{"id":"f28809b2-5964-441d-8031-73bb68acf2cd","arxiv_id":"2507.06828","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised ultrasound despeckling method that creates multi-scale perturbed views and enforces cross-scale consistency to separate low-rank anatomy from speckle.","lead":"Speckle2Self is a self-supervised method that removes ultrasound speckle from single noisy B-mode images, without needing clean images or paired scans. It perturbs the input at multiple scales, trains a network to find anatomy shared across scales, and reports improved structural preservation on simulated and real carotid ultrasound images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MSP is assumed to perturb only speckle, but downsampling also low-pass filters the tissue signal; the consistency loss may therefore reward smoothing, so the reported structure metrics may partly measure blur rather than true speckle disentanglement.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: MSP is assumed to leave the clean structure unchanged, but downsampling-upsampling necessarily low-pass filters tissue information, so the cross-scale consistency loss may simply enforce smoothness. This is the single most consequential threat to the central claim because the entire training signal derives from agreement among the MSP views. The reader's verdict of CONDITIONAL already reflects this uncertainty, and the proposed concrete test (comparing Speckle2Self against simple low-pass baselines on synthetic data with edge-preservation metrics) would either substantiate or retire the concern. I therefore agree with the reader and do not propose a verdict change; the conditional status should remain until the smoothing-versus-speckle distinction is tested directly. Independent supports in the paper—the physics-based MUST simulator, cross-configuration tests, and the downstream segmentation experiment—are real evidence, but they do not remove the need to verify that the learned output is not primarily a low-pass filtered version of the input.","tokens_in":23928,"tokens_out":2073,"duration_ms":27242,"concrete_test":"On the S-III synthetic test set (which has clean ground truth), compute PSNR, SSIM, LPIPS, GLCM homogeneity, and an edge-preservation metric such as Pratt's Figure of Merit for: (a) Speckle2Self output, (b) a Gaussian low-pass filter with σ tuned to match Speckle2Self's homogeneity score, and (c) the most aggressive MSP variant P_{0.25}(Y) (bilinear downsample to 1/4 and upsample back). If either (b) or (c) matches Speckle2Self within a small margin on SSIM/LPIPS/homogeneity while having substantially lower edge preservation, then the reported gains are consistent with low-pass smoothing, and Eq. (7) must be revised to account for the scale-dependent loss of tissue structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is Eq. (7), P_k(Y)=X+N^(k): MSP is claimed to change only the speckle component while leaving the clean anatomical structure X untouched. This is not what downsampling-upsampling does. Rescaling by 0.25 and interpolating back removes high-frequency tissue content, so the component shared across the three scales is a low-pass filtered version of X, not X itself. The training objective (Eq. 11) combines an MSE reconstruction loss to each perturbed input with an L1 consistency loss among the three outputs; a network can minimize both by emitting a smoothed image that sits near all three filtered views, especially because MSE concentrates on low-frequency fidelity. The Discussion concedes 'slight blurring along tissue boundaries,' and the real-data metric used (GLCM homogeneity) explicitly rewards smoothness. On synthetic data, Speckle2Self's PSNR is not the best among self-supervised methods (Table 4), while its SSIM, homogeneity, and LPIPS gains are exactly the metrics most improved by low-pass filtering. Thus the evidence does not yet establish the abstract's claim that MSP 'preserves the shared anatomical structure' in a way that isolates speckle, rather than imposing a smoothness prior. The core premise of the method is therefore not secure, independent of the missing baselines and statistical rigor noted by the reader.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Speckle2Self, a self-supervised method for ultrasound speckle reduction that requires only single noisy B-mode images. The method applies a multi-scale perturbation (MSP) operation, implemented as downsampling-upsampling at scales {1.0, 0.5, 0.25}, to create three variants of the input; it assumes each variant can be written as X + N^(k) with shared clean structure X and scale-specific speckle N^(k), and trains three encoders plus a shared decoder with an MSE reconstruction loss and an L1 cross-scale consistency loss, motivated by low-rank/sparse decomposition. The paper evaluates the method on three synthetic MUST datasets, in-vivo carotid images from a Clarius scanner, cross-configuration and cross-device generalization, a loss/interpolation/scale ablation, and a downstream carotid segmentation experiment.","tokens_in":24221,"tokens_out":6956,"duration_ms":71981,"significance":"The proposed setting is practically important: ultrasound speckle is tissue-dependent and deterministic, so N2N-style paired observations and blind-spot assumptions are often inapplicable, and a method that works from single images would fill a real gap. The paper's strengths are its extensive comparison set, the use of a realistic simulator, the cross-device generalization attempt, the ablation study, and the reported 7 ms inference time. The central weakness is that the MSP assumption has not been separated from low-pass smoothing: the metrics on which Speckle2Self wins are substantially the ones improved by blur, and the in-vivo evidence is a single smoothing-favoring metric. If the authors can provide evidence that the method preserves high-frequency tissue structure while suppressing speckle, this would be a valuable contribution to the field.","major_comments":[{"comment":"The decomposition P_k(Y) = X + N^(k) is the load-bearing assumption, but downsampling-upsampling does not leave X unchanged: rescaling by 0.25 and interpolating back removes high-frequency tissue signal, so the component shared across scales is a low-pass filtered version of X, not X itself. With the loss in Eq. (11), the network can minimize both the MSE reconstruction term and the L1 consistency term by emitting a smoothed image that sits near all three filtered views; MSE is dominated by low-frequency fidelity. The Discussion's admission of 'slight blurring along tissue boundaries' is therefore not a cosmetic limitation but direct evidence for this mechanism, and the pattern in Table 4 (PSNR not best on S-I/S-II while SSIM/LPIPS/homogeneity improve) is consistent with low-pass filtering. Please provide a test that disentangles speckle suppression from smoothing, e.g., quantify P_k(X) - X on the synthetic clean targets, compare with a model trained on LPF perturbations of matched variation, or report a high-frequency/edge-preservation metric.","section":"Section 3.3, Eq. (7)"},{"comment":"The in-vivo evaluation rests on a single no-reference metric, GLCM homogeneity, which explicitly rewards local smoothness; no error bars, confidence intervals, or paired statistical tests are reported for the 104 test images. Given the smoothing concern above, this single metric cannot support the statement that Speckle2Self 'consistently outperforms' other methods on real data. Please report per-image distributions or mean +/- std, add a structure-sensitive no-reference metric (e.g., edge/gradient preservation), and ideally a blinded clinician preference study.","section":"Section 4.4.4, Table 4 (Carotid column)"},{"comment":"The baseline set omits the most relevant self-supervised ultrasound methods, including the deep ultrasound denoising approach already cited as Goudarzi and Rivaz (2023) and the AP-BSN/C-BSN variants (Lee et al. 2022b; Jang et al. 2023) designed for spatially correlated noise; the plane-wave-based N2N variants cited in the Introduction are also not compared. The cross-device evaluation in Section 4.6.2 is qualitative only, with no metric reported on the Siemens and Cephasonics frames before or after fine-tuning. Without these quantitative comparisons, the claims of consistent superiority over self-supervised methods and strong cross-device generalization are not fully supported.","section":"Sections 4.4.1 and 4.6.2"},{"comment":"The SVD singular-value decay analysis shows that Speckle2Self outputs have faster singular-value decay, but any low-pass filter produces this effect; without a control for detail preservation, the experiment does not demonstrate that the method preserves tissue structure while removing speckle. On the synthetic data where ground truth exists, please complement this analysis with a quantitative measure of how much of the clean image's high-frequency structure is retained.","section":"Section 4.3"}],"minor_comments":[{"comment":"The heading 'Nosie2Void' contains a typo and should read 'Noise2Void'.","section":"Section 2.4"},{"comment":"The heading 'Discusssion' contains a typo and should read 'Discussion'.","section":"Section 5"},{"comment":"The text refers to the 'Folld fill algorithm'; this should be 'Flood Fill algorithm'.","section":"Section 4.7"},{"comment":"In the sentence describing Figure 9, 'the proposed MPS' should be 'the proposed MSP'.","section":"Section 4.5.2"},{"comment":"The sentence 'This denied result will enhance image consistency and thereby boost the development of reproducible computer-aided diagnosis' is malformed and unclear; it should be rewritten.","section":"Section 6"},{"comment":"The reference for N2S(*) is given as 'Krull et al. (2019)', but Noise2Self should cite Batson and Royer (2019).","section":"Section 4.4.1"},{"comment":"All synthetic and carotid metrics are reported as point estimates without standard deviations or confidence intervals, so the statistical significance of the performance gaps cannot be assessed.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The main threat is the smoothing confound around Eq. (7); I do not see circularity or integrity problems. The manuscript fits the journal's scope and is a plausible contribution if the authors can provide evidence that the method preserves high-frequency tissue structure rather than simply low-pass filtering. It would be useful to ask the authors to release code and checkpoints, and to include the ultrasound-specific self-supervised baselines they already cite."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this one is worth a look if you track self-supervised denoising for medical imaging. The core idea: apply downsampling-upsampling at three scales to a single noisy B-mode ultrasound image, train a multi-encoder/shared-decoder network to reconstruct each perturbed input, and enforce L1 consistency among outputs. The claim is that the shared component across scales is the clean anatomy while speckle changes, so the network learns to suppress speckle without clean data. What's actually new is the domain-specific adaptation of downsampling-based self-supervision to deterministic, tissue-dependent ultrasound speckle, plus the three-encoder architecture and a solid set of experiments on synthetic and real carotid images.\n\nCredit where due: the paper is transparent about its limitations—it admits the low-rank prior is an approximation and that MSP causes slight boundary blurring. The ablation study is thoughtful: it tests loss configurations, interpolation schemes, scale combinations, and compares MSP against low-pass filters as input perturbations. The cross-device generalization and downstream segmentation experiments are a plus. The synthetic results consistently beat classical filters and older self-supervised methods on SSIM, LPIPS, and homogeneity, even if PSNR is not the best.\n\nSoft spots, in order of importance. First, the most relevant baselines are missing: AP-BSN and C-BSN, which also train on downsampled image pairs, are cited but not evaluated. Without them, the novelty claim is under-supported. Second, the load-bearing assumption in Eq. (7) is not strictly true. Downsampling by 0.25 and upsampling back removes high-frequency tissue information, so the cross-scale shared structure is a low-pass filtered version of X, not X itself. The consistency loss can therefore be minimized by outputting a smoothed image. The paper acknowledges 'slight blurring along tissue boundaries,' but the real-data evaluation uses GLCM homogeneity, a metric that explicitly rewards smoothness, and there are no error bars or significance tests. On synthetic data, the wins are exactly on structure/perceptual metrics, not PSNR, which is consistent with a smoothing effect. This doesn't sink the paper—the method still outperforms BM3D and NLM on SSIM/LPIPS, so it's not just Gaussian blur—but the interpretation that MSP isolates 'sparse noise' needs more support. A simple low-pass baseline (e.g., Gaussian filter with tuned sigma) on the same synthetic data would clarify how much of the gain is smoothing. Third, the reported numbers are single runs; some error bars and a significance test would help.\n\nWho this is for: researchers working on self-supervised denoising for ultrasound or other structured noise. It is a solid contribution that a serious editor should send to review, with the expectation that the baselines and the smoothing analysis get addressed.","headline":"A solid self-supervised despeckling paper whose central MSP assumption is partly violated by low-pass filtering; worth reviewing, but it needs AP-BSN/C-BSN baselines and a smoothing check.","tokens_in":24731,"tokens_out":4958,"would_cite":false,"duration_ms":49825,"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":"A single noisy ultrasound image is enough to train a speckle-removal network.","keywords":["ultrasound speckle reduction","self-supervised denoising","multi-scale perturbation","low-rank prior","single-image denoising","carotid ultrasound","B-mode imaging","speckle"],"falsifier":"On a simulated phantom with a known clean image X and sharp point scatterers, run Speckle2Self and compare the output to X and to X passed through the same downsampling-upsampling at scale 0.25; if the output is closer (higher SSIM) to the smoothed version than to X, the assumption that MSP preserves structure fails and the method is doing low-pass estimation rather than true despeckling.","tokens_in":23722,"feed_emoji":"🩻","tokens_out":6751,"duration_ms":68335,"temperature":0.7,"pith_summary":"Speckle in ultrasound is deterministic and tissue-dependent, so standard self-supervised denoising that assumes independent pixel noise cannot remove it, and clean references are physically unavailable. This paper claims that a simple multi-scale perturbation of a single noisy B-mode image, downsampling and then upsampling at three scales, creates several views that share the same anatomical structure while carrying different speckle patterns. A network trained to reconstruct each view while enforcing consistency among reconstructions implicitly separates the low-rank clean tissue from the sparse speckle, without any clean data or paired observations. The authors show the result outperforms classical filters and existing self-supervised methods on simulated and human carotid ultrasound, generalizes to unseen probe frequencies and devices, and improves downstream vessel segmentation. If true, this makes speckle reduction available wherever a single B-mode frame exists, including portable and low-cost scanners.","feed_headline":"One noisy ultrasound image can train a speckle-removal network","feed_subtitle":"Cross-scale consistency separates tissue structure from speckle, with no clean reference or paired scans needed.","key_machinery":"The load-bearing mechanism is the multi-scale perturbation (MSP): controlled downsampling and upsampling of the input at scale factors 1.0, 0.5, and 0.25, which is assumed to leave the clean tissue structure X unchanged while reshuffling the speckle pattern. Around this, the framework is a multi-encoder network, one encoder per scale sharing a single decoder, trained with a reconstruction loss (MSE to each perturbed view) and a consistency loss (L1 between all pairs of reconstructed outputs). The low-rank prior on clean tissue, the assumption that anatomy lies on a low-rank subspace while speckle is sparse or high-rank, is what gives the consistency loss its meaning: the network is rewarded for outputs that are common across scales and penalized for outputs that follow scale-specific speckle.","core_discovery":"On the paper's own terms, the discovery is that ultrasound speckle can be removed from a single noisy image by treating the clean anatomy as a low-rank signal and the speckle as sparse, scale-dependent variation. The multi-scale perturbation (MSP) operation, with scale factors 1.0, 0.5, and 0.25, is claimed to preserve the shared anatomical structure X while producing distinct speckle realizations N(k) for each scale (Eq. 7). Three scale-specific encoders and a shared decoder are trained with an MSE reconstruction loss toward each perturbed input plus an L1 consistency loss between all pairs of reconstructions. The consistency loss forces the network to find what is invariant across scales, and the authors report that MSE reconstruction plus L1 consistency is the only loss combination that removes speckle, while symmetric combinations blur structure or retain speckle. The outcome is a despeckled image that, in their experiments, beats established filters (SRAD, NLM, BM3D, OBNLM) and prior self-supervised methods (N2N, N2V, Noise2Self, Neighbor2Neighbor, DIP, ZS-N2N) on SSIM, LPIPS, and homogeneity, with only a supervised method trained on clean targets scoring higher.","pith_inferences":["If cross-scale agreement is the right training signal, the same perturbation idea could be tested on other deterministic, spatially correlated image degradations, such as CT metal artifacts or MRI ghosting, where independent-noise self-supervision fails.","The reported boundary blurring suggests the method may be estimating a low-pass-filtered version of the anatomy rather than the true X; a direct test would compare denoised output against both clean X and its downsampled-upsampled version on synthetic data.","Because speckle is itself a deterministic signal used in elastography and motion tracking, the network's ability to separate speckle from tissue could be repurposed to extract speckle maps, not only to discard them.","A cheap diagnostic for the mechanism is replacing the three-scale perturbation with two scales plus stronger perturbation to see whether performance tracks the degree of speckle decorrelation."],"forward_implications":["Speckle reduction no longer requires clean images, paired noisy observations, or plane-wave hardware; one B-mode frame suffices for training, with about 7 ms per image at inference.","Because no tissue-specific noise model is assumed, the same trained network transfers to unseen central frequencies (3.75 to 9.38 MHz) and to images from two unseen commercial scanners, with zero-shot use and small fine-tuning gains.","Reducing speckle improves downstream analysis: on carotid artery segmentation with a rule-based flood-fill segmenter, AUC rose from 0.589 to 0.816 and IoU from 0.175 to 0.497 after Speckle2Self denoising.","The method is architecture-light (about three million parameters) and operates on standard B-mode output, so it can be deployed on portable and point-of-care systems where clean data do not exist."],"supporting_citations":[{"why":"Supplies the additive-noise approximation of B-mode speckle after log compression (Eq. 2) that the method's decomposition Y = X + N relies on.","marker":"Coupé et al. (2009)"},{"why":"Provides the robust PCA low-rank-plus-sparse decomposition that motivates modeling clean tissue as low-rank and speckle as sparse.","marker":"Wright et al. (2009)"},{"why":"Applies low-rank matrix approximation to ultrasound speckle reduction, the prior that Speckle2Self extends to a learned, self-supervised setting.","marker":"Zhu et al. (2017)"},{"why":"Supports treating clean tissue as low-rank and speckle as sparse in ultrasound despeckling.","marker":"Sagheer and George (2017)"},{"why":"Defines the Noise2Noise paradigm whose paired-independent-observation requirement is infeasible for deterministic speckle; the paper's motivation builds on this limitation.","marker":"Lehtinen (2018)"},{"why":"Defines blind-spot Noise2Void training whose spatial-independence assumption fails for clustered, tissue-dependent speckle; Speckle2Self is positioned against it.","marker":"Krull et al. (2019)"},{"why":"Provides the simulator used to generate the synthetic ultrasonic images with known clean ground truth for evaluation.","marker":"Garcia (2022)"},{"why":"Provides the GLCM homogeneity metric used to evaluate real carotid images where no clean ground truth exists.","marker":"Haralick et al. (1973)"}],"fun_headline_variants":["Self-supervised speckle removal from a single noisy ultrasound image","One noisy ultrasound scan is enough to train speckle removal","Ultrasound despeckling without clean data, from one noisy image","Self-supervised speckle reduction using only one noisy ultrasound","No clean ultrasound needed: train despeckling on a single noisy frame"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training signal rests on the premise that downsampling and upsampling changes only the speckle while leaving the true anatomy untouched, but downsampling also removes fine tissue detail, so the shared structure is really a smoothed image, which likely explains the reported boundary blurring.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised speckle removal from a single noisy ultrasound image","One noisy ultrasound scan is enough to train speckle removal","Ultrasound despeckling without clean data, from one noisy image","Self-supervised speckle reduction using only one noisy ultrasound","No clean ultrasound needed: train despeckling on a single noisy frame"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000343,"raw_usage":{"total_tokens":1956,"prompt_tokens":1083,"completion_tokens":873,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":783}},"tokens_in":699,"tokens_out":873,"duration_ms":8898,"temperature":1.0,"reasoning_tokens":783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:53:08.650813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a simulated phantom with a known clean image X and sharp point scatterers, run Speckle2Self and compare the output to X and to X passed through the same downsampling-upsampling at scale 0.25; if the output is closer (higher SSIM) to the smoothed version than to X, the assumption that MSP preserves structure fails and the method is doing low-pass estimation rather than true despeckling.","supporting_citations":[],"review_version":1}