{"id":"6aa98cd1-5eca-40b8-aa83-0bcb6f9b8c4e","arxiv_id":"2509.10341","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GARD, a Gamma-based diffusion denoiser with a noise-reduced fidelity term, improves OCT despeckling over baselines on a paired dataset.","lead":"A new denoising method for retinal OCT images combines a Gamma-distributed diffusion model with a fidelity term that guides reconstruction toward a smoothed version of the input. On paired noisy/less-noisy scans it reports modest but statistically significant gains over existing methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA comparison is not yet fair: external deep baselines are off-the-shelf pretrained, and the only fairly trained baseline margin (vs DDGM) is 0.09 dB PSNR in Table 1; retraining could erase the gap.","rationale":"The paper's contribution is a plausible adaptation of DDGM to OCT denoising, and the paired noisy/less-noisy evaluation with registered B-scans is a genuine strength. The ablation structure (DDPM vs DDGM, with and without NRFT) is the right way to isolate the Gamma mechanism, and the authors are transparent that external baselines are used off-the-shelf. My concern is specifically about the load-bearing comparative claim. For 'outperforms state-of-the-art' to be supported, SOTA methods must be allowed to perform at their best on the target domain. Off-the-shelf pretrained weights do not satisfy this. The small margin over the fairly trained DDGM (0.09 dB from the table) makes it especially plausible that retrained SCUNet would close the gap. The numerical inconsistencies in Section 4 (0.31/0.34/0.23 dB versus Table 1's 0.15/0.40/0.09 dB) reinforce the need for a careful recheck. This does not imply the method is invalid; the within-family comparisons and qualitative results may be correct. The reader's conditional verdict is reasonable, and the required condition is simply to retrain baselines and correct the reported numbers.","tokens_in":8395,"tokens_out":9523,"duration_ms":82042,"concrete_test":"Retrain SCUNet (and ideally N2V2 and Speckle2Speckle) on the same 2000-volume investigational OCT training set used for GARD, with identical optimizer settings and number of iterations as described in Section 3, then evaluate on the paired Spectralis dataset and recompute PSNR, SSIM, MSE, and the Wilcoxon p-values. Additionally, recompute the Section 4 PSNR improvement claims from the Table 1 means. If any retrained baseline matches or exceeds GARD, or if the quoted improvements do not reproduce, the central SOTA claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GARD 'significantly outperforms state-of-the-art deep learning models' is not yet established, because the external deep baselines SCUNet, N2V2, and Speckle2Speckle are evaluated with their publicly available pretrained weights rather than trained or fine-tuned on the OCT domain (Section 3, Baselines). SCUNet was trained on natural-image denoising data; using it directly on retinal OCT B-scans underestimates what the same architecture could achieve after domain adaptation. The only comparisons on equal footing are the diffusion variants, and there GARD's margin over vanilla DDGM is 0.09 dB PSNR (28.25 vs 28.16 in Table 1), not the 0.23 dB stated in Section 4. The same section quotes a 0.31 dB gain over SCUNet, while Table 1 implies 0.15 dB. If the external baselines are retrained on GARD's 2000-volume training set, the reported SOTA margin could plausibly disappear, leaving only a small within-family improvement. The paired-data evaluation itself is a strength; the weakness is the asymmetry of the comparison and the inconsistency of the quoted effect sizes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents GARD, a denoising method for retinal OCT B-scans that combines a Denoising Diffusion Gamma Model (DDGM) with a Noise-Reduced Fidelity Term (NRFT). The forward process in Eq. (1) adds zero-mean Gamma noise, and the reverse process in Eq. (2) follows Nachmani et al.; the NRFT replaces fidelity to the noisy input with fidelity to an NLM-filtered version (Eq. 3) via a proximal update (Eq. 4). DDIM-style deterministic sampling is used to accelerate inference. The method is trained on 2000 volumes from an investigational device and evaluated on 13 volumes with paired noisy (single sweep) and less-noisy (30-frame ART) B-scans from a commercial Spectralis device. Table 1 reports mean PSNR/SSIM/MSE for GARD, baselines (NLM, SCUNet, N2V2, Speckle2Speckle), and diffusion ablations, with Wilcoxon tests. GARD ranks first on all metrics, and qualitative results on cross-vendor data are shown. The authors claim significant improvement over state-of-the-art methods and sharper detail preservation.","tokens_in":8649,"tokens_out":6609,"duration_ms":50954,"significance":"The topic is relevant: OCT despeckling is an active clinical imaging problem, and paired noisy/less-noisy data with near-perfect registration are a valuable resource. The idea of adapting Gamma diffusion models to medical imaging is a novel extension, and the NRFT is a sensible mechanism to avoid noise reinforcement. The strengths of the paper include the paired-data evaluation design, the ablation across diffusion variants and fidelity terms, the use of a Wilcoxon test, and the release of source code. However, the significance is currently limited by the small single-device evaluation set and, more importantly, by the unfair comparison with external deep baselines that were not retrained on OCT. The actual margin over the strongest fairly trained baseline (vanilla DDGM) is 0.09 dB PSNR in Table 1, which is small and may not be practically meaningful. If the external baselines were retrained on the authors' training set, the state-of-the-art claim could plausibly disappear.","major_comments":[{"comment":"The deep-learning baselines SCUNet, N2V2, and Speckle2Speckle are evaluated using publicly available pretrained weights without any retraining or fine-tuning on OCT data, as stated in Section 3. This is not an equal-footing comparison: SCUNet was trained on natural images, and it is unsurprising that a model trained on 2000 OCT volumes outperforms it off-the-shelf. The claim in Section 4 that GARD 'significantly outperforms state-of-the-art deep learning models' is therefore not established. Please either retrain these baselines on the same training set (or a comparable OCT corpus) and report the results, or explicitly rephrase the claim as a comparison against off-the-shelf pretrained models.","section":"Section 3 (Baselines), Table 1"},{"comment":"The effect sizes quoted in Section 4 are inconsistent with Table 1. The text states PSNR improvements of 0.31 dB over SCUNet, 0.34 dB over standard DDPM, and 0.23 dB over vanilla DDGM, but Table 1 gives GARD=28.25, SCUNet=28.10, DDPM=27.85, DDGM=28.16 (differences of 0.15, 0.40, and 0.09 dB, respectively). The 0.31 and 0.23 dB values do not match the table; please correct the text and discuss the actual margins, especially the small 0.09 dB gain over the vanilla DDGM, which is the only fairly trained diffusion baseline comparable to GARD.","section":"Section 4 vs Table 1"},{"comment":"The statistical significance tests are reported as p<0.01 for all metrics and methods, but the paper does not specify the unit of analysis. If the 247 B-scans (13 volumes × 19 B-scans) are treated as independent samples, the test ignores within-volume correlation and overstates significance; if the unit is 13 volumes, the sample size is very small and the test's power is low. Please state the unit, report exact p-values, and, if B-scans are used, account for clustering (e.g., by volume-level averaging or a mixed-effects model).","section":"Section 4 (Wilcoxon tests)"},{"comment":"The central modeling assumption is that display-ready OCT noise is approximately additive and Gamma-distributed, but this is not empirically validated. The hyperparameter θ0 is set to 0.1 because it 'resulted in a noise most similar to typical OCT noise' (Section 3), which is subjective. The paper should provide a quantitative comparison of the assumed Gamma distribution against measured residual noise statistics (e.g., on paired noisy/less-noisy B-scans), and also demonstrate that the reverse process removes real speckle beyond what the NRFT guide already achieves. Without such validation, the claimed advantage of the Gamma model over the Gaussian DDPM is only supported by a 0.09 dB PSNR difference in Table 1.","section":"Section 2 (Eq. 1) and Section 3 (Implementation)"}],"minor_comments":[{"comment":"The displayed optimization is garbled: 'z+e ˜y−z +µ' does not define a proper objective. Please rewrite the proximal update with explicit notation for the variables and the quadratic term.","section":"Section 2, Eq. (4)"},{"comment":"For the quantitative evaluation set, specify the total number of B-scan pairs (13 volumes × 19 B-scans) and clarify whether the metrics in Table 1 are computed per B-scan and then averaged, or per volume.","section":"Section 3, Datasets"},{"comment":"The sentence 'GARD is significantly better for all metrics and methods with p<0.01' could be made more informative by reporting exact p-values or a supplementary table, and by stating whether any multiple-comparison correction was applied.","section":"Section 4 (statistical reporting)"},{"comment":"The definition of \\bar{g}_t and its shape parameter \\bar{k}_t should be spelled out; it is only implicit that the sum of Gamma variables with the same scale is Gamma-distributed, and this property should be stated explicitly for the reader.","section":"Section 2, Eq. (2)"},{"comment":"The column label 'Original Crop' is confusing because the actual crop is shown in the second column; adjust the caption or labels to match the content.","section":"Figure 1"},{"comment":"The abbreviation for the fidelity term in [17] is written both as 'CPDM' and 'CDPM'; use one notation consistently throughout the paper.","section":"Section 3 and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the unfair baseline comparison and the inconsistent effect sizes, which together undermine the central claim of significant improvement over state-of-the-art deep learning models. The authors should be required to either retrain the external baselines on their OCT data or substantially temper the claims. The small 0.09 dB PSNR gain over the vanilla DDGM also raises the question of whether the practical benefit is meaningful; the paper would be more credible if it reported confidence intervals and focused on the NRFT contribution. The statistical test unit needs clarification before the significance claims can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The key point: GARD combines Denoising Diffusion Gamma Models with a Noise-Reduced Fidelity Term that uses an NLM-filtered version of the noisy input as a guide. That combination is new, and the paired noisy/less-noisy dataset with registered B-scans is a solid evaluation setup. The code is public, which helps reproducibility.\n\nThe ablation is the best part. NRFT improves the Gamma model but slightly hurts the Gaussian DDPM. That is a clean empirical demonstration of a match between the noise model and the fidelity guidance. The writing is clear, and the single-device limitation is acknowledged up front.\n\nNow the soft spots, in order of importance. First, the SOTA comparison is not yet fair. SCUNet, N2V2, and Speckle2Speckle are evaluated with their pretrained weights, not trained or fine-tuned on OCT. That is an off-the-shelf test, not a head-to-head comparison. The claim that GARD 'significantly outperforms state-of-the-art deep learning models' is therefore overstated. Retraining those baselines on the same 2000-volume training set could plausibly erase the gap.\n\nSecond, the effect sizes quoted in Section 4 do not match Table 1. The text reports 0.31 dB over SCUNet and 0.23 dB over vanilla DDGM; Table 1 gives 0.15 dB and 0.09 dB, respectively. That is a concrete inconsistency, and it matters because the true margin over the next-best diffusion variant is tiny.\n\nThird, the evaluation is 13 volumes from one device type. The Gamma-noise assumption for post-processed display images is plausible but not empirically verified; a simple fit of noise statistics on their own data would strengthen it. The Wilcoxon tests are fine, but with 247 B-scans from 13 volumes, treating each B-scan as independent inflates the effective sample size.\n\nConclusion: this is a reasonable application paper with one genuinely interesting ablation and a reproducible setup. The core idea is worth exploring, but the current evidence supports a modest within-family improvement, not a SOTA claim. I would send it to peer review, but require retrained baselines, corrected numbers, and preferably multi-device validation. For readers working on diffusion-based medical image restoration, it is worth a look, but I would not cite it as evidence for a large gain.","headline":"A solid OCT denoising paper with a novel DDGM+NRFT combination, but the SOTA comparison is unfair and the real gain over vanilla DDGM is ~0.1 dB; worth refereeing after major fixes.","tokens_in":9204,"tokens_out":4414,"would_cite":false,"duration_ms":37787,"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":"GARD claims that replacing the Gaussian noise assumption in diffusion-based denoising with a Gamma model of speckle, and guiding the reverse process with a non-locally filtered reference, substantially improves retinal OCT despeckling.","keywords":["optical coherence tomography","speckle denoising","diffusion probabilistic models","Gamma distribution","retinal imaging","image restoration","deep learning","noise-reduced fidelity term"],"falsifier":"A decisive check is to repeat the paired noisy/less-noisy evaluation on raw, linear-scale OCT B-scans before log or fourth-root compression, where speckle is known to be multiplicative. If GARD's advantage over the Gaussian DDPM shrinks or vanishes there, the additive-Gamma model is doing the load-bearing work; if the advantage persists, the NRFT guide, not the Gamma mechanism, drives the gains.","tokens_in":8196,"feed_emoji":"👁️","tokens_out":6705,"duration_ms":52780,"temperature":0.7,"pith_summary":"GARD is a diffusion-based denoiser for retinal OCT built on the premise that speckle noise in display-ready OCT is Gamma-distributed rather than Gaussian, and that a diffusion model which matches this statistic will remove noise without erasing fine anatomy. The paper claims that pairing this Gamma diffusion model with a Noise-Reduced Fidelity Term, a guide image obtained by non-local-means filtering the noisy scan, gives the best despeckling on a paired dataset of single-sweep and 30-frame-averaged B-scans, beating classical filters and state-of-the-art deep-learning baselines on PSNR, SSIM, and MSE. The authors argue that the fidelity term matters because enforcing consistency with the original noisy image, as prior diffusion denoisers do, tends to re-inject noise, whereas guiding with a less-noisy reference preserves structure while letting the model synthesize realistic high-frequency detail. If right, the method points to a practical path for getting clinically useful OCT quality from fewer acquisitions or lower-cost devices.","feed_headline":"Beats OCT denoisers with a Gamma-noise diffusion model","feed_subtitle":"Assuming speckle is Gamma-distributed and steering with a noise-reduced reference beats existing OCT denoisers.","key_machinery":"The mechanism is the Gamma diffusion process from Eq. (1), where each forward step adds a centered Gamma-distributed variable with shape $k_t = \\beta_t/(\\alpha_t \\theta_0^2)$ and scale $\\theta_t = \\sqrt{\\bar{\\alpha}_t} \\theta_0$, chosen so that sums of independent Gamma variables with common scale remain Gamma-distributed, permitting closed-form sampling of $x_t$ from $x_0$. The reverse process in Eq. (2), with a U-Net trained to predict the noise component, is accelerated by setting $\\sigma_t = 0$ (adapting the Denoising Diffusion Implicit Model approach for deterministic sampling), letting the denoiser start at $t=70$ and skip timesteps. The Noise-Reduced Fidelity Term (Eqs. 3-4) replaces the noisy input with its NLM-filtered version $\\tilde{y} = \\mathrm{NLM}(y)$ and, at each reverse step, solves a Newton optimization that balances the guide image against the diffusion output, retaining low-frequency anatomy while the diffusion fills in high-frequency detail.","core_discovery":"The paper's central claim is that a Denoising Diffusion Gamma Model (DDGM), in which the forward noising process adds Gamma-distributed random variables rather than Gaussian ones, is a better statistical match to OCT speckle than standard DDPMs, and that this match translates into measurable denoising gains. The second claim is that the reverse diffusion should be steered not by the noisy input but by a non-local-means-filtered version of it, through the Noise-Reduced Fidelity Term; this prevents high-frequency noise from being reinforced. On a prospectively collected paired dataset, GARD reports the highest PSNR (28.25 dB), SSIM (0.58), and lowest MSE (103.95) among all compared methods, with Wilcoxon signed-rank tests marking every difference significant at p<0.01, and qualitative inspection showing crisper retinal-layer edges and better-preserved small reflective structures. The ablation results support the mechanism: the NRFT improves the Gamma DDGM but slightly degrades a Gaussian DDPM, which the authors read as evidence that the Gamma noise model and the noise-reduced guide work together.","pith_inferences":["If the Gamma-statistics premise transfers across devices, the $\\theta_0$ hyperparameter and noise schedule would need per-device calibration; the paper's cross-vendor qualitative results hint at generalizability but do not quantify it, so a device-specific tuning study is the natural next test.","Because the NLM guide is a fixed preprocessing step, a learned or adaptively weighted guide could plausibly improve the fidelity term further; this is an extension the paper does not explore.","The same additive-Gamma-after-compression argument should apply to other coherent imaging modalities, such as ultrasound or other OCT systems, so a direct transfer test on log-compressed ultrasound speckle would reveal how much of the gain is generic rather than retinal-specific."],"forward_implications":["GARD achieves the best PSNR, SSIM, and MSE on the paired noisy/less-noisy OCT dataset, with all differences significant at p<0.01, outperforming classical NLM and deep-learning baselines such as SCUNet, N2V2, and Speckle2Speckle.","The Noise-Reduced Fidelity Term improves the Gamma DDGM but slightly hurts a Gaussian DDPM, implying the noise-reduced guide is most effective when the diffusion model's noise statistics already match speckle.","A fidelity term that forces consistency with the original noisy image (CPDM) lowers performance, supporting the paper's argument that guiding with a less-noisy reference, not the noisy input, is the right design.","Because inference starts at t=70 and samples every 10th timestep via the Gamma-adapted DDIM, the method is substantially faster than a full 1000-step reverse diffusion, making clinical use more plausible.","Higher-quality images after denoising could enable faster acquisitions and cheaper OCT devices while preserving diagnostic detail, the application-level motivation."],"supporting_citations":[{"why":"Supplies the Denoising Diffusion Gamma Model: the forward process in Eq. (1), the reverse process in Eq. (2), and the Gamma additivity property used for fast sampling.","marker":"[20]"},{"why":"Provides the Denoising Diffusion Implicit Model acceleration that GARD adapts by setting $\\sigma_t = 0$ for deterministic sampling and skipped timesteps.","marker":"[27]"},{"why":"Defines non-local means filtering, used to build the less-noisy guide image $\\tilde{y}$ for the Noise-Reduced Fidelity Term.","marker":"[2]"},{"why":"Introduces the content-preserving fidelity term used as a baseline and as the comparison point showing why fidelity to the noisy input is harmful.","marker":"[17]"},{"why":"Brings regularized reverse diffusion for medical image denoising to the comparison and supplies the DDPM baseline and fidelity-term idea.","marker":"[5]"},{"why":"State-of-the-art blind denoising baseline (SCUNet) that GARD outperforms on the paired dataset.","marker":"[30]"},{"why":"Self-supervised Noise2Void2 baseline representing the specialized training-without-clean-targets approach.","marker":"[11]"},{"why":"Speckle2Speckle baseline for unsupervised speckle filtering, used as a task-specific comparison.","marker":"[7]"}],"fun_headline_variants":["Gamma diffusion model sharpens OCT images beyond SOTA","Despeckle OCT with Gamma diffusion and noise-reduced guidance","GARD: Gamma diffusion beats OCT denoisers in fidelity","Diffusion model with Gamma noise restores finer retinal details"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that speckle in the display-ready, post-processed OCT images fed to the model is approximately additive and Gamma-distributed, so the diffusion forward and reverse processes match the real noise; if that statistical model is wrong for a given image domain, the learned reverse diffusion would not remove real speckle and the reported gains would rest on the NLM guide alone.","fun_headline_variants_meta":{"raw":{"variants":["Gamma diffusion model sharpens OCT images beyond SOTA","Despeckle OCT with Gamma diffusion and noise-reduced guidance","GARD: Gamma diffusion beats OCT denoisers in fidelity","Diffusion model with Gamma noise restores finer retinal details"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1544,"prompt_tokens":1009,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":466}},"tokens_in":625,"tokens_out":535,"duration_ms":4614,"temperature":1.0,"reasoning_tokens":466,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:54:26.412284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to repeat the paired noisy/less-noisy evaluation on raw, linear-scale OCT B-scans before log or fourth-root compression, where speckle is known to be multiplicative. If GARD's advantage over the Gaussian DDPM shrinks or vanishes there, the additive-Gamma model is doing the load-bearing work; if the advantage persists, the NRFT guide, not the Gamma mechanism, drives the gains.","supporting_citations":[{"cited_title":"In: MICCAI 2023","cited_arxiv_id":null,"evidence_quote":"Introduces the content-preserving fidelity term used as a baseline and as the comparison point showing why fidelity to the noisy input is harmful."},{"cited_title":"Machine Intelligence Research20(6), 822–836 (Dec 2023)","cited_arxiv_id":null,"evidence_quote":"State-of-the-art blind denoising baseline (SCUNet) that GARD outperforms on the paired dataset."},{"cited_title":"In: Computer Vision – ECCV 2022 Workshops","cited_arxiv_id":null,"evidence_quote":"Self-supervised Noise2Void2 baseline representing the specialized training-without-clean-targets approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Speckle2Speckle baseline for unsupervised speckle filtering, used as a task-specific comparison."}],"review_version":1}