Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Super-Resolution Optical Coherence Tomography Using Diffusion Model-Based Plug-and-Play Priors

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that a plug-and-play diffusion model, alternating likelihood and denoising steps inside an MCMC sampler, reconstructs 4x-undersampled OCT B-scans at 256x256 with sharper anatomical structure and less speckle than a…

desk verdict Solid, honest application of PnP-DM to OCT super-resolution, but the test-set degradation model is underspecified and that undercuts the headline numbers. read the letter →

arxiv 2505.14916 v1 pith:PSGS3AV6 submitted 2025-05-20 eess.IV cs.CV

classification eess.IVcs.CV
keywords opticalcoherencetomographysuper-resolutionplug-and-playdiffusionmodelMCMCposteriorsamplingsparsemeasurementscornealimagingimageinverseproblems
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 tries to establish that a diffusion-model-based plug-and-play reconstruction, alternating data-consistency updates with denoising prior steps, can turn 4x-downsampled, noisy OCT B-scans into high-resolution images that preserve corneal microstructure better than a conventional 2D-UNet. It matters because high-speed OCT acquisition undersamples to avoid eye-motion artifacts, but undersampling degrades image quality; if this approach holds, clinical high-speed OCT could retain diagnostic detail. The evidence is quantitative: on 100 fish-eye corneal images, all PnP-DM variants beat 2D-UNet on PSNR, SSIM, and LPIPS, with the best variant reaching PSNR 32.50 versus 30.07.

What carries the argument

The central mechanism is the Split Gibbs Sampler, which alternates two steps: a likelihood step that samples from a Gaussian whose mean and precision are computed through the SVD of the block-averaging operator $P_f$ (equations 5--6), and a prior step that applies the EDM denoising stochastic differential equation (equation 7) using a pretrained diffusion score network. The coupling parameter $\rho$ anneals exponentially from 10 to 0.3, which the paper states accelerates mixing and helps avoid local minima in the highly ill-posed inverse problem.

What would settle it

Acquire real high-speed OCT B-scans with an irregular or undersampled A-scan pattern, reconstruct them with PnP-DM and with 2D-UNet, and compare both to co-registered conventional full-rate scans: if PnP-DM no longer preserves the lens and iris boundaries or its PSNR drops below the UNet's, the claim as stated would fail.

Watch

Extended reading notes

Core claim

Under a linear forward model where the observed low-resolution image is a 4x block-average of the high-resolution image plus Gaussian noise, the PnP-DM algorithm -- which alternates a Gaussian likelihood update computed via the singular value decomposition of the forward operator with an EDM diffusion denoiser acting as an implicit image prior -- reconstructs 256x256 OCT B-scans from 64x64 inputs such that anatomical boundaries such as the iris and crystalline lens appear sharper and speckle is better suppressed than with a supervised 2D-UNet. The paper further claims that the choice of diffusion prior formulation has little perceptual effect, with VP, VE, DDPM, and EDM variants all outperforming the conventional baseline.

Load-bearing premise

The entire evaluation rests on the assumption that real OCT undersampling behaves like a 4x block-averaging downsampling with Gaussian noise, because all test inputs are generated from high-resolution images by exactly that operator.

Editorial extensions

If this is right

  • Sparse 4x-undersampled OCT scans can be reconstructed at 256x256 with preserved anatomical boundaries, making high-speed acquisition more clinically usable.
  • The same PnP-DM pipeline works with several different diffusion priors (VP, VE, DDPM, and EDM) with little perceptual difference, so the prior can be swapped without retraining the reconstruction loop.
  • Using an out-of-domain prior trained on human faces still removes speckle better than conventional baselines, suggesting the method is partially robust to prior mismatch.
  • The method produces multiple posterior samples per image, and the paper evaluates metrics on the mean of 100 samples, enabling a distribution of reconstructions rather than a single point estimate.

Reading between the lines

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

  • If the forward operator were replaced with a more realistic model for non-uniform or motion-correlated undersampling, the likelihood step would need re-derivation, but the plug-and-play loop would carry over -- this is an editorial extension, not a paper claim.
  • The 100-iteration MCMC at about 80 seconds per image is a practical bottleneck; patch-based or latent-space diffusion priors could scale the method to full 1024x1024x128 volumes, which the paper itself lists as a limitation.
  • The posterior samples could be used to produce pixel-wise uncertainty maps for clinical decision support, a use not explored in the paper.
  • Because the diffusion prior is domain-agnostic, the same algorithm could plausibly transfer to other undersampled biomedical modalities with linear forward models, such as MRI, provided the SVD of the forward operator is computable.
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.

Referee Report

4 major / 5 minor

Summary. The paper applies the plug-and-play diffusion model (PnP-DM) framework of Wu et al. to super-resolution of OCT B-scan corneal images. The forward model is a linear 4x block-averaging downsampling operator with Gaussian noise, and reconstruction alternates between a likelihood step using the operator's SVD and a prior step using an EDM denoiser. Experiments on fish-eye corneal OCT data compare PnP-DM with different diffusion priors against bicubic interpolation and a 2D-UNet, reporting that PnP-DM achieves higher PSNR, SSIM, and LPIPS. An ablation study replaces the in-domain prior with an FFHQ-trained prior. The central methodological machinery is standard and borrowed correctly, but the evaluation protocol has important ambiguities and inconsistencies that currently limit confidence in the headline claim.

Significance. If the reported results hold, the paper demonstrates a useful application of diffusion-based plug-and-play priors to OCT super-resolution, with the merit of treating reconstruction as an explicit inverse problem rather than using a purely supervised mapping. The ablation with an out-of-domain FFHQ prior is informative and supports the need for domain-matched priors. The strengths are the clear problem formulation, use of a principled MCMC-based posterior sampling scheme, and the comparison across several diffusion prior variants. However, the methodological novelty is modest, and the current evaluation does not yet establish that the advantage over 2D-UNet generalizes beyond the specific synthetic degradation assumed in Eq. (2).

major comments (4)
  1. [Section 3.1 and Eq. (2)] The manuscript never specifies how the 64x64 low-resolution test inputs for Table 1 were generated. Eq. (2) models the forward operator as a 4x block-averaging filter, and the likelihood step in Eqs. (5)-(6) is computed from its SVD. Section 3.1 only says that the 1024x1024 volumes were 'downsampled to 256x256 resolution by sparsely sampling along the depth and fast scanning axes with a step size of 4,' which describes sparse decimation, not block averaging. If the 64x64 inputs were produced by sparse decimation or any operator other than block averaging, the likelihood step would be solving a different inverse problem than the one that generated the test data, and the reported PSNR/SSIM/LPIPS gains over 2D-UNet in Table 1 could be an artifact of that mismatch. The authors should state explicitly how the LR test B-scans were created, and for the 'under sampled test B-scans' mentioned in Section 3.1, they should describe the acquisition protocol and the forward model used in the likelihood update.
  2. [Section 3.3 and Table 1] The text in Section 3.3 states that the evaluation metrics are 'PSNR, SSIM, and root mean square error (RMSE),' but Table 1 reports LPIPS and no RMSE column. Section 3.5 similarly claims results are reported in terms of 'PSNR, SSIM, and RMSE.' This inconsistency must be corrected. In addition, the abstract's claim of 'better noise suppression' is not tied to any quantitative metric; the authors should either add a noise-specific metric or qualify the claim as qualitative.
  3. [Section 3.3 and Table 1] The diffusion-based methods are stochastic, and the paper states that 100 random samples were drawn and averaged for each image, yet Table 1 reports only point estimates with no variance, confidence intervals, or statistical significance tests. Without such uncertainty quantification, the observed differences between PnP-DM variants and against 2D-UNet cannot be assessed. Reporting mean and standard deviation across the 100 samples, or across the 100 test images, would substantially strengthen the quantitative claims.
  4. [Sections 2.4 and 2.2] The reconstruction depends on several free parameters that are only stated as fixed values: the coupling parameter schedule (rho_0 = 10, rho_min = 0.3, alpha = 0.9), the measurement noise variance sigma_y^2, and the number of PnP-DM iterations. No sensitivity analysis or model-selection criterion is provided for these choices. Since the reported advantage over 2D-UNet could in principle depend on these settings, the authors should show that the qualitative conclusions are stable across reasonable variations of these parameters.
minor comments (5)
  1. [Introduction] There are typos in the Introduction, including 'interative plug-and-play' and 'imaing,' which should be corrected.
  2. [Eq. (7)] Equation (7) is garbled in the manuscript; the SDE is not written in a readable form. Please rewrite it with standard notation for the score function and the noise schedule.
  3. [Fig. 1 caption] The caption says the diffusion prior is 'trained on the same dataset,' but Section 3.1 states that the training data are ex vivo volumes while evaluation is on live-fish validation data. Please clarify which dataset is used for training the prior.
  4. [Section 3.2] The implementation details mention that a 'pre-trained score function' trained with VP-SDE is converted to EDM formulation, but the experiments also use VE-SDE and iDDPM variants. Please specify the source of the pretrained weights for each variant.
  5. [Section 3.1] The phrase 'under sampled test B-scans' is used without explaining how undersampling was performed physically. Please provide acquisition details or, if the test inputs were simulated, state so explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PnP-DM reconstruction is posterior sampling under an explicitly stated forward model with a diffusion prior trained on separate OCT volumes, and the headline comparison to 2D-UNet is an empirical evaluation rather than a derivation from the method's own assumptions.

full rationale

The derivation chain is self-contained. The forward model is stated in Eq. (2) as block-averaging downsampling with Gaussian noise, and the likelihood step in Eqs. (3)-(6) follows from that model by standard Gaussian conditioning. The prior step is an EDM diffusion denoiser, trained on separate ex vivo fish-eye OCT volumes as described in Section 3.1, with test B-scans taken from live fish-eye volumes; no test-set HR images are used to fit the diffusion prior or the likelihood parameters. The claimed advantage over 2D-UNet is an empirical result reported in Table 1, not an algebraic consequence of the model, and the FFHQ-prior ablation in Section 3.6 provides an external control showing that the OCT-trained prior is what drives the structural fidelity. The only potentially confusing statements are the abstract's mention of a 'deep learning-based up-sampling pipeline' and the Fig. 1 caption saying the prior is 'trained on the same dataset'; these are data-preparation/domain-matching descriptions, not circular definitions. Likewise, the apparent mismatch between Eq. (2)'s block-averaging operator and Section 3.1's 'sparsely sampling' language is an evaluation-protocol ambiguity and a correctness risk, but it does not make the prediction equivalent to the model's inputs by construction. The citation of Wu et al. [23] for the PnP-DM algorithm includes an overlapping author, but the algorithm is applied rather than derived here, and the quantitative evaluation is independent of the citation's internal correctness; hence the self-citation is not load-bearing in a circular sense. No step in the paper reduces a claimed prediction to a fitted parameter or to a self-citation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (rho schedule, iterations) and on the assumption that the synthetic block-averaging forward model faithfully represents real OCT sparse acquisition. No new physical entities are introduced.

free parameters (3)
  • Coupling parameter decay schedule (rho_0, rho_min, alpha) = 10, 0.3, 0.9
    Exponential decay schedule for the coupling parameter rho in the Split Gibbs Sampler; chosen by hand without sensitivity analysis, yet it influences mixing and convergence.
  • Measurement noise variance sigma_y^2 = Not specified
    Appears in the likelihood model (Eq. 2) but no numeric value or estimation procedure is given; required to define the data consistency step.
  • Number of PnP-DM iterations = 100
    The test protocol runs 100 iterations to 'ensure convergence', but no convergence diagnostics are shown.
assumptions (5)
  • domain assumption Sparse OCT measurements follow the linear block-averaging model y ~ N(P_f x, sigma_y^2 I).
    Invoked in Eq. (2); all experiments generate low-resolution inputs from high-resolution images with this model, so the claim depends on this idealization.
  • domain assumption A diffusion model trained on high-resolution OCT fish cornea images is an effective prior for the target distribution.
    The EDM prior is trained on ex vivo fish cornea data and applied to in vivo fish eye data; the paper assumes this transfer works.
  • standard math The PnP-DM algorithm and its theoretical guarantees from Wu et al. [23] are correct.
    Section 2 explicitly attributes the framework to Wu et al.; the paper does not re-derive the algorithm.
  • standard math The singular value decomposition of the block-averaging operator is computed correctly following Kawar et al. [24].
    Used in the likelihood step to reduce computational cost; no verification is provided that the SVD implementation is accurate for the chosen problem size.
  • domain assumption Downsampling by sparse sampling along depth and fast axes with step 4 produces representative high-resolution training images.
    Section 3.1 describes creating 256x256 high-resolution images from 1024x1024 volumes; the assumption is that these are good targets for super-resolution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Super-Resolution Optical Coherence Tomography Using Diffusion Model-Based Plug-and-Play Priors." pith.science (2026). https://pith.science/paper/PSGS3AV6

@misc{pith2026250514916,
  author       = {Pith},
  title        = {Pith review of: Super-Resolution Optical Coherence Tomography Using Diffusion Model-Based Plug-and-Play Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSGS3AV6}},
  note         = {Machine review of arXiv:2505.14916}
}
read the original abstract

We propose an OCT super-resolution framework based on a plug-and-play diffusion model (PnP-DM) to reconstruct high-quality images from sparse measurements (OCT B-mode corneal images). Our method formulates reconstruction as an inverse problem, combining a diffusion prior with Markov chain Monte Carlo sampling for efficient posterior inference. We collect high-speed under-sampled B-mode corneal images and apply a deep learning-based up-sampling pipeline to build realistic training pairs. Evaluations on in vivo and ex vivo fish-eye corneal models show that PnP-DM outperforms conventional 2D-UNet baselines, producing sharper structures and better noise suppression. This approach advances high-fidelity OCT imaging in high-speed acquisition for clinical applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 27 canonical work pages

  1. [1]

    Automatic and real-time tissue sensing for autonomous intestinal anastomosis using hybrid MLP-DC-CNN classifier-based optical coherence tomography,

    Y . Wang, S. Wei, R. Zuo, et al., "Automatic and real-time tissue sensing for autonomous intestinal anastomosis using hybrid MLP-DC-CNN classifier-based optical coherence tomography," Biomed. Opt. Express, vol. 15, no. 4, p. 2543, 2024

  2. [2]

    Reimagining partial thickness keratoplasty: An eye mountable robot for autonomous big bubble needle insertion

    Y . Wang, J. D. Opfermann, J. Yu, et al., "Reimagining partial thickness keratoplasty: An eye mountable robot for autonomous big bubble needle insertion," arXiv preprint arXiv:2410.14577, 2024

  3. [3]

    Topology-based deep-learning segmentation method for deep anterior lamellar keratoplasty (DALK) surgical guidance using M -mode OCT data,

    J. Yu, H. Yi, Y . Wang, et al., "Topology-based deep-learning segmentation method for deep anterior lamellar keratoplasty (DALK) surgical guidance using M -mode OCT data," in Proc. SPIE 13310, Optical Fibers and Sensors for Medical Diagnostics, Treatment, and Environmental Applications XXV , p. 1331003, 2025

  4. [4]

    Kalman filter/deep-learning hybrid automatic boundary tracking of optical coherence tomography data for deep anterior lamellar keratoplasty (DALK),

    H. Yi, J. Yu, Y . Wang, et al., "Kalman filter/deep-learning hybrid automatic boundary tracking of optical coherence tomography data for deep anterior lamellar keratoplasty (DALK)," in Proc. SPIE 13310, p. 133100A, 2025

  5. [5]

    The neural networks-based needle detection for medical retinal surgery,

    J. Xu, J. Y u, J. Yao, et al., "The neural networks-based needle detection for medical retinal surgery," in ICCAID, vol. 12605, pp. 674-678. SPIE, 2023

  6. [6]

    Live porcine eye model studies of subretinal injection using handheld endoscopy OCT integrated injector ,

    Y . Wang, MS. Singh, J. Y u, et al., "Live porcine eye model studies of subretinal injection using handheld endoscopy OCT integrated injector ," in Investigative Ophthalmology & Visual Science , 65, no. 7, pp. 5499-5499, 2024

  7. [7]

    Distributed scanning volumetric SDOCT for motion corrected corneal biometry,

    R. P. McNabb, et al., "Distributed scanning volumetric SDOCT for motion corrected corneal biometry," Biomed. Opt. Express, vol. 3, no. 9, pp. 2050–2065, 2012

  8. [8]

    Review on retrospective procedures to correct retinal motion artefacts in OCT imaging,

    L. Sánchez Brea, et al., "Review on retrospective procedures to correct retinal motion artefacts in OCT imaging," Appl. Sci., vol. 9, no. 13, pp. 2700–2727, 2019

Show all 29 references
  1. [9]

    Handheld simultaneous scanning laser ophthalmoscopy and optical coherence tomography system,

    F. LaRocca, et al., "Handheld simultaneous scanning laser ophthalmoscopy and optical coherence tomography system," Biomed. Opt. Express, vol. 4, no. 11, pp. 2307–2321, 2013

  2. [10]

    Contactless optical coherence tomography of the eyes of freestanding individuals with a robotic scanner,

    M. Draelos, P. Ortiz, R. Qian, et al., "Contactless optical coherence tomography of the eyes of freestanding individuals with a robotic scanner," Nat. Biomed. Eng., vol. 5, pp. 726–736, 2021

  3. [11]

    Higher-order regression three-dimensional motion-compensation method for real-time optical coherence tomography volumetric imaging of the cornea,

    R. Zuo, K. Irsch, and J. U. Kang, "Higher-order regression three-dimensional motion-compensation method for real-time optical coherence tomography volumetric imaging of the cornea," J. Biomed. Opt., vol. 27, no. 6, p. 066006, 2022

  4. [12]

    Deep-learning-based 2.5D flow field estimation for maximum intensity projections of 4D optical coherence tomography,

    M.-H. Laves, L. A. Kahrs, T. Ortmaier, et al., "Deep-learning-based 2.5D flow field estimation for maximum intensity projections of 4D optical coherence tomography," in Medical Imaging 2019: Image-Guided Procedures, Robotic Interventions, and Modeling, SPIE, p. 26, 2019

  5. [13]

    Spatio -temporal deep learning methods for motion estimation using 4D OCT image data,

    M. Bengs, N. Gessert, M. Schlüter, et al., "Spatio -temporal deep learning methods for motion estimation using 4D OCT image data," Int. J. Comput. Assist. Radiol. Surg., vol. 15, no. 6, pp. 943– 952, 2020

  6. [14]

    High resolution in vivo 4D -OCT fish eye imaging using deep learning,

    R. Zuo, S. Wei, Y . Wang, et al., "High resolution in vivo 4D -OCT fish eye imaging using deep learning," in Optica Biophotonics Congress: Biomedical Optics 2024, paper CTu3E.4

  7. [15]

    Wavelets-Based Digital Watermarking Scheme for Medical Image Data,

    J. Xu, J. Yu, H. Tang, "Wavelets-Based Digital Watermarking Scheme for Medical Image Data," in ICCGIV, pp. 98-192. IEEE, 2022

  8. [16]

    High-resolution in vivo 4D-OCT fish-eye imaging using 3D-UNet with multi-level residue decoder,

    R. Zuo, S. Wei, Y . Wang, et al., "High-resolution in vivo 4D-OCT fish-eye imaging using 3D-UNet with multi-level residue decoder," Biomed. Opt. Express, vol. 15, pp. 5533–5546, 2024

  9. [17]

    GAN prior embedded network for blind face restoration in the wild,

    T. Yang, P. Ren, X. Xie, and L. Zhang, "GAN prior embedded network for blind face restoration in the wild," in Proc. CVPR, pp. 672–681, 2021

  10. [18]

    Towards real-world blind face restoration with generative facial prior,

    X. Wang, Y . Li, H. Zhang, and Y . Shan, "Towards real-world blind face restoration with generative facial prior," in Proc. CVPR, pp. 9168–9178, 2021

  11. [19]

    V olumetric (3D) compressive sensing spectral domain optical coherence tomography,

    D. Xu, Y . Huang, and J. U. Kang, "V olumetric (3D) compressive sensing spectral domain optical coherence tomography," Biomed. Opt. Express, vol. 5, no. 11, p. 3921, 2014

  12. [20]

    Super-resolution technology to simultaneously improve optical & digital resolution of optical coherence tomography via deep learning,

    S. Cao, X. Yao, N. Koirala, et al., "Super-resolution technology to simultaneously improve optical & digital resolution of optical coherence tomography via deep learning," in Proc. IEEE EMBC, pp. 1879–1882, 2020

  13. [21]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, "Denoising diffusion probabilistic models," in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 6840–6851, 2020

  14. [22]

    High -resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, et al., "High -resolution image synthesis with latent diffusion models," in Proc. CVPR, pp. 10684–10695, 2022

  15. [23]

    Principled probabilistic imaging using diffusion models as plug- and-play priors

    Z. Wu, Y . Sun, Y . Chen, et al., "Principled probabilistic imaging using diffusion models as plug- and-play priors." in Adv. Neural Inf. Process. Syst., vol. 37, pp. 118389-118427, 2024

  16. [24]

    Denoising diffusion restoration models,

    B. Kawar, M. Elad, S. Ermon, and J. Song, "Denoising diffusion restoration models," in Adv. Neural Inf. Process. Syst., 2022

  17. [25]

    Split -and-augmented Gibbs sampler —application to large-scale inference problems,

    M. V ono, N. Dobigeon, and P. Chainais, "Split -and-augmented Gibbs sampler —application to large-scale inference problems," IEEE Trans. Signal Process., vol. 67, no. 6, pp. 1648–1661, 2019

  18. [26]

    Elucidating the design space of diffusion -based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, "Elucidating the design space of diffusion -based generative models," arXiv preprint arXiv:2206.00364, 2022

  19. [27]

    Score -based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl -Dickstein, D. P. Kingma, et al., "Score -based generative modeling through stochastic differential equations," in Proc. ICLR, 2021

  20. [28]

    PyTorch : An imperative style, high -performance deep learning library,

    A. Paszke, S. Gross, F. Massa, et al., "PyTorch : An imperative style, high -performance deep learning library," in Adv. Neural Inf. Process. Syst., vol. 32, pp. 8024–8035, 2019

  21. [29]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," arXiv preprint arXiv:1412.6980, 2017

Pith tools

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