{"id":"6e4ed7a6-0e22-46b5-982c-413cc6e7d703","arxiv_id":"2507.19035","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-path CNN that learns noise and context separately and fuses them gives modest PSNR gains over U-Net for Gaussian noise on CT, MRI, OCT, and fundus images, but not over RED-CNN for AWGN or speckle.","lead":"This paper introduces a dual-path neural network that separates noise estimation from context reconstruction and fuses the two to denoise medical images. It reports small PSNR gains over a U-Net baseline on synthetic Gaussian noise across four modalities, though the CT validation split leaks patient identity and the motivating CNCL method is never compared.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3.35% headline rests on Table IV PSNR values that cannot be reconciled with the stated noise parameters, and no patient-disjoint validation is reported; the released code must reproduce the table before the central claim can be trusted.","rationale":"I read the abstract, the noise-simulation description in Section III-B3, the PSNR definition in Section III-A2, the data-splitting description in Section III-B2, and Tables III and IV. The strongest claim is a single percentage on a pooled Gaussian PSNR comparison. For that claim to be true, the reported PSNR values must be actual PSNRs of denoised validation images under the stated noise parameters. That condition is the least secure part of the paper: the numbers have no noisy-input reference, are inconsistent with either common normalization of the image range, and the CT/MRI validation splits are not patient-disjoint. The reader identified the CT split leakage as the weakest assumption; that is a real and plausible cause of inflated PSNR, but the deeper problem is that the quantitative evaluation itself is not reproducible from the text. The AWGN and speckle rows in Table IV further suggest a systematic metric or normalization issue rather than a narrow data-split problem. Because the central claim is not supported by internally consistent evidence as written, I would move the verdict from CONDITIONAL to REJECT pending the concrete rerun: if the code reproduces the table with corrected splits and a clear noisy-input baseline, the claim can be revisited; if not, the headline result is unsubstantiated.","tokens_in":11162,"tokens_out":11646,"duration_ms":126587,"concrete_test":"Using the released code (Zenodo 10.5281/zenodo.15836053), regenerate a small Gaussian-noise validation set with the exact Table III parameters and compute PSNR of (i) the noisy input, (ii) the baseline U-Net output, and (iii) the DPL output, using the paper's own PSNR routine and recording whether images are normalized to [0,1] and whether the maximum intensity is set to 1 or 255. If the noisy-input PSNR is about 71 dB (8-bit) or about 23 dB ([0,1]) while the outputs are 47-49 dB, or if the AWGN variance contradicts the ordering in Table IV, rerun Table IV with patient-disjoint CT and MRI splits and report the pooled Gaussian comparison; the 3.35% claim stands only if DPL still outperforms U-Net under that corrected evaluation.","verdict_should_be":"REJECT","load_bearing_attack":"The central quantitative claim (Abstract; Table IV 'All' under Gaussian noise: baseline 47.65 dB, DPL 49.25 dB, a 3.35% gain) depends on a PSNR evaluation that is never tied to the image range used for noise simulation. Section III-B3 adds Gaussian noise with variance 0.005; Section III-A2 defines PSNR with 'maximumIntensity (typically 255)'. If pixel values are in [0,255], the noisy input itself has PSNR 71.1 dB, so a trained denoiser should not fall to 47.65 dB. If images are instead normalized to [0,1] before noise, the noisy input has PSNR 23.0 dB and the claimed outputs imply an output MSE roughly 250 times smaller than the injected noise variance, which is not achievable by ordinary denoising without oracle-level information or memorization. Table IV also shows AWGN with a stated scale of 0.0001, i.e. much smaller variance, yet lower PSNR than the Gaussian case, which is again internally inconsistent. No PSNR of the noisy input, no error bars, and no significance tests are reported. Additionally, Section III-B2 describes CT and MRI validation splits that overlap in patient identity with training (30 CT scans per each of 5 patients; one MRI scan per patient while other scans of the same patients are in training). Together these issues mean the reported 47-49 dB values are not credible as standard PSNR measurements, and the 3.35% gain may be an artifact of evaluation leakage or a nonstandard metric rather than of the dual-path architecture.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dual Path Learning (DPL), a two-stage CNN for medical image denoising in which a noise-estimator U-Net and a context-estimator U-Net operate in parallel and their outputs are concatenated and fused by a third U-Net. The model is trained with the sum of three MSE losses and evaluated on synthetic Gaussian, AWGN, and speckle noise across CT, MRI, OCT, and fundus images, both per modality and pooled. The headline claim is a 3.35% PSNR improvement over a baseline U-Net under Gaussian noise when all modalities are pooled. Code is released via Zenodo.","tokens_in":11516,"tokens_out":9206,"duration_ms":90312,"significance":"If the reported results were correct, the paper would provide a useful modular, modality-agnostic denoising architecture and one of the broader multi-modality, multi-noise evaluations in this area. The architecture is clearly specified, the code is publicly available, and the pooled multi-modality training setup is a worthwhile contribution. However, the quantitative support is currently undermined by an apparent nonstandard PSNR definition, validation splits that are not patient-disjoint, and the absence of error bars or significance tests. These issues directly affect the paper's central claim, so the significance is conditional on a corrected and rerun evaluation.","major_comments":[{"comment":"Eq. (2) defines PSNR as 10*log10(maximumIntensity/MSE) without squaring the peak intensity, and Section III-B3 adds Gaussian noise with variance 0.005 without stating the pixel range used for the PSNR computation. With images normalized to [0,1] and maximumIntensity=255, the noisy input would have PSNR 10*log10(255/0.005) ≈ 47.1 dB under this formula, essentially equal to the baseline U-Net's 47.65 dB in Table IV, while the conventional PSNR with peak value 1 would be roughly 24 dB lower. Because the Abstract's 3.35% gain is computed from these Table IV values, the authors must state the intensity range, correct Eq. (2) to 10*log10(MAX^2/MSE), and recompute all reported PSNRs and percentages; as written, the absolute values are not comparable to standard PSNR results in the literature.","section":"III-A2, Eq. (2), and Table IV"},{"comment":"The CT validation set is formed by randomly selecting 30 scans from each of the same five patients whose remaining scans are used for training, and the MRI validation similarly selects one scan per patient from the same 146 patients represented in training. This is not a patient-disjoint split, so the reported gains could be inflated by the models' ability to memorize patient-specific anatomy and acquisition characteristics. Please re-run at least the CT and MRI experiments with patient-disjoint splits, or explicitly relabel the results as slice-level and provide a separate patient-disjoint analysis to support the generalization claim.","section":"III-B2"},{"comment":"No error bars, standard deviations over validation slices, confidence intervals, or significance tests are reported for any PSNR or SSIM entry. Figure 6 gives only aggregate means and standard deviations, and Figure 7 shows percentage improvements without a paired statistical test; therefore differences such as the AWGN 'All' row (DPL 33.87 vs RED-CNN 35.99) cannot be distinguished from random variation. Please report per-slice distributions and paired significance tests for the main comparisons.","section":"IV, Table IV"},{"comment":"The noise simulation parameters in Table III are ambiguous ('Covar:loc 0.01' and 'Covar:scale 0.0001' for AWGN; no specification of whether images are in [0,1] or [0,255]), and all noise is synthetic. The paper's conclusion of robustness and generalizability to medical imaging is therefore not supported for real clinical acquisition noise. Additionally, in the pooled 'All' columns of Table IV, DPL is below RED-CNN for both AWGN (33.87 vs 35.99 dB) and speckle (37.12 vs 40.89 dB), so the Abstract's wording 'demonstrating its robustness and generalizability' should be narrowed to the Gaussian-noise setting and qualified as a synthetic-noise evaluation.","section":"III-B3 and IV"},{"comment":"Since DPL is presented as an extension of CNCL [14], the experiments should include CNCL as a baseline. Table IV compares against RED-CNN and classical methods but not against the approach that motivates the dual-path design, leaving the incremental contribution unquantified. Please add this comparison or justify its omission and adjust the corresponding claims.","section":"II and IV"}],"minor_comments":[{"comment":"The sentence 'Section 2 discusses the related work, and Section 2 details the data preparation and model' should refer to distinct sections (II and III).","section":"Section I"},{"comment":"The RED-CNN row for the Gaussian condition contains '49.2837.7744.93', which appears to be missing spaces between the MRI, OCT, and FUNDUS entries.","section":"Table IV"},{"comment":"The caption contains the typo 'Dundus'; it should be 'FUNDUS'.","section":"Figure 5"},{"comment":"The reference to 'Genf [14]' should be 'Geng [14]', and the author list contains 'Pedro Friere' while the affiliation block gives 'Pedro Freire'; please standardize the spelling.","section":"References and affiliations"},{"comment":"DPL is trained with batch size 32 while the baseline U-Net and RED-CNN are trained with batch size 64; please clarify whether this difference affects the comparison or justify the choice.","section":"Section III-E"},{"comment":"The limitations paragraph acknowledges the small OCT dataset and the 10,000-epoch training budget; these acknowledged limitations should be reflected more explicitly in the conclusion and in the Abstract's generality claims.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The apparent PSNR-formula issue is reproducible from the manuscript alone: Eq. (2) omits the square, and Table IV values align with the noisy-input PSNR computed with that formula. I recommend asking the authors to provide a reproducibility appendix that reruns the metrics with a standard PSNR definition and patient-disjoint splits. If the Zenodo code reproduces Table IV as printed, the metric issue will be immediately visible; if it does not, that is an additional serious concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper with the stress-test note in hand. The short version: the dual-path idea is a sensible extension of CNCL, and the paper is honest about its limitations, but the reported PSNR values don't survive contact with the noise parameters in Table III. That makes the headline 3.35% gain unsupported.\n\nHere's the concrete problem. Gaussian noise is listed with variance 0.005. PSNR is defined with a maximum intensity of 255. If the images are in [0,255], the noisy input already has a PSNR around 71 dB, so a trained denoiser that outputs 49 dB is doing harm, not denoising. If the images are normalized to [0,1] before noise, the noisy input has a PSNR around 23 dB, and a 49 dB output means the output MSE is roughly 400 times smaller than the injected noise variance. That is not plausible for generic denoising without oracle-level information or memorization. The AWGN results are even stranger: a scale of 0.0001 should be nearly invisible, yet the table reports lower PSNR than the Gaussian case. No noisy-input PSNR, error bars, or significance tests are reported, so the table is not anchored to a standard measurement.\n\nWhat the paper does well: the method is clearly specified, the code is released, and the multi-modality, multi-noise benchmark is a useful contribution even if the numbers shift. The authors state their limitations plainly, noting DPL underperforms on OCT and does not beat RED-CNN on AWGN or speckle. The patient-leakage issue in the CT and MRI splits is real but secondary; if the code reproduces the table, that is fixable with a cleaner split.\n\nMy recommendation: do not reject on the architecture. Send to peer review with a hard requirement to reproduce Table IV from the released code, clarify the PSNR definition, and rerun with patient-disjoint validation. If the numbers don't reproduce, that is a rejection. As it stands, the central quantitative claim is not credible, but the idea and the dataset setup deserve a second look.","headline":"The dual-path architecture is a reasonable CNCL variant, but the PSNR table contradicts the paper's own noise parameters, so the 3.35% claim is not credible until the code reproduces it.","tokens_in":12045,"tokens_out":3394,"would_cite":false,"duration_ms":33310,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Denoising medical images works better when a network learns noise and context in two separate paths, then fuses them.","keywords":["medical image denoising","dual-path learning","convolutional neural network","U-Net","PSNR","Gaussian noise","multimodal imaging","CT MRI OCT fundus"],"falsifier":"Retrain DPL and the baseline U-Net on CT with a patient-disjoint split, for instance training on four of the five patients and validating on the fifth, then average over all five choices and compare pooled Gaussian PSNR. If DPL's 3.35% advantage over the baseline disappears or falls below the per-image noise variance, the stated cross-modality gain is an artifact of the split rather than a property of the architecture.","tokens_in":10986,"feed_emoji":"🩻","tokens_out":4648,"duration_ms":45245,"temperature":0.7,"pith_summary":"This paper tries to establish that medical image denoising improves when a network learns noise and context separately and then fuses them, rather than learning a single direct noisy-to-clean mapping. It reports that this dual-path design beats a standard U-Net by 3.35% in PSNR on Gaussian noise when trained on pooled CT, MRI, OCT, and fundus images, and also beats RED-CNN and BM3D under that noise. If true, this would mean one unified model can serve multiple modalities and noise types, reducing the need for separate denoisers per modality.","feed_headline":"Dual-path CNN beats U-Net at medical image denoising","feed_subtitle":"Two learned paths, one for noise and one for context, lift Gaussian PSNR by 3.35% across four modalities.","key_machinery":"The carrying mechanism is a two-phase dual-path U-Net: a noise estimator produces a noise map $\\eta$ whose subtraction from the input yields $X''$, a context estimator recovers the clean context $X'$ directly, and a fusion reconstructor U-Net concatenates $X'$ and $X''$ to produce the final denoised image. The three sub-networks are trained jointly with three MSE losses, one for each path and one for the fusion output, and the paper argues this separation of noise and context lets the fusion step combine complementary features that a single direct mapping would blur together.","core_discovery":"The paper's central claim is that DPL, a Dual-Path Learning architecture, outperforms a single-path U-Net on medical image denoising. Two parallel U-Nets estimate the noise map $\\eta$ and the clean context $X'$; subtracting the noise estimate gives an intermediate denoised image $X'' = X - \\eta$, and a third U-Net fuses $X'$ and $X''$ into the final output. On pooled Gaussian-noise data, this design reaches 49.25 dB PSNR versus 47.65 dB for the baseline U-Net, a 3.35% improvement, and also exceeds RED-CNN and BM3D under Gaussian noise. Under AWGN and speckle noise, DPL remains competitive with, but does not surpass, RED-CNN.","pith_inferences":["Inference: the reported advantage is measured on synthetically corrupted images; real scanner noise is structured and signal-dependent, so the gap over the baseline could shrink when tested on genuine low-dose CT or OCT acquisitions.","Inference: because the CT validation set draws slices from the same five patients used for training, a strict patient-disjoint evaluation is needed before the 3.35% pooled gain can be taken as evidence of generalizable denoising rather than patient memorization.","Inference: the same noise/context separation could be applied to other restoration tasks such as deblurring, super-resolution, or artifact removal, where the 'noise' path is replaced by the specific degradation being estimated."],"forward_implications":["A single DPL model trained on combined CT, MRI, OCT, and fundus data can denoise all four modalities under Gaussian noise, reducing the need for per-modality denoisers.","Replacing the U-Net blocks in DPL with other autoencoder designs should be straightforward, making the dual-path scheme a flexible template rather than a fixed network.","Under AWGN and speckle noise, DPL stays competitive with RED-CNN while outperforming the baseline and most classical filters, suggesting the architecture generalizes across noise types even where it does not lead.","DPL's gains come at a higher training cost: about 46.7 hours on CT versus 16 hours for the baseline U-Net and 17 hours for RED-CNN.","The paper reports consistent improvement over the baseline on roughly 80% of individual Gaussian-noised images, not just on the pooled mean."],"supporting_citations":[{"why":"Supplies the content-noise complementary learning idea that DPL extends into a CNN-based fusion framework.","marker":"[14]"},{"why":"Provides the U-Net backbone used for the noise estimator, context estimator, and fusion reconstructor.","marker":"[7]"},{"why":"RED-CNN is the deep-learning baseline against which DPL reports its Gaussian-noise gains.","marker":"[8]"},{"why":"Source of the 650 CT slices from five patients used in training and validation.","marker":"[11]"},{"why":"Source of the brain MRI scans across 146 patients used for the MRI portion of the study.","marker":"[19]"},{"why":"Source of the fundus images used to test cross-modality generalization.","marker":"[4]"},{"why":"One of the OCT datasets used in the small OCT evaluation.","marker":"[15]"},{"why":"The other OCT dataset used in the small OCT evaluation.","marker":"[6]"}],"fun_headline_variants":["Dual-path denoising: noise map + context fusion lifts PSNR by 3.35%","Two U-Nets beat one on Gaussian medical image denoising","Fuse noise and context in dual paths for 3.35% PSNR gain over U-Net","DPL: dual-path learning improves medical denoising by 3.35% on Gaussian","Noise and context paths: better medical denoising than single U-Net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The CT validation slices are drawn from the same five patients as the training slices, so the reported PSNR gains may partly reflect the model memorizing patient-specific anatomy instead of learning general denoising.","fun_headline_variants_meta":{"raw":{"variants":["Dual-path denoising: noise map + context fusion lifts PSNR by 3.35%","Two U-Nets beat one on Gaussian medical image denoising","Fuse noise and context in dual paths for 3.35% PSNR gain over U-Net","DPL: dual-path learning improves medical denoising by 3.35% on Gaussian","Noise and context paths: better medical denoising than single U-Net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2861,"prompt_tokens":895,"completion_tokens":1966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":1848}},"tokens_in":511,"tokens_out":1966,"duration_ms":16135,"temperature":1.0,"reasoning_tokens":1848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:02:49.951982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DPL and the baseline U-Net on CT with a patient-disjoint split, for instance training on four of the five patients and validating on the fifth, then average over all five choices and compare pooled Gaussian PSNR. If DPL's 3.35% advantage over the baseline disappears or falls below the per-image noise variance, the stated cross-modality gain is an artifact of the split rather than a property of the architecture.","supporting_citations":[{"cited_title":"Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN)","cited_arxiv_id":"1702.00288","evidence_quote":"RED-CNN is the deep-learning baseline against which DPL reports its Gaussian-noise gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the 650 CT slices from five patients used in training and validation."},{"cited_title":"Triplet Cross-Fusion Learning for Unpaired Image Denoising in Optical Coherence To- mography","cited_arxiv_id":null,"evidence_quote":"One of the OCT datasets used in the small OCT evaluation."},{"cited_title":"Sparsity based denoising of spectral domain optical coherence tomography images","cited_arxiv_id":null,"evidence_quote":"The other OCT dataset used in the small OCT evaluation."}],"review_version":2}