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method inherits a forward model from the authors' prior work and relies on heuristics of deep image priors. No new physical entities or fitted material constants are introduced; the free parameters are network architecture and optimizer settings chosen by hand.

free parameters (6)

Deep decoder channel count k_i = 128 for all layers i=0..5
Fixed network width chosen by hand; not tuned per sample, but it controls the expressiveness of the image prior.
Input canvas size n0 = 8x8
Starting resolution for the deep decoder; chosen by hand.
Upsampling schedule = Double canvas size each layer to reach 256x256
Part of the deep decoder architecture; chosen by hand.
Optimizer learning rate = 3e-4 (Adam)
Fixed for all experiments; no per-sample tuning but still a hyperparameter.
Number of epochs = 4000
Fixed stopping point; chosen by hand.
Random wavefronts per optimization step M = 200 per step, 4000 for data generation
Trade-off between statistical noise and computational cost; chosen by hand.

assumptions (5)

domain assumption Compressive propagation with finite random wavefronts models partially coherent illumination
Eq. (5) and Section II; approximation from [15], with M=200 per step. The error relative to true partially coherent imaging is not quantified.
domain assumption Pupil functions c and p for the condenser annulus and phase ring are known and fixed
Used in forward model Eq. (5); values inherited from [7]; no in-paper calibration or uncertainty analysis.
domain assumption Deep decoder favors phase images with natural-image statistics
Central to the method; heuristic from deep image prior literature [13,14], not proven for phase distributions.
domain assumption Every phase image has a background region with phase near zero
Used to justify the LeakyReLU final activation in Section II; may fail for crowded fields of view.
domain assumption The nonconvex optimization converges to an acceptable solution
Discussion states convergence analysis should be done based on [8]; no guarantee is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Phase retrieval via Zernike phase contrast microscopy with an untrained neural network." pith.science (2026). https://pith.science/paper/WX6HLOYR

@misc{pith2026250114290,
  author       = {Pith},
  title        = {Pith review of: Phase retrieval via Zernike phase contrast microscopy with an untrained neural network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WX6HLOYR}},
  note         = {Machine review of arXiv:2501.14290}
}

read the original abstract

Zernike's phase contrast microscopy (PCM) is among the most widely used techniques for observing phase objects, but it lacks quantitative nature, as it cannot directly provide phase information. Current methods for computationally extracting phase distributions from PCM images, however, rely heavily on empirical regularization parameter tuning. In this paper we extend an existing approach by employing an untrained neural network as an image prior, removing the need for manual regularization. We quantitatively demonstrate improved accuracy and robustness in phase retrieval compared to existing methods, using numerical and experimental PCM images. Our results confirm the feasibility of applying deep priors for phase retrieval in incoherent illumination setups.

Figures

Figures reproduced from arXiv: 2501.14290 by the authors.

**Figure 1.** FIG. 1: Schematic of the optimization. Our method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** FIG. 2: Comparison between phase retrieval by UNN [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: FIG. 3: (a) Artificial circle samples of different [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]

**Figure 5.** Figure 5: FIG. 5: Visualization of estimated phase images during [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 4.** Figure 4: FIG. 4: A preview of phase samples. “Beads” group is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

[7]

Kurata, K

R. Kurata, K. Toda, G. Ishigane, M. Naruse, R. Horisaki, and T. Ideguchi, Single-image phase retrieval for off-the- shelf zernike phase-contrast microscopes, Optics Express 32, 2202 (2024)

work page 2024
[1]

Zernike, How i discovered phase contrast, Science 121, 345 (1955)

F. Zernike, How i discovered phase contrast, Science 121, 345 (1955)

1955
[2]

N. T. Shaked, Z. Zalevsky, and L. L. Satterwhite, Biomedical optical phase microscopy and nanoscopy (Academic Press, 2012)

work page 2012
[3]

Maurer, A

C. Maurer, A. Jesacher, S. Bernet, and M. Ritsch-Marte, Phase contrast microscopy with full numerical aperture illumination, Optics Express 16, 19821 (2008), publisher: Optica Publishing Group

work page 2008
[4]

P. Gao, B. Yao, I. Harder, N. Lindlein, and F. J. Torcal- Milla, Phase-shifting Zernike phase contrast microscopy for quantitative phase measurement, Optics Letters 36, 4305 (2011)

work page 2011
[5]

Z. Wang, L. Millet, M. Mir, H. Ding, S. Unarunotai, J. Rogers, M. U. Gillette, and G. Popescu, Spatial light interference microscopy (SLIM), Optics Express 19, 1016 (2011), publisher: Optica Publishing Group

work page 2011
[6]

Z. Yin, T. Kanade, and M. Chen, Understanding the phase contrast optics to restore artifact-free microscopy images for segmentation, Medical Image Analysis 16, 1047 (2012)

work page 2012
[8]

Jagatap and C

G. Jagatap and C. Hegde, Algorithmic guarantees for inverse imaging with untrained network priors, Advances in neural information processing systems 32 (2019)

work page 2019

Show all 20 references

[9]

F. Wang, Y. Bian, H. Wang, M. Lyu, G. Pedrini, W. Os- ten, G. Barbastathis, and G. Situ, Phase imaging with an untrained neural network, Light: Science & Applications 9, 77 (2020)

2020
[10]

K. C. Zhou and R. Horstmeyer, Diffraction tomography with a deep image prior, Optics express28, 12872 (2020)

2020
[11]

Bostan, R

E. Bostan, R. Heckel, M. Chen, M. Kellman, and L. Waller, Deep phase decoder: self-calibrating phase mi- croscopy with an untrained deep neural network, Optica 7, 559 (2020)

2020
[12]

Mashiko, J

R. Mashiko, J. Tanida, M. Naruse, and R. Horisaki, Ex- trapolated speckle-correlation imaging with an untrained deep neural network, Appl. Opt. 62, 8327 (2023)

2023
[13]

Ulyanov, A

D. Ulyanov, A. Vedaldi, and V. Lempitsky, Deep image prior, in Proceedings of the IEEE conference on computer vision and pattern recognition(2018) pp. 9446–9454

2018
[14]

Heckel and P

R. Heckel and P. Hand, Deep decoder: Concise image representations from untrained non-convolutional net- works, arXiv preprint arXiv:1810.03982 (2018)

2018 arXiv
[15]

Horisaki, T

R. Horisaki, T. Aoki, Y. Nishizaki, A. R¨ ohm, N. Chauvet, J. Tanida, and M. Naruse, Compressive propagation with coherence, Optics Letters 47, 613 (2022)

2022
[16]

D. P. Kingma, Adam: A method for stochastic optimiza- tion, arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[17]

S. H. Chan, X. Wang, and O. A. Elgendy, Plug-and-play ADMM for image restoration: Fixed-point convergence and applications, IEEE Transactions on Computational Imaging 3, 84 (2016)

2016
[18]

E. J. Candes, M. B. Wakin, and S. P. Boyd, Enhanc- ing sparsity by reweighted l1 minimization, Journal of Fourier analysis and applications 14, 877 (2008)

2008
[19]

Bhaduri, C

B. Bhaduri, C. Edwards, H. Pham, R. Zhou, T. H. Nguyen, L. L. Goddard, and G. Popescu, Diffraction phase microscopy: principles and applications in materi- als and life sciences, Advances in Optics and Photonics 6, 57 (2014). 6

2014
[20]

T. H. Nguyen, M. Kandel, H. M. Shakir, C. Best- Popescu, J. Arikkath, M. N. Do, and G. Popescu, Halo- free phase contrast microscopy, Scientific reports7, 44034 (2017)

2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [7]

Kurata, K

R. Kurata, K. Toda, G. Ishigane, M. Naruse, R. Horisaki, and T. Ideguchi, Single-image phase retrieval for off-the- shelf zernike phase-contrast microscopes, Optics Express 32, 2202 (2024)

work page 2024

[2] [1]

Zernike, How i discovered phase contrast, Science 121, 345 (1955)

F. Zernike, How i discovered phase contrast, Science 121, 345 (1955)

1955

[3] [2]

N. T. Shaked, Z. Zalevsky, and L. L. Satterwhite, Biomedical optical phase microscopy and nanoscopy (Academic Press, 2012)

work page 2012

[4] [3]

Maurer, A

C. Maurer, A. Jesacher, S. Bernet, and M. Ritsch-Marte, Phase contrast microscopy with full numerical aperture illumination, Optics Express 16, 19821 (2008), publisher: Optica Publishing Group

work page 2008

[5] [4]

P. Gao, B. Yao, I. Harder, N. Lindlein, and F. J. Torcal- Milla, Phase-shifting Zernike phase contrast microscopy for quantitative phase measurement, Optics Letters 36, 4305 (2011)

work page 2011

[6] [5]

Z. Wang, L. Millet, M. Mir, H. Ding, S. Unarunotai, J. Rogers, M. U. Gillette, and G. Popescu, Spatial light interference microscopy (SLIM), Optics Express 19, 1016 (2011), publisher: Optica Publishing Group

work page 2011

[7] [6]

Z. Yin, T. Kanade, and M. Chen, Understanding the phase contrast optics to restore artifact-free microscopy images for segmentation, Medical Image Analysis 16, 1047 (2012)

work page 2012

[8] [8]

Jagatap and C

G. Jagatap and C. Hegde, Algorithmic guarantees for inverse imaging with untrained network priors, Advances in neural information processing systems 32 (2019)

work page 2019