Pith. sign in

REVIEW 4 major objections 4 minor 72 references

Model-free estimation of the Cram\'er-Rao bound for deep-learning microscopy in complex media

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces a model-free way to compute the Cramér–Rao bound from experimental data alone, and shows that a convolutional neural network approaches this fundamental precision limit when localizing a target hidden behind a…

desk verdict Model-free Fisher information estimation is a real contribution; the central claim that a CNN approaches the CRB is plausible but rests on a translational-invariance assumption the paper should check more carefully. read the letter →

arxiv 2505.22330 v1 pith:LZXGCLPS submitted 2025-05-28 physics.optics

Ilya Starshynov , Maximilian Weimar , Lukas M. Rachbauer , Günther Hackl , Daniele Faccio , Stefan Rotter , Dorian Bouchet This is my paper
classification physics.optics
keywords Cramér-RaoboundFisherinformationmodel-freeestimationdeeplearningmicroscopyscatteringmediaspeckletargetlocalizationindependentcomponentanalysis
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 the Cramér–Rao bound, the fundamental precision limit for any unbiased estimator, can be computed directly from measured data even when no physical model of the imaging process exists. To show this, the authors estimate the Fisher information matrix of a high-dimensional speckle image by decorrelating pixels with independent component analysis, approximating each marginal distribution with equal-frequency histograms, and taking finite differences between numerically shifted images. They apply this model-free bound to benchmark neural networks trained to localize a reflective target hidden behind a dynamic scattering medium, and find that a convolutional network with coordinate layers approaches the bound across a range of scattering strengths. If correct, the method turns the Cramér–Rao bound from a theoretical tool into a practical benchmark for any deep-learning microscope, and identifies when a network is or is not extracting all available information.

What carries the argument

The load-bearing object is the Fisher information matrix estimator built from three ingredients: (i) a linear transformation — independent component analysis (ICA) for non-Gaussian data, or PCA when the data are Gaussian — that maps correlated pixel intensities into approximately independent components \(Y = AX\); (ii) equal-frequency histogram estimates of the marginal densities \(p_k(Y_k;\$\theta$)\), which make finite-difference derivatives numerically stable; and (iii) a centered finite-difference scheme \(\partial p_k/\partial \theta_i \simeq [p_k(\$\theta$ + \hat{e}_i \$\Delta$\$\theta$) - p_k(\$\theta$ - \hat{e}_i \$\Delta$\$\theta$)]/(2\$\Delta$\$\theta$)\), where the shifted distributions are obtained by numerical translation of measured images, exploiting transverse translational invariance instead of physically moving the target. The step size \(\$\Delta$\$\theta$\) is not tuned freely but chosen from a plateau criterion based on the second derivative of the estimated Fisher information, and the final per-component estimator is \(\hat{J}^k_{ii} = \frac{1}{H(\$\Delta$\$\theta$)^2} \sum_j (\sqrt{h_{j}^{+,k}} - \sqrt{h_{j}^{-,k}})^2\).

What would settle it

Measure the same localization task with the target physically displaced to several off-center positions, estimate the Fisher information from those real measurements (using each position's own data), and compare with the value obtained by numerically shifting the central-target images; disagreement beyond the quoted error bars would show the translation-invariance assumption fails and the published bound is position-dependent.

Watch

Extended reading notes

Core claim

The central discovery is that the Fisher information of an unknown, high-dimensional data distribution can be recovered from experimental samples alone, and that this suffices to place a quantitative bound on the precision of neural-network estimators. Concretely, the paper shows that after a linear decorrelation step (ICA for non-Gaussian speckle statistics), the likelihood approximately factorizes, so the total Fisher information is a sum over single-component contributions; each marginal density is estimated by equal-frequency histograms, and its derivative with respect to the target position is obtained by shifting images of a centrally placed target by a small step \(\$\Delta$\$\theta$\) and comparing histograms. On synthetic Gaussian and non-Gaussian data with known Fisher information, the estimator returns the true value, and in the experiment the Cramér–Rao bound computed this way is approached by a CoordConv network, with standard deviations rising from about 2.5 \(\mu\)m in free space to about 12.8 \(\mu\)m at an optical thickness of \(b=5\).

Load-bearing premise

The whole comparison rests on the assumption that the imaging statistics are translation invariant, so that the Fisher information at any target position equals that computed by numerically shifting images of a target at the field-of-view center — yet the setup itself includes a tilted DMD that blurs differently at different y positions, and the field of view is finite.

Editorial extensions

If this is right

  • Any deep-learning imaging system can be benchmarked against the true physical limit of the measurement, without knowing the noise model, by comparing its achieved variance to the model-free Cramér–Rao bound.
  • Network architectures can be ranked not just by validation loss but by whether they saturate the information-theoretic bound; the paper shows CoordConv reaches the bound while other architectures trade off bias and variance.
  • The method quantifies how much harder localization becomes as scattering thickens, giving a target curve (approximately linear CRB versus optical thickness) that algorithmic improvements can be measured against.
  • Because the bound is estimated from data, it naturally includes all real noise sources (shot noise, vibrations, speckle fluctuations) and any prior information encoded in the measurement, so super-resolution effects are automatically captured.
  • The approach extends to more than two parameters, paving the way to benchmarking networks that estimate many object properties from a single image, although the paper only demonstrates the two-coordinate case.

Reading between the lines

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

  • The same recipe — decorrelate, histogram, finite-difference — should transfer to any estimation task where shifted versions of the measurement can be synthesized, such as phase or depth estimation, and could be validated by comparing the data-derived bound with an analytic bound when one exists.
  • The translation-invariance shortcut is the most fragile part: the DMD's y-dependent defocus and the finite field of view suggest the Fisher information may vary across the parameter space, so the paper's single-value bound is best read as an average local precision; a direct test with physically displaced targets would bound the error.
  • One could turn the method into an experimental design tool: since the Fisher information is computed from data, it can be measured for different illuminations, scattering strengths, or object shapes and used to choose the configuration that maximizes information.
  • The roughly 30% uncertainty in the CRB estimate (10% at b=0, 40% at b=5) means that claims of a network 'approaching' the bound are only meaningful at the tens-of-percent level; comparisons between architectures are more reliable than absolute statements.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces a model-free estimator of the Cramér-Rao bound for parameters estimated from measured images, using ICA to decorrelate pixels, equal-frequency histograms to estimate marginal densities, and finite differences to approximate derivatives of the log-likelihood. The estimator is validated on simulated Gaussian and non-Gaussian data with known Fisher information, then applied to experimental data in which a DMD target is localized behind a dynamically scattering sample at optical thicknesses b=0 to 5. The performance of four ANN architectures is compared, and the CoordConv network is reported to approach the CRB after an empirical bias correction. The paper includes a detailed supplementary analysis of step-size selection, mutual information after ICA, ANN hyperparameters, and model generalization, and the data and processing scripts are deposited in a public repository.

Significance. If correct, the method would provide a general, model-free benchmark for deep-learning microscopy and metrology in complex media. The central validation on synthetic data is convincing: the estimator recovers the known Fisher information within a few percent for both Gaussian and non-Gaussian correlated data. The experimental pipeline is described in unusual detail, and the public data and code deposit supports reproducibility. However, the main experimental claim rests on a comparison between a centrally computed CRB and ANN precision averaged over a field of view, as well as on numerical choices whose uncertainty is not fully propagated; these issues need to be resolved before the benchmark claim can be accepted as stated.

major comments (4)
  1. [Supplementary S3.3; Methods (Experimental setup); Supplementary S4.6; Discussion] The CRB used in Fig. 4(b) is estimated from images of a target placed at the central position, with the finite differences obtained by numerically shifting those images, under the assumption of transverse translational invariance. The Methods state that the DMD is tilted by a few degrees in the y direction, producing a y-dependent defocus, and Supplementary S4.6 shows that the ANN uncertainty depends strongly on the target position for b >= 3.3, with central positions easier to predict. If the likelihood is not translation invariant, the Fisher information is theta-dependent, and a CRB computed at the center is not the relevant lower bound for off-center positions; averaging the ANN standard deviation over 25 physical positions can then make the network appear closer to the central CRB than it actually is. The Discussion's assertion that the system has translational invariance needs direct experimental support, for example by computing the CRB at several physically translated positions or by demonstrating invariance of the estimated Fisher information, otherwise the claim that the CNN approaches the CRB is not established.
  2. [Supplementary S3.2, Eqs. (S1)-(S2)] The Fisher information estimator assumes that, after ICA, the transformed distribution factorizes as a product of marginals. The manuscript's own mutual information analysis, however, reports max(sum_j g_ij) = 35.3 for b = 0, with the text admitting that for this case modeling the data with a product of marginals is a rough approximation. Since the b = 0 point is included in Fig. 4(b) and used in the central comparison, the residual dependence between ICA components could bias that CRB value. The authors should quantify this effect, for example by computing the Fisher information from joint distributions for the most dependent components or by excluding b = 0 from the quantitative claim, rather than stating that ICA errors are unlikely to affect the estimates.
  3. [Supplementary S3.5; Fig. 4(b)] The estimated CRB carries an uncertainty of typically about 30%, rising to about 40% at b = 5, due to the finite-difference step-size selection, but Fig. 4(b) plots the CRB as a single value without error bars while the ANN precision is shown as full distributions. This makes the central claim that the network 'approaches' the bound quantitatively weak, particularly in the strong-scattering regime where the comparison matters most. Please propagate the step-size uncertainty into the CRB, show confidence bands, and report the ratio of ANN standard deviation to CRB together with its uncertainty.
  4. [Methods (ANN structure and optimization); Supplementary S4.4] The main text says the bias correction spline B(x,y) is fit on 'a part of a training set,' whereas Supplementary S4.4 says patterns are 'extracted from the test dataset.' If the same test data are used both to fit the bias correction and to compute the reported standard deviations, the resulting variance estimates will be optimistically biased. Please clarify which data are used for the bias-correction fit and, if necessary, evaluate the corrected estimator on a separate held-out set.
minor comments (4)
  1. [Supplementary S3.3, Eq. (S7)] In Eq. (S7), the second square root term contains h^{+,k}_j twice instead of h^{-,k}_j; the final estimator in Eq. (S8) is correct, but the intermediate formula should be fixed.
  2. [Fig. 4(b); Supplementary Fig. S10] The label 'No scattering' is used in Supplementary Fig. S10; for consistency with the main text, it should be given as b = 0 in all figure panels.
  3. [Results, 'Achievable precision for different scattering strengths'] The sentence 'In some cases, it even seems that the Cramér-Rao bound can even be overpassed' contains a duplicated 'even' and should be reworded.
  4. [Discussion] The statement that the method estimates the Cramér-Rao bound 'solely from experimental data' should be qualified, since the estimator depends on the choice of ICA, histogram binning, and finite-difference step size, and the experimental demonstration additionally relies on the translation-invariance assumption discussed above.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Cramér-Rao bound is estimated from the data likelihood, not from the ANN, and is validated on synthetic data with known Fisher information.

full rationale

The derivation is self-contained. The Fisher-information estimator is constructed from the measured frame statistics via ICA decorrelation, equal-frequency histograms and a centered finite-difference scheme (Supplementary S3.1-S3.3, Eq. S8), and is benchmarked on correlated Gaussian and non-Gaussian synthetic data with analytically known Fisher information (Fig. 2; S3.5 reports mu_FI/J = 1.01 +/- 0.05 and 0.97 +/- 0.03). The ANN precision is measured independently from the width of the softmax-position histograms on physically translated test frames (Methods; S4.1), and the CRB values plotted in Fig. 4(b) are not fitted to those widths. The assumption of transverse translational invariance, used in S3.3 to obtain theta +/- Delta theta by numerically shifting central-position images, is an external validity condition rather than a circular reduction: it is the same assumption under which the augmented training distribution matches the physical test distribution, and the paper states explicitly that it studies 'a system with translational invariance for which the Fisher information does not depend on theta' (Discussion). The DMD-tilt caveat in Methods (tilted by a few degrees in y, leading to a slight defocus for different y positions) is a correctness risk for that assumption, not a circularity. The bias-correction spline in S4.4 is fitted to test-set predictions before comparison; although this is an evaluation leak, subtracting a position-dependent constant does not change the per-position standard deviation being compared with the CRB, so it cannot manufacture the central sigma-vs-CRB agreement. Self-citations (main-text refs. 49-51; supplementary ref. 6) are background or interpretational and not load-bearing. No step of the derivation reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies mainly on standard estimation theory plus two domain assumptions (ICA factorization and translational invariance) and a test-set-based bias correction. No new physical entities are introduced.

free parameters (3)
  • Finite-difference step size Δθ = Varies per dataset; selected by plateau criterion, order 0.2 to 1 μm in Fig. S3
    The estimated Fisher information depends strongly on this step size; the paper selects it with a stability rule and reports ~30 percent uncertainty from this choice.
  • Number of equal-frequency histogram bins = Not reported in the provided text
    The Fisher information estimator in S3.3 uses equal-frequency histograms, but the bin count used for the final estimates is not stated, so the dependence of the result on this choice cannot be assessed.
  • Bias-correction spline B(x,y) = Fit to test-set predictions (S4.4)
    Used to remove ANN bias before comparing to the unbiased Cramér-Rao bound; because it is fitted to test data, the comparison is not fully independent.
assumptions (4)
  • standard math Cramér-Rao inequality and Fisher information definitions (Eqs. 1 and 2)
    Foundation of the method; assumes standard regularity conditions such as differentiable likelihood and an unbiased estimator.
  • domain assumption ICA approximately factorizes the likelihood: p(Y;θ) ≈ ∏ p_k(Y_k;θ)
    Used in Supplementary Eqs. S1 and S2; the residual dependence is not zero, and at b=0 the reported mutual information sum is large (35.3), so the Fisher information estimate is approximate.
  • domain assumption Transverse translational invariance of the imaging statistics
    Used in S3.3 to generate θ±Δθ data by shifting central images instead of physically moving the target; not quantitatively validated, and the DMD tilt induces y-dependent defocus.
  • ad hoc to paper Unbiasedness of the ANN after empirical bias correction
    The comparison to the unbiased Cramér-Rao bound assumes the spline correction in S4.4 removes estimator bias; the uncertainty in the corrected estimate is ignored.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Model-free estimation of the Cram\'er-Rao bound for deep-learning microscopy in complex media." pith.science (2026). https://pith.science/paper/LZXGCLPS

@misc{pith2026250522330,
  author       = {Pith},
  title        = {Pith review of: Model-free estimation of the Cram\'er-Rao bound for deep-learning microscopy in complex media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZXGCLPS}},
  note         = {Machine review of arXiv:2505.22330}
}
read the original abstract

Artificial neural networks have become important tools to harness the complexity of disordered or random photonic systems. Recent applications include the recovery of information from light that has been scrambled during propagation through a complex scattering medium, especially in the challenging case where the deterministic input-output transmission matrix cannot be measured. This naturally raises the question of what the limit is that information theory imposes on this recovery process, and whether neural networks can actually reach this limit. To answer these questions, we introduce a model-free approach to calculate the Cram\'er-Rao bound, which sets the ultimate precision limit at which artificial neural networks can operate. As an example, we apply this approach in a proof-of-principle experiment using laser light propagating through a disordered medium, evidencing that a convolutional network approaches the ultimate precision limit in the challenging task of localizing a reflective target hidden behind a dynamically-fluctuating scattering medium. The model-free method introduced here is generally applicable to benchmark the performance of any deep-learning microscope, to drive algorithmic developments and to push the precision of metrology and imaging techniques to their ultimate limit.

Figures

Figures reproduced from arXiv: 2505.22330 by the authors.

Figure 1
Figure 1. FIG. 1. Principle of the experiment. (a) A reflective target is placed behind a dynamic scattering sample. Our goal is to infer [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Demonstration of the Fisher information estimation procedure using numerically-generated data. (a) Representation of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Comparison of different ANN architectures for the sample with optical thickness [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Dependence of the ANN precision on the optical thickness of the scattering sample. (a) Examples of reconstructed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 65 canonical work pages

  1. [1]

    E. Moen, D. Bannon, T. Kudo, W. Graf, M. Covert, and D. Van Valen, Deep learning for cellular image analysis, Nature Methods16, 1233 (2019)

    work page 2019

  2. [2]

    L. Tian, B. Hunt, M. A. L. Bell, J. Yi, J. T. Smith, M. Ochoa, X. Intes, and N. J. Durr, Deep Learning in Biomedical Optics, Lasers in Surgery and Medicine 53, 748 (2021)

    work page 2021

  3. [3]

    Gebhart, R

    V. Gebhart, R. Santagati, A. A. Gentile, E. M. Gauger, D. Craig, N. Ares, L. Banchi, F. Marquardt, L. Pezzè, and C. Bonato, Learning quantum systems, Nature Reviews Physics 5, 141 (2023)

    work page 2023

  4. [4]

    Y. Tang, J. Kurths, W. Lin, E. Ott, and L. Kocarev, In- troduction to Focus Issue: When machine learning meets complex systems: Networks, chaos, and nonlinear dynam- ics, Chaos: An Interdisciplinary Journal of Nonlinear Sci- ence 30, 063151 (2020)

    work page 2020

  5. [5]

    Barbastathis, A

    G. Barbastathis, A. Ozcan, and G. Situ, On the use of deep learning for computational imaging, Optica6, 921 (2019)

    2019

  6. [6]

    C. Zuo, J. Qian, S. Feng, W. Yin, Y. Li, P. Fan, J. Han, K. Qian, and Q. Chen, Deep learning in optical metrology: a review, Light: Science & Applications11, 39 (2022)

    work page 2022

  7. [7]

    Gigan, O

    S. Gigan, O. Katz, H. B. d. Aguiar, E. R. Andresen, A. Aubry, J. Bertolotti, E. Bossy, D. Bouchet, J. Brake, S. Brasselet, Y. Bromberg, H. Cao, T. Chaigne, Z. Cheng, W. Choi, T. Čižmár, M. Cui, V. R. Curtis, H. Defienne, M. Hofer, R. Horisaki, R. Horstmeyer, N. Ji, A. K. LaVi- olette, J. Mertz, C. Moser, A. P. Mosk, N. C. Pégard, R. Piestun, S. Popoff, D....

    work page 2022

  8. [8]

    Choi, Deep optical imaging within complex scattering media, Nature Reviews Physics2, 141 (2020)

    S.Yoon, M.Kim, M.Jang, Y.Choi, W.Choi, S.Kang,and W. Choi, Deep optical imaging within complex scattering media, Nature Reviews Physics2, 141 (2020)

    work page 2020

Show all 72 references
  1. [9]

    A. P. Mosk, A. Lagendijk, G. Lerosey, and M. Fink, Con- trolling waves in space and time for imaging and focusing in complex media, Nature Photonics6, 283 (2012)

    2012

  2. [10]

    S. M. Popoff, G. Lerosey, R. Carminati, M. Fink, A. C. Boccara, and S. Gigan, Measuring the Transmission Ma- trix in Optics: An Approach to the Study and Control of Light Propagation in Disordered Media, Physical Review Letters 104, 100601 (2010)

    2010

  3. [11]

    D. B. Conkey, A. M. Caravaca-Aguirre, and R. Piestun, High-speed scattering medium characterization with ap- plication to focusing light through turbid media, Optics Express 20, 1733 (2012)

    2012

  4. [12]

    H. Yu, T. R. Hillman, W. Choi, J. O. Lee, M. S. Feld, R. R. Dasari, and Y. Park, Measuring Large Optical Transmission Matrices of Disordered Media, Physical Re- view Letters111, 153902 (2013)

    2013

  5. [13]

    S. M. Popoff, G. Lerosey, M. Fink, A. C. Boccara, and S. Gigan, Image transmission through an opaque material, Nature Communications1, 81 (2010)

    2010

  6. [14]

    Y. Choi, T. D. Yang, C. Fang-Yen, P. Kang, K. J. Lee, R. R. Dasari, M. S. Feld, and W. Choi, Overcoming the Diffraction Limit Using Multiple Light Scattering in a Highly Disordered Medium, Physical Review Letters107, 023902 (2011)

    2011

  7. [15]

    Horstmeyer, H

    R. Horstmeyer, H. Ruan, and C. Yang, Guidestar- assisted wavefront-shaping methods for focusing light into biological tissue, Nature Photonics9, 563 (2015)

    2015

  8. [16]

    W. Denk, J. H. Strickler, and W. W. Webb, Two-Photon Laser Scanning Fluorescence Microscopy, Science248, 73 (1990)

    1990

  9. [17]

    N. G. Horton, K. Wang, D. Kobat, C. G. Clark, F. W. Wise, C. B. Schaffer, and C. Xu, In vivo three-photon mi- croscopy of subcortical structures within an intact mouse brain, Nature Photonics7, 205 (2013)

    2013

  10. [18]

    Huang, E

    D. Huang, E. A. Swanson, C. P. Lin, J. S. Schuman, W. G. Stinson, W. Chang, M. R. Hee, T. Flotte, K. Gre- gory, C. A. Puliafito, and J. G. Fujimoto, Optical Coher- ence Tomography, Science254, 1178 (1991)

    1991

  11. [19]

    A. H. Kashani, C.-L. Chen, J. K. Gahm, F. Zheng, G. M. Richter, P. J. Rosenfeld, Y. Shi, and R. K. Wang, Optical coherence tomography angiography: A comprehensive re- 10 viewofcurrentmethodsandclinicalapplications,Progress in Retinal and Eye Research60, 66 (2017)

    2017

  12. [20]

    Bertolotti, E

    J. Bertolotti, E. G. van Putten, C. Blum, A. Lagendijk, W. L. Vos, and A. P. Mosk, Non-invasive imaging through opaque scattering layers, Nature491, 232 (2012)

    2012

  13. [21]

    O. Katz, P. Heidmann, M. Fink, and S. Gigan, Non- invasive single-shot imaging through scattering layers and around corners via speckle correlations, Nature Photonics 8, 784 (2014)

    2014

  14. [22]

    T. Ando, R. Horisaki, and J. Tanida, Speckle-learning- based object recognition through scattering media, Opt. Express 23, 33902 (2015)

    2015

  15. [23]

    Horisaki, R

    R. Horisaki, R. Takagi, and J. Tanida, Learning-based imaging through scattering media, Optics Express 24, 13738 (2016)

    2016

  16. [24]

    S. Li, M. Deng, J. Lee, A. Sinha, and G. Barbastathis, Imaging through glass diffusers using densely connected convolutional networks, Optica5, 803 (2018)

    2018

  17. [25]

    Y. Li, Y. Xue, and L. Tian, Deep speckle correlation: a deep learning approach toward scalable imaging through scattering media, Optica5, 1181 (2018)

    2018

  18. [26]

    Y. Sun, Z. Xia, and U. S. Kamilov, Efficient and accurate inversion of multiple scattering with deep learning, Optics Express 26, 14678 (2018)

    2018

  19. [27]

    Turpin, I

    A. Turpin, I. Vishniakou, and J. d. Seelig, Light scat- tering control in transmission and reflection with neural networks, Opt. Express26, 30911 (2018)

    2018

  20. [28]

    Rahmani, D

    B. Rahmani, D. Loterie, G. Konstantinou, D. Psaltis, and C. Moser, Multimode optical fiber transmission with a deep learning network, Light Sci Appl7, 69 (2018)

    2018

  21. [29]

    M. Lyu, H. Wang, G. Li, S. Zheng, and G. Situ, Learning- based lensless imaging through optically thick scattering media, Advanced Photonics1, 036002 (2019)

    2019

  22. [30]

    Caramazza, O

    P. Caramazza, O. Moran, R. Murray-Smith, and D. Fac- cio, Transmission of natural scene images through a mul- timode fibre, Nat. Commun.10, 2029 (2019)

    2019

  23. [31]

    C. Zhu, E. A. Chan, Y. Wang, W. Peng, R. Guo, B. Zhang, C. Soci, and Y. Chong, Image reconstruction through a multimode fiber with a simple neural network architecture, Sci. Rep.11, 896 (2021)

    2021

  24. [32]

    Resisi, S

    S. Resisi, S. M. Popoff, and Y. Bromberg, Image trans- mission through a dynamically perturbed multimode fiber by deep learning, Laser & Photonics Reviews15, 2000553 (2021)

    2021

  25. [33]

    Y. Li, S. Cheng, Y. Xue, and L. Tian, Displacement- agnostic coherent imaging through scatter with an inter- pretable deep neural network, Optics Express 29, 2244 (2021)

    2021

  26. [34]

    Starshynov, A

    I. Starshynov, A. Turpin, P. Binner, and D. Faccio, Sta- tistical dependencies beyond linear correlations in light scattered by disordered media, Phys. Rev. Res.4, L022033 (2022)

    2022

  27. [35]

    Rahmani, I

    B. Rahmani, I. Oguz, U. Tegin, J. liang Hsieh, D. Psaltis, and C. Moser, Learning to image and compute with mul- timode optical fibers, Nanophotonics11, 1071 (2022)

    2022

  28. [36]

    B. Bai, Y. Li, Y. Luo, X. Li, E. Çetintaş, M. Jarrahi, and A. Ozcan, All-optical image classification through un- knownrandomdiffusersusingasingle-pixeldiffractivenet- work, Light: Science & Applications12, 69 (2023)

    2023

  29. [37]

    Abdulaziz, S

    A. Abdulaziz, S. P. Mekhail, Y. Altmann, M. J. Padgett, and S. McLaughlin, Robust real-time imaging through flexible multimode fibers, Scientific Reports 13, 11371 (2023)

    2023

  30. [38]

    Huang, Z

    Z. Huang, Z. Gu, M. Shi, Y. Gao, and X. Liu, Op-fcnn: an optronic fully convolutional neural network for imaging through scattering media, Opt. Express32, 444 (2024)

    2024

  31. [39]

    Giovannetti, S

    V. Giovannetti, S. Lloyd, and L. Maccone, Advances in quantum metrology, Nature Photonics5, 222 (2011)

    2011

  32. [40]

    Polino, M

    E. Polino, M. Valeri, N. Spagnolo, and F. Sciarrino, Photonic quantum metrology, AVS Quantum Science2, 024703 (2020)

    2020

  33. [41]

    S. D. Cohen, H. L. D. de S. Cavalcante, and D. J. Gau- thier, Subwavelength Position Sensing Using Nonlinear Feedback and Wave Chaos, Physical Review Letters107, 254103 (2011)

    2011

  34. [42]

    del Hougne, M

    P. del Hougne, M. F. Imani, M. Fink, D. R. Smith, and G. Lerosey, Precise Localization of Multiple Noncoopera- tive Objects in a Disordered Cavity by Wave Front Shap- ing, Physical Review Letters121, 063901 (2018)

    2018

  35. [43]

    del Hougne, Robust position sensing with wave fin- gerprints in dynamic complex propagation environments, Physical Review Research2, 043224 (2020)

    P. del Hougne, Robust position sensing with wave fin- gerprints in dynamic complex propagation environments, Physical Review Research2, 043224 (2020)

    2020

  36. [44]

    Jauregui-Sánchez, H

    Y. Jauregui-Sánchez, H. Penketh, and J. Bertolotti, Tracking moving objects through scattering media via speckle correlations, Nature Communications 13, 5779 (2022)

    2022

  37. [45]

    H. L. Van Trees, K. L. Bell, and Z. Tian,Detection Es- timation and Modulation Theory, Part I (John Wiley & Sons, 2013)

    2013

  38. [46]

    R. J. Ober, S. Ram, and E. S. Ward, Localization Accu- racy in Single-Molecule Microscopy, Biophysical Journal 86, 1185 (2004)

    2004

  39. [47]

    Deschout, F

    H. Deschout, F. C. Zanacchi, M. Mlodzianoski, A. Di- aspro, J. Bewersdorf, S. T. Hess, and K. Braeckmans, Pre- cisely and accurately localizing single emitters in fluores- cence microscopy, Nature Methods11, 253 (2014)

    2014

  40. [48]

    X. Song, J. Xu, F. Liu, T. X. Han, and Y. C. Eldar, Intelligent Reflecting Surface Enabled Sensing: Cramér- Rao Lower Bound optimization, in2022 IEEE Globecom Workshops (GC Wkshps) (2022) pp. 413–418

    2022

  41. [49]

    Bouchet, R

    D. Bouchet, R. Carminati, and A. P. Mosk, Influence of the Local Scattering Environment on the Localization Precision of Single Particles, Physical Review Letters124, 133903 (2020)

    2020

  42. [50]

    Bouchet, S

    D. Bouchet, S. Rotter, and A. P. Mosk, Maximum infor- mation states for coherent scattering measurements, Na- ture Physics17, 564 (2021)

    2021

  43. [51]

    Horodynski, D

    M. Horodynski, D. Bouchet, M. Kühmayer, and S. Rot- ter, Invariance property of the fisher information in scat- tering media, Phys. Rev. Lett.127, 233201 (2021)

    2021

  44. [52]

    I. T. Jolliffe and J. Cadima, Principal component anal- ysis: a review and recent developments, Philosophical transactions of the royal society A: Mathematical, Physi- cal and Engineering Sciences374, 20150202 (2016)

    2016

  45. [53]

    Hyvärinen and E

    A. Hyvärinen and E. Oja, Independent component analy- sis: algorithms and applications, Neural Networks13, 411 (2000)

    2000

  46. [54]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville,Deep Learn- ing (MIT Press, 2016)

    2016

  47. [55]

    R. Liu, J. Lehman, P. Molino, F. P. Such, E. Frank, A. Sergeev, and J. Yosinski, An intriguing failing of con- volutional neural networks and the coordconv solution, in Proceedings of the 32nd International Conference on Neu- ral Information Processing Systems, NIPS’18 (Curran...

    2018

  48. [56]

    Huang, Z

    G. Huang, Z. Liu, L. V. D. Maaten, and K. Q. Wein- berger, Densely connected convolutional networks, in2017 IEEE Conference on Computer Vision and Pattern Recog- 11 nition (CVPR) (IEEE Computer Society, Los Alamitos, CA, USA, 2017) pp. 2261–2269

    2017

  49. [57]

    H. H. Barrett, J. L. Denny, R. F. Wagner, and K. J. Myers, Objective assessment of image quality. II. Fisher information, Fouriercrosstalk, andfiguresofmeritfortask performance, Journal of the Optical Society of America A 12, 834 (1995)

    1995

  50. [58]

    Bouchet, J

    D. Bouchet, J. Dong, D. Maestre, and T. Juffmann, Fun- damental Bounds on the Precision of Classical Phase Mi- croscopes, Physical Review Applied15, 024047 (2021)

    2021

  51. [59]

    Szameit, Y

    A. Szameit, Y. Shechtman, E. Osherovich, E. Bullkich, P. Sidorenko, H. Dana, S. Steiner, E. B. Kley, S. Gazit, T. Cohen-Hyams, S. Shoham, M. Zibulevsky, I. Yavneh, Y. C. Eldar, O. Cohen, and M. Segev, Sparsity-based single-shot subwavelength coherent diffractive imaging, Natur...

    2012

  52. [60]

    Bouchet, J

    D. Bouchet, J. Seifert, and A. P. Mosk, Optimizing il- lumination for precise multi-parameter estimations in co- herent diffractive imaging, Optics Letters46, 254 (2021)

    2021

  53. [61]

    S. Feng, C. Kane, P. A. Lee, and A. D. Stone, Cor- relations and fluctuations of coherent wave transmis- sion through disordered media, Phys. Rev. Lett.61, 834 (1988)

    1988

  54. [62]

    Akkermans and G

    E. Akkermans and G. Montambaux,Mesoscopic Physics of Electrons and Photons (Cambridge University Press, 2007)

    2007

  55. [63]

    Pierrat, R

    I.Starshynov, A.Paniagua-Diaz, N.Fayard, A.Goetschy, R. Pierrat, R. Carminati, and J. Bertolotti, Non-Gaussian Correlations between Reflected and Transmitted Intensity Patterns Emerging from Opaque Disordered Media, Phys- ical Review X8, 021041 (2018)

    2018

  56. [64]

    H. Cao, T. Čižmár, S. Turtaev, T. Tyc, and S. Rot- ter, Controlling light propagation in multimode fibers for imaging, spectroscopy, and beyond, Advances in Optics and Photonics15, 524 (2023)

    2023

  57. [65]

    Stibůrek, P

    M. Stibůrek, P. Ondráčková, T. Tučková, S. Turtaev, M. Šiler, T. Pikálek, P. Jákl, A. Gomes, J. Krejčí, P. Kol- bábková, H. Uhlířová, and T. Čižmár, 110µm thin endo- microscope for deep-brain in vivo observations of neuronal connectivity, activity and blood flow dynamics, Natu...

    2023

  58. [66]

    Depth” parameter denotes the number of layers, and the “Out size

    H. Sarafraz, T. Nöbauer, H. Kim, F. Soldevila, S. Gigan, and A. Vaziri, Speckle-enabled in vivo demixing of neural activity in the mouse brain, Biomedical Optics Express 15, 3586 (2024). Acknowledgements I.S. and D.F. acknowledge financial support from the UK Engineering and P...

    2024

  59. [67]

    I. T. Jolliffe and J. Cadima, Principal component analysis: a review and recent developments, Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences374, 20150202 (2016)

    2016

  60. [68]

    Hyvärinen and E

    A. Hyvärinen and E. Oja, Independent component analysis: algorithms and applications, Neural Networks13, 411 (2000)

    2000

  61. [69]

    T. M. Cover and J. A. Thomas,Elements of information theory (John Wiley & Sons, 1999)

    1999

  62. [70]

    Huang, Z

    G. Huang, Z. Liu, L. V. D. Maaten, and K. Q. Weinberger, Densely connected convolutional networks, in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE Computer Society, Los Alamitos, CA, USA,

  63. [71]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in International Conference on Learning R...

    2021

  64. [72]

    Hüpfl, F

    J. Hüpfl, F. Russo, L. M. Rachbauer, D. Bouchet, J. Lu, U. Kuhl, and S. Rotter, Continuity equation for the flow of Fisher information in wave scattering, Nature Physics20, 1294 (2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.