{"id":"074f02a2-4b6d-47a7-a8aa-254fe0a19b5f","arxiv_id":"2505.14916","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion-based plug-and-play prior improves OCT image super-resolution over a 2D-UNet on fish-eye cornea data, yet the test setup uses synthetic block-averaged degradation.","lead":"This paper applies a plug-and-play diffusion model to super-resolution of optical coherence tomography (OCT) corneal images. The authors report sharper reconstructions than a standard 2D-UNet baseline, but the evaluation relies on synthetic downsampled inputs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LR test inputs for Table 1 are never specified as block-averaged; §3.1 says 'sparsely sampling,' so the forward model in Eq. (2) may not match the actual degradation, undermining the reported PnP-DM advantage.","rationale":"The reader's verdict is CONDITIONAL and largely justified. The reader identified the forward model as the weakest assumption, but focused on synthetic vs real degradation. My stress-test finds a more immediate, internal version of the same issue: the text does not verify that the actual LR generation used in Table 1 matches the P_f in Eq. (2). §3.1's 'sparsely sampling' language conflicts with the block-averaging definition, and the 64×64 generation step is not described at all. This is load-bearing because the likelihood update (Eqs. 5–6) is derived from P_f; using a mismatched operator in evaluation makes the quantitative comparison uninterpretable. The paper's independent support is limited: no code, no data release, no machine-checked proofs; the ablation with FFHQ prior is a nice sanity check but does not resolve the forward-model ambiguity. I therefore agree the paper should be CONDITIONAL until the operator is clarified and the comparison re-run if necessary. Since the reader already reached this verdict, no adjustment is needed.","tokens_in":6522,"tokens_out":6234,"duration_ms":50915,"concrete_test":"Request the authors' code or ask them to state precisely, for Table 1, how each 64×64 LR image was generated from its 256×256 HR counterpart. Compare that operator with the block-averaging P_f used in Eq. (2). If they differ, re-run the PnP-DM inference with the correct operator (e.g., a decimation matrix) and recompute Table 1; the claim stands only if PnP-DM still outperforms 2D-UNet under the correct forward model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—PnP-DM beats 2D-UNet on Table 1—depends on the 64×64 low-resolution test inputs being generated by the exact forward operator used in the likelihood step. Eq. (2) defines P_f as a 4× block-averaging downsampler, and §2.2 computes the data-consistency update from that operator. However, §3.1 only states that high-resolution images were 'down sampled to 256×256 resolution by sparsely sampling along the depth and fast scanning axes with a step size of 4'; it never states how the 64×64 LR inputs for the super-resolution task were created. If those LR inputs were generated by sparse decimation rather than block averaging, then the likelihood term in Eq. (5)-(6) uses the wrong forward model. The PnP-DM reconstructions would then be solving a different inverse problem than the one that produced the test data, and the reported PSNR/SSIM/LPIPS gains over 2D-UNet could reflect an artefact of this mismatch (e.g., block-averaging assumption smoothing the decimated input). The paper also does not report the operator used for the real 'under sampled test B-scans' mentioned in §3.1, so the evaluation protocol is ambiguous. This is the most load-bearing issue because it directly affects the validity of the only quantitative evidence for the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6827,"tokens_out":4482,"duration_ms":37186,"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":[{"comment":"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.","section":"Section 3.1 and Eq. (2)"},{"comment":"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.","section":"Section 3.3 and Table 1"},{"comment":"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.","section":"Section 3.3 and Table 1"},{"comment":"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.","section":"Sections 2.4 and 2.2"}],"minor_comments":[{"comment":"There are typos in the Introduction, including 'interative plug-and-play' and 'imaing,' which should be corrected.","section":"Introduction"},{"comment":"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.","section":"Eq. (7)"},{"comment":"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.","section":"Fig. 1 caption"},{"comment":"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.","section":"Section 3.2"},{"comment":"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.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a domain application of the existing PnP-DM method of Wu et al., so the novelty lies mainly in the OCT context and the empirical evaluation. The central derivation appears sound, but the evaluation protocol must be clarified and strengthened before publication. The self-citation count is high but not inappropriate given the authors' prior work in the area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does exactly what it says on the tin: it applies Wu et al.'s plug-and-play diffusion framework to OCT super-resolution, compares four diffusion priors, and shows PnP-DM beating a 2D-UNet on fish-eye cornea scans. The attribution is clean, the math is standard and correctly borrowed, and the ablation with a face-trained prior is a nice touch—it demonstrates that a domain-matched prior matters and that even a mismatched diffusion prior preserves some structure. That empirical comparison is a genuine contribution, albeit a narrow one.\n\nThe load-bearing problem is the forward model ambiguity. Eq. (2) defines the degradation as 4x block averaging, and the likelihood step uses the SVD of that operator. But Section 3.1 only says the training data were downsampled by \"sparsely sampling\" with step size 4, and it never states how the 64x64 test inputs for Table 1 were generated. If those test inputs came from sparse decimation rather than block averaging, then the PnP-DM is solving a different inverse problem than the one that produced the data, and the reported PSNR/SSIM/LPIPS gains could be partly an artifact of the mismatch. The stress-test note is right to flag this. It is not fatal—the authors may well have used block averaging for the test set—but the paper needs to say so explicitly and, ideally, validate on real undersampled acquisitions.\n\nThe other issues are minor but real: the text says results are reported as RMSE, while Table 1 shows LPIPS; there are no error bars or variance estimates; the abstract overclaims motion compensation and clinical relevance when the evaluation is on fish-eye data; and no code or data are provided. All fixable in revision.\n\nThis paper deserves a serious referee. It is a legitimate application of a proven method to a clinically relevant problem, the evidence is plausible once the degradation protocol is pinned down, and the literature is cited fairly. I'd recommend sending it out, with the expectation that the authors tighten the evaluation description and fix the small inconsistencies.","headline":"Solid, honest application of PnP-DM to OCT super-resolution, but the test-set degradation model is underspecified and that undercuts the headline numbers.","tokens_in":7378,"tokens_out":3910,"would_cite":false,"duration_ms":33186,"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":"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…","keywords":["optical coherence tomography","super-resolution","plug-and-play diffusion model","MCMC posterior sampling","sparse measurements","corneal imaging","image inverse problems"],"falsifier":"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.","tokens_in":6331,"feed_emoji":"👁","tokens_out":3946,"duration_ms":33514,"temperature":0.7,"pith_summary":"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.","feed_headline":"Plug-and-play diffusion beats UNet on OCT super-resolution","feed_subtitle":"MCMC sampling with a diffusion prior reconstructs 4x-undersampled corneal B-scans with sharper boundaries and less speckle.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Introduces the PnP-DM plug-and-play diffusion framework that this paper applies to OCT super-resolution.","marker":"[23]"},{"why":"Supplies the SVD-based implementation of the block-averaging forward operator used in the likelihood step.","marker":"[24]"},{"why":"Provides the Split Gibbs Sampler that alternates the likelihood and prior steps in the reconstruction loop.","marker":"[25]"},{"why":"Defines the EDM diffusion formulation used for the prior denoising step.","marker":"[26]"},{"why":"Provides the score-based SDE models, including VP-SDE and VE-SDE, that serve as diffusion priors after conversion to EDM.","marker":"[27]"},{"why":"Supplies the iDDPM diffusion prior variant and the original denoising diffusion probabilistic model formulation.","marker":"[21]"}],"fun_headline_variants":["Diffusion prior beats UNet on OCT super-resolution","Plug-and-play diffusion sharpens sparse OCT images","MCMC diffusion prior improves OCT resolution beyond UNet","4x-undersampled OCT super-resolved with diffusion model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion prior beats UNet on OCT super-resolution","Plug-and-play diffusion sharpens sparse OCT images","MCMC diffusion prior improves OCT resolution beyond UNet","4x-undersampled OCT super-resolved with diffusion model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1683,"prompt_tokens":814,"completion_tokens":869,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":804}},"tokens_in":430,"tokens_out":869,"duration_ms":8467,"temperature":1.0,"reasoning_tokens":804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:26:55.369898+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Principled probabilistic imaging using diffusion models as plug- and-play priors","cited_arxiv_id":null,"evidence_quote":"Introduces the PnP-DM plug-and-play diffusion framework that this paper applies to OCT super-resolution."},{"cited_title":"Denoising diffusion restoration models,","cited_arxiv_id":null,"evidence_quote":"Supplies the SVD-based implementation of the block-averaging forward operator used in the likelihood step."},{"cited_title":"Split -and-augmented Gibbs sampler —application to large-scale inference problems,","cited_arxiv_id":null,"evidence_quote":"Provides the Split Gibbs Sampler that alternates the likelihood and prior steps in the reconstruction loop."},{"cited_title":"Score -based generative modeling through stochastic differential equations,","cited_arxiv_id":null,"evidence_quote":"Provides the score-based SDE models, including VP-SDE and VE-SDE, that serve as diffusion priors after conversion to EDM."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the iDDPM diffusion prior variant and the original denoising diffusion probabilistic model formulation."}],"review_version":1}