{"id":"dd373b61-4cad-4002-8696-15687982ad5d","arxiv_id":"2412.04296","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Stylizing training images to the test-domain style with structure-preserving diffusion improves segmentation scores, but the protocol uses target-domain images for style transfer.","lead":"This paper uses a one-shot, structure-preserving diffusion stylization method called OSASIS to repaint medical training images so they look like the test hospital's images, then trains a segmentation network on the repainted images. The reported polyp and skin lesion Dice scores improve, but the protocol feeds test-domain images into the style transfer stage, so the paper's central claim that the target domain is never used does not hold.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own protocol contradicts its central generalization claim: Algorithm 1 trains the style mapping on Ytest images, so target-domain data are used in the pipeline.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the paper uses test-domain images as style references during the style-transfer stage, so the claim of 'target domain absent from training' applies only to the segmentation model, not to the overall method. This is not a peripheral flaw; it directly invalidates the central contribution as framed. The manuscript itself is explicit about the protocol (Algorithm 1, Section 5.4), so the concern is grounded in the paper's own text rather than an external assumption. The reported numerical gains may be real for a one-shot style adaptation setting, but they do not establish the advertised domain-generalization result. Additional weaknesses, such as missing error bars and very small test sets, would matter more if the central claim survived, but they do not change the verdict. Since the reader already recommends REJECT for essentially this reason, no verdict adjustment is needed.","tokens_in":14765,"tokens_out":2276,"duration_ms":24541,"concrete_test":"Re-run the CVC-ClinicDB → CVC-ColonDB experiment exactly as in Section 5.4, but set yin_B to a single random image from a third non-evaluation domain (e.g., Kvasir-SEG or a held-out split of another polyp dataset) or to a CVC-ClinicDB image, so that no test-domain image is used anywhere in Algorithm 1. Keep segmentation training and evaluation unchanged and report Dice/IoU on CVC-ColonDB. If the style-transfer gains over the direct baselines vanish or reverse, the reported improvements are contingent on test-domain access during style transfer; if they persist, the generalization claim could be re-stated and verified with a third truly held-out domain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that segmentation remains robust 'even when the target domain is not included in the training set' (Contribution 3 and the Conclusion). This is only true for the segmentation model, not for the full pipeline. Algorithm 1 (line 1) explicitly selects yin_B ∈ Yi from Ytest as the style target and trains the mapping G against it; Section 5.4 confirms that 20 CVC-ColonDB test images were used as this style reference, and a similar protocol is implied for the skin lesion experiment. Consequently, G is a one-shot style adaptor fitted to the test domain, not a generalizer trained without test-domain access. Tables 1 and 2 therefore compare 'direct training' against 'training after test-style transformation,' which is a domain-adaptation setup, not the zero-target-access setup advertised. The observed ~10% Dice gains may be genuine for that setup, but they do not support the paper's headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a medical image segmentation pipeline that combines diffusion-based style transfer with a Structure-Preserving Network (SPN), following the OSASIS method, to stylize training images into a target-domain style before training a segmentation model. The authors claim this improves segmentation robustness and accuracy across domains even when the target domain is absent from the training set, and they present experiments on colonoscopy polyp segmentation and skin lesion segmentation comparing U-Net, U-Net++, and PraNet with and without style transfer. The central domain-generalization claim, however, is contradicted by the experimental protocol: Algorithm 1 explicitly uses images from Ytest to train the style mapping, and Section 5.4 confirms that 20 CVC-ColonDB test images were used as the style target. As a result, the reported comparisons are between direct training and a pipeline that has access to test-domain style information, which is a domain-adaptation setup rather than the zero-target-access domain-generalization setup advertised in the abstract and conclusion.","tokens_in":15106,"tokens_out":3623,"duration_ms":37303,"significance":"If substantiated, a plug-in stylization method that improves cross-domain segmentation without target-domain data would be practically valuable for medical imaging, especially in low-resource clinical settings. The paper also has the merit of applying a well-known method (OSASIS) to medical segmentation and evaluating multiple segmentation backbones with several metrics. However, the central claim is not supported by the evidence as presented: the protocol uses test-domain images in the style-transfer training stage, the test sets are very small (20 images), the tables report single runs without error bars, and the only baseline is direct training without comparison to existing domain generalization or domain adaptation methods. The observed improvements may be genuine for a test-time-adaptation variant, but they do not establish the paper's headline claim of domain generalization without target-domain access.","major_comments":[{"comment":"The paper's central claim that the method works 'even when the target domain is not included in the training set' (Contribution 3, Abstract, Conclusion) is contradicted by the protocol. Algorithm 1 selects a random image yin_B ∈ Yi from Ytest as Domain B and trains the style mapping G against it, and Section 5.4 confirms that 20 CVC-ColonDB test images were used as the style target. The style-transfer stage therefore has direct access to test-domain images, so the evaluation is a domain-adaptation setting, not the domain-generalization setting claimed. The observed gains in Tables 1 and 2 may be valid for that adaptive setup, but they do not support the headline claim.","section":"§4.6, Algorithm 1, line 1; §5.4"},{"comment":"The test set in the polyp experiment consists of 20 selected images from CVC-ColonDB, and the skin lesion experiment similarly uses 20 test images. Tables 1 and 2 report single-run Dice, IoU, and other metrics with no error bars, confidence intervals, or multiple-seed results. With n = 20, the reported improvements of roughly 10% in Dice and IoU cannot be distinguished from run-to-run variability, so the quantitative claims are not statistically established.","section":"§5.4, Tables 1–2"},{"comment":"The only baseline is the 'direct approach' of training the segmentation model on the original source images. There is no comparison to existing domain generalization or domain adaptation methods (e.g., DANN, CORAL, IRM, or other style-transfer-based DG approaches), nor to the OSASIS method itself. Consequently, the claim of 'superior performance metrics compared to baseline models' is limited to a comparison with no stylization and does not demonstrate an improvement over the state of the art for cross-domain medical image segmentation.","section":"§5.1, §5.4"},{"comment":"The methodology section (Equations 1–13 and the SPN description) is essentially a restatement of the OSASIS method of [16], which the paper acknowledges in Section 2.1. The paper does not clearly delineate which components are novel contributions beyond applying OSASIS to medical image segmentation and attaching a segmentation head. Since the title and introduction claim a 'novel medical image segmentation method', the novelty attribution needs to be explicit and justified.","section":"§4, §2.1"}],"minor_comments":[{"comment":"The captions of Figures 5 and 7 contain garbled text sequences such as '/uni00000027/uni0000004c/...' that appear to be rendering artifacts; these should be corrected.","section":"Figures 5 and 7"},{"comment":"The equations for DDIM forward and reverse processes are typeset incorrectly (e.g., the square root and epsilon terms are malformed), making them hard to read; they should be reformatted to match standard notation.","section":"§4.1, Equations (1)–(2)"},{"comment":"The cycle consistency loss uses F and G without defining these mappings earlier in the text; in Algorithm 1, G is the overall style mapping, but the equation's notation is unclear. Please define F and G explicitly.","section":"§4.5.1, Equation (11)"},{"comment":"The segmentation loss is defined with fθ in the equation, but the accompanying text says 'denotes the segmentation model'; also 'Segment' is not defined. Use one consistent symbol for the segmentation model.","section":"§4.5.2, Equation (14)"},{"comment":"The description 'Freeze parameters of DiffAEDA for Domain A and Set DiffAEDB for Domain B to train mode' is ambiguous; later, backpropagation updates 'the Structure-Preserving Network and mapping functions G, specifically zB_sem.' Please clarify exactly which parameters are trainable and which are frozen.","section":"§4.6, Algorithm 1, line 2"},{"comment":"The paper states that 178 images were chosen from CVC-ClinicDB; it is unclear whether this is the full training set or a subset, and if a subset, how the selection was made. Please specify the exact data split and the number of test images used for evaluation.","section":"§5.4"}],"recommendation":"reject","confidential_remarks":"The paper appears to be a direct application of OSASIS to medical segmentation, and the central domain-generalization claim is undermined by the use of Ytest images in the style-transfer training stage. The authors could potentially reframe the work as test-time adaptation, but that would require a substantial revision of the claims and a new experimental protocol. As it stands, the load-bearing issue is not a presentation problem but a mismatch between the claimed contribution and the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an application of OSASIS to medical segmentation, and the empirical gains may be real, but the paper's central claim that the target domain is absent from training is contradicted by its own protocol. Algorithm 1 takes yin_B from Ytest and trains the style mapping G against it, and Section 5.4 says 20 CVC-ColonDB images were used as the style target. So the comparison is 'train on source' vs 'train on source stylized to match test images' — that is one-shot domain adaptation, not domain generalization. The gains are not surprising under that setup; they tell us little about unseen-domain behavior.\n\nWhat is genuinely new: the paper transfers the OSASIS pipeline to polyp and skin-lesion segmentation and evaluates three backbones (UNet, UNet++, PraNet). That specific empirical result, e.g., PraNet Dice rising from 0.759 to 0.842 on CVC-ClinicDB to CVC-ColonDB, is not in the cited literature. The authors also provide code and report several metrics beyond Dice/IoU. As an engineering test of whether one-shot stylization can improve cross-domain segmentation when a test-style image is available, the idea is reasonable.\n\nThe soft spots beyond the protocol issue: no error bars or multiple runs; the test sets are tiny (20 images); the only baseline is direct training without style transfer, so we don't see how this compares to standard DG or DA methods like DANN or simple fine-tuning on the target hospital. The style-transfer hyperparameters (lambda_1..3, T1, T2, n) are not tied to any sensitivity analysis. The method itself is borrowed from OSASIS; the only new contribution is the application and the measurements.\n\nIs it worth engaging? I think yes, with the framing fixed. The core method is published elsewhere; what this paper adds is evidence that the same approach can help medical segmentation. If the authors reframe the contribution as one-shot domain adaptation with access to a single target-domain image, add variance estimates and standard baselines, and stop claiming the target is unseen, it could be a credible short paper. As submitted, the headline claim does not survive a close reading.\n\nMy recommendation: send to peer review only if the authors are willing to fix the framing and add those controls. Otherwise it is a desk-reject candidate because the central claim is unsupported.","headline":"OSASIS applied to medical segmentation, but the protocol uses test images for style transfer, so the 'target-free' claim is unsupported.","tokens_in":15474,"tokens_out":2835,"would_cite":false,"duration_ms":26064,"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":"This paper claims that adding a structure-preserving network to a one-shot diffusion stylizer keeps lesion location and size fixed, and that repainting training images in a target style raises segmentation Dice and IoU by about 10% on…","keywords":["medical image segmentation","domain generalization","style transfer","diffusion models","structure-preserving network","polyp segmentation","skin lesion segmentation","one-shot stylization"],"falsifier":"Train the style mapper on images from a third polyp dataset, keeping all CVC-ColonDB images out of both training stages, and then test on CVC-ColonDB; if the Dice and IoU gains disappear, the reported improvement came from test-domain leakage into the style-mapping stage rather than from structure-preserving stylization.","tokens_in":14579,"feed_emoji":"🩺","tokens_out":7636,"duration_ms":71612,"temperature":0.7,"pith_summary":"This paper argues that the reason style transfer has not helped medical image segmentation much is that diffusion models alter lesion structure, and it proposes a Structure-Preserving Network (SPN) to stop that. Adding SPN to a one-shot diffusion stylizer lets a training set be repainted in a target domain's style while lesion location, size, and shape stay fixed. Segmentation models trained on those stylized images then score higher Dice, IoU, and related metrics on colonoscopy polyp and skin-lesion benchmarks than the same models trained on the original images. The paper further claims this holds even when the target domain is absent from the segmentation model's training set, so the stylizer acts as a plug-in preprocessing step for any existing segmentation network.","feed_headline":"Stylized training images lift polyp segmentation Dice by ~10%","feed_subtitle":"A structure-preserving stylizer keeps lesions intact and lifts Dice, IoU, and related metrics on unseen domains.","key_machinery":"The central mechanism is the Structure-Preserving Network (SPN), a small convolutional block that applies a 1$\\times$1 convolution to the input image's latent representation and adds the result to the diffusion latent at every reverse step. The diffusion model is a pretrained DDPM sampled with DDIM, conditioned on a semantic code from a Diffusion Autoencoder (DiffAE) encoder, so style comes from the target image while content comes from the source. Training combines a CLIP directional adversarial loss, a cycle-consistency loss, and an SPN structure-preservation loss; the SPN is what carries the paper's claim that lesion location and size are invariant between original and stylized images.","core_discovery":"The paper's central claim is that structure-aware one-shot stylization, implemented as OSASIS with an extra Structure-Preserving Network, turns a diffusion model into a safe domain translator for medical images: it changes appearance and background to match a target domain without moving, resizing, or reshaping the lesion. On colonoscopy polyp segmentation, PraNet with stylized inputs raises Dice from 0.7593 to 0.8415 and IoU from 0.6520 to 0.7712 over direct training; similar gains appear for U-Net and U-Net++, and the skin-lesion experiment shows the same pattern. The authors attribute the gain to the SPN's 1$\\times$1 convolution being added to the diffusion latents at every reverse step, which preserves the input structure that plain diffusion stylization tends to erase.","pith_inferences":["Editorial inference: the experiments conflate two forms of generalization. Because Algorithm 1 uses a test-set image to train the style mapping G, the target domain is absent only from the segmentation model's training, not from the overall pipeline; a stricter test would hold out entire target datasets from both stages.","Editorial inference: the SPN is a weak structural prior, so the method's ability to preserve lesion geometry could be measured directly by computing centroid shift and area change between source and stylized images; the paper reports qualitative examples but not such a quantitative structure-preservation metric.","Editorial extension: if the method generalizes as claimed, it should transfer to other medical imaging modalities such as CT or MRI and to multi-target setups where one stylizer is trained on several target styles at once; those settings would stress-test whether SPN preserves structure independent of style.","Editorial extension: because the stylizer needs only one target image, a natural clinical deployment is to restyle the existing labeled source corpus to match a new device and then retrain, which could be evaluated on prospective data from that device."],"forward_implications":["If the central claim holds, any existing segmentation network, such as U-Net, U-Net++, or PraNet, can be improved by prepending this stylizer, with no changes to the segmentation architecture.","The stylizer needs only one target-domain image to define the style, so a hospital could adapt a model to a new device by supplying a single representative scan.","Because SPN keeps lesion geometry fixed, the stylized images can be used as training data without introducing the label misalignment that naive style transfer would cause.","The roughly 10-point Dice and IoU gains reported for PraNet on polyps imply that the stylized training corpus, not the original corpus, is what segmentation models should be trained on when a new acquisition condition is encountered."],"supporting_citations":[{"why":"Supplies the OSASIS method and the Structure-Preserving Network that the paper adapts to medical segmentation.","marker":"[16]"},{"why":"Provides the pretrained DDPM that generates and denoises images in the stylization pipeline.","marker":"[10]"},{"why":"Defines the DDIM sampling equations used for the forward and reverse diffusion steps.","marker":"[37]"},{"why":"Provides the DiffAE semantic encoder that captures content so style transfer does not erase semantics.","marker":"[38]"},{"why":"Supplies the CLIP directional loss used as the adversarial style-matching term.","marker":"[39]"},{"why":"CVC-ClinicDB is the training source for the colonoscopy polyp experiment.","marker":"[43]"},{"why":"CVC-ColonDB supplies the test set and the style-target images in the polyp experiment.","marker":"[44]"},{"why":"HAM10000 provides the skin-lesion images used for the second experiment.","marker":"[45]"}],"fun_headline_variants":["Diffusion stylization with structure preservation lifts polyp Dice by 10%","Keep lesions intact: stylized images improve polyp segmentation","Structure-aware stylization boosts segmentation Dice by 10%","One-shot style transfer that preserves structure improves segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a test-domain image may be used to train the style-transfer stage without violating the claim that the target domain is absent from training; if target-domain images must be withheld from the entire pipeline, the experiments no longer test what the paper promises.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion stylization with structure preservation lifts polyp Dice by 10%","Keep lesions intact: stylized images improve polyp segmentation","Structure-aware stylization boosts segmentation Dice by 10%","One-shot style transfer that preserves structure improves segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1466,"prompt_tokens":923,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":476}},"tokens_in":539,"tokens_out":543,"duration_ms":5426,"temperature":1.0,"reasoning_tokens":476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:32:42.545477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the style mapper on images from a third polyp dataset, keeping all CVC-ColonDB images out of both training stages, and then test on CVC-ColonDB; if the Dice and IoU gains disappear, the reported improvement came from test-domain leakage into the style-mapping stage rather than from structure-preserving stylization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the OSASIS method and the Structure-Preserving Network that the paper adapts to medical segmentation."},{"cited_title":"Preechakul, N","cited_arxiv_id":null,"evidence_quote":"Provides the DiffAE semantic encoder that captures content so style transfer does not erase semantics."},{"cited_title":"Bernal, F","cited_arxiv_id":null,"evidence_quote":"CVC-ClinicDB is the training source for the colonoscopy polyp experiment."},{"cited_title":"Tajbakhsh, S","cited_arxiv_id":null,"evidence_quote":"CVC-ColonDB supplies the test set and the style-target images in the polyp experiment."},{"cited_title":"Tschandl, C","cited_arxiv_id":null,"evidence_quote":"HAM10000 provides the skin-lesion images used for the second experiment."}],"review_version":1}