{"id":"6b0e22c7-e281-453d-b989-8c57266638a4","arxiv_id":"2508.09189","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A Swin Transformer plus CNN decoder reports 0.9555 recall and 0.9849 accuracy on Kvasir-SEG, but lower F1 and precision than DUCK-Net, with claimed attention innovations absent from the architecture.","lead":"A hybrid Swin Transformer and CNN model for polyp segmentation reports higher recall and accuracy than DUCK-Net on Kvasir-SEG. The paper claims state-of-the-art results, but its own table shows lower F1 and precision, and the described decoder lacks the attention mechanisms it claims.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set early stopping and selective reporting leave the SOTA claim unsupported: Table 2 shows lower F1/precision than DUCK-Net, and the reported recall/accuracy gains come from a test set used for model selection.","rationale":"The reader's REJECT verdict is well supported. The paper's abstract promises a hybrid architecture with boundary-aware attention and artifact resilience, but the described decoder is a plain convolutional U-Net-style decoder with concatenation skip connections and no attention mechanism; the artifact-robustness claim has no experiments. More importantly, the only quantitative comparison table contradicts the headline claim on F1 and precision, and the model's mDice/mIoU are missing. The cleanest load-bearing flaw is the use of the test set for early stopping: Section 4 explicitly says the Dice score was inspected on the test set after every epoch, so the reported recall and accuracy are model-selection artifacts. This invalidates the claim of superiority over DUCK-Net, whose numbers were not obtained under the same protocol. A retraining experiment with a proper validation split and a fully held-out test set would settle the question. I therefore agree with the reader's weakest assumption and see no reason to adjust the REJECT verdict.","tokens_in":10062,"tokens_out":3662,"duration_ms":38346,"concrete_test":"Retrain the model from scratch with a separate validation split (e.g., 800 train/100 val/100 test on Kvasir-SEG), use validation Dice for early stopping only, and never touch the test set until after training. Then report test-set F1, mDice, mIoU, precision, recall, and accuracy for both Hybrid and DUCK-Net retrained under the identical split and protocol. If Hybrid's recall/accuracy gains shrink or disappear, or its F1/mDice remain below DUCK-Net, the central claim of superiority is falsified. Also verify whether the released code/architecture actually contains the promised boundary-aware attention modules.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'superior performance over existing solutions'—rests entirely on the Kvasir-SEG numbers in Table 2. Two concrete flaws are load-bearing. First, Section 4 states that early stopping 'inspected the Dice score on the test set after every epoch' and stopped at epoch 63. That means the test set was used for model selection, so the reported recall (0.9555) and accuracy (0.9849) are not unbiased estimates of generalization; they are selected values. Any comparison with DUCK-Net's published numbers is therefore invalid unless DUCK-Net was selected under the same protocol. Second, even accepting the table at face value, the model is not superior: F1 0.9499 < DUCK-Net 0.9502, precision 0.9422 < 0.9628, and mDice/mIoU for Hybrid are absent. The only claimed gains are recall (+0.0176) and accuracy (+0.0007), metrics that are less central for polyp segmentation and that are exactly the ones vulnerable to the test-set selection bias. The architecture section also describes no boundary-aware attention module despite the abstract's attribution, so the mechanism claimed for the recall gain is not in the described model. The claim of artifact resilience is never tested. Without a clean held-out evaluation and complete metrics, the SOTA claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid Swin-Transformer/CNN architecture for polyp segmentation and reports results on Kvasir-SEG, claiming superior performance over state-of-the-art methods, including DUCK-Net. The architecture is a standard encoder–decoder with skip connections, and the reported evaluation uses a 900/100 train/test split. The central claims are that the model improves recall and accuracy over DUCK-Net and that it is robust to endoscopic artifacts and ill-defined polyp boundaries.","tokens_in":10449,"tokens_out":2781,"duration_ms":28845,"significance":"If the reported results were obtained under a sound protocol and the architecture matched the described innovations, the work could be of interest for real-time polyp segmentation. The paper is clearly written and the architecture is easy to follow. However, the evaluation protocol is flawed, the reported metrics do not support the superiority claim even at face value, and several abstract-level claims are not backed by experiments or by the described architecture. The current manuscript does not provide reliable evidence for its central claim.","major_comments":[{"comment":"The early-stopping criterion is explicitly computed on the test set: 'inspecting the Dice score on the test set after every epoch' and 'cross-validation was applied solely on the test set within the Kvasir dataset.' This means the reported test metrics are not unbiased estimates of generalization; they are selected using the same data on which they are reported. Any comparison with DUCK-Net's published numbers is therefore invalid unless DUCK-Net was selected under the same protocol. A separate validation split is required before any claim of superiority can be made.","section":"Section 4 and §5.1"},{"comment":"Even accepting the table at face value, the model is not superior: F1 is 0.9499 versus DUCK-Net's 0.9502, precision is 0.9422 versus 0.9628, and mDice/mIoU for the proposed model are missing (dash). The only claimed gains are recall and accuracy, and the recall value is inconsistent between the abstract (0.9555) and Table 2 (0.955). To support the central claim, the authors need to report the complete metric set, ideally with confidence intervals, and compare all models under the same evaluation protocol.","section":"Table 2 and §6"},{"comment":"The abstract attributes the performance to 'boundary-aware attention mechanisms' and robustness to 'specular highlights, motion blur, and fluid occlusions,' but the architecture described in §3.1–3.3 is a standard Swin encoder with CNN decoder blocks and concatenation skip connections; there is no boundary-aware attention module, and no artifact-robustness experiments are reported anywhere in the manuscript. These claims are therefore unsupported by the described method and experiments.","section":"Abstract versus §3"},{"comment":"The paper states that the model was further tested on CVC-ClinicDB, ETIS, CVC-ColonDB, and Endotect, and the conclusion claims SOTA performance across DSC, IoU, precision, recall, F2, and FPS. However, no results for any external dataset are presented, and Table 2 does not report mDice, mIoU, F2, or FPS for the proposed model. These claims should be substantiated or removed.","section":"§5.1 and §7"}],"minor_comments":[{"comment":"There are typos and formatting issues: 'Kavasir-SEG' should be 'Kvasir-SEG', '33 convolutional' should be '3×3 convolutional', 'coscusative performance' appears to be a typo, and Figure 4's caption says 'fifty epochs' while Section 4 states training ran for 100 epochs.","section":"Throughout"},{"comment":"The reference list contains entries that appear to be irrelevant to the manuscript (e.g., [4], [15], [46], [51]) and some citations are not properly matched to the text. The references should be pruned and verified.","section":"References"}],"recommendation":"reject","confidential_remarks":"The evaluation-protocol issue alone is disqualifying: using the test set for early stopping invalidates the reported generalization numbers. The paper's own Table 2 also fails to show the claimed superiority. Retraining with a proper validation split, reporting complete metrics on all datasets, and adding the claimed artifact experiments would require a substantial revision that goes beyond the current manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper's central claim does not hold up. The architecture is a Swin Transformer backbone with a plain upsampling CNN decoder—both standard in the cited literature—and the three 'innovations' promised in Section 3 (adaptive fusion, context-preserving skips, cross-attention bridge) never appear in the equations. What's left is an incremental configuration on Kvasir-SEG.\n\nWhat it does well: the related work is broad and the clinical framing is accurate. The decoder equations are simple enough to reimplement, and the author is transparent about the training setup, which is exactly how we can see the problem.\n\nThe soft spots are load-bearing. Section 4 states early stopping 'inspected the Dice score on the test set after every epoch.' That means the reported recall (0.9555) and accuracy (0.9849) are selected values, not unbiased estimates of generalization; any comparison with DUCK-Net's published numbers is invalid unless DUCK-Net used the same selection procedure. Second, even taking Table 2 at face value, the model isn't superior: F1 is 0.9499 versus DUCK-Net's 0.9502, precision is 0.9422 versus 0.9628, and mDice/mIoU are absent. The only claimed gains are recall and accuracy, the two metrics most vulnerable to test-set selection. Third, the abstract attributes boundary-aware attention and artifact resilience, but no such mechanism is described and no artifact robustness experiments are reported. The conclusion mentions FPS and parameter counts that never appear in the results.\n\nThese are not minor omissions. The empirical claim is unsupported, and the mechanism claimed for the recall gain is not in the described model. The paper is readable and the citations are relevant, but the novelty is minimal.\n\nWho is this for? A reader scanning the hybrid-polyp-segmentation literature might skim the related work, but the results should not be cited. I would not bring it to a reading group, and I would not cite it. If I were the editor, I'd desk reject: the evaluation protocol is broken and the architecture contribution is too thin to justify referee time.","headline":"Test-set early stopping and a missing architecture description leave the SOTA claim unsupported; the paper is a clear but overclaimed incremental baseline.","tokens_in":10867,"tokens_out":4555,"would_cite":false,"duration_ms":44934,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a hybrid Swin Transformer-CNN network outperforms existing polyp-segmentation models, lifting recall to 0.9555 and accuracy to 0.9849 on Kvasir-SEG.","keywords":["colorectal cancer","polyp segmentation","colonoscopy","CNN","Transformer","Swin Transformer","hybrid architecture","Kvasir-SEG"],"falsifier":"Rerun the same 900/100 Kvasir-SEG training recipe but stop learning based on a held-out validation set rather than the test set, then recompute recall and accuracy on the test images. If recall falls below DUCK-Net's 0.9379 or accuracy below 0.9842, the paper's central superiority claim fails; a quicker check is to compute the missing mDice/mIoU cells in Table 2 and see whether they also exceed DUCK-Net.","tokens_in":10000,"feed_emoji":"🩺","tokens_out":10237,"duration_ms":99615,"temperature":0.7,"pith_summary":"Colonoscopists miss a fifth or more of polyps, and segmentation models that look accurate on benchmarks can still fail on flat lesions with blurred edges. The paper argues that a hybrid network—Swin Transformer encoder plus lightweight CNN decoder—cuts those misses: on Kvasir-SEG it reports recall of 0.9555 (1.76 percentage points above DUCK-Net) and accuracy of 0.9849. The authors trace the gain to boundary-aware attention and artifact resilience, and claim the model also runs at the highest frames per second among the compared methods. If those numbers hold, the architecture would be a practical option for real-time colonoscopy assistance rather than just an offline benchmark improvement.","feed_headline":"Hybrid Transformer-CNN hits 0.9555 recall on polyps","feed_subtitle":"Beats DUCK-Net on Kvasir-SEG recall and accuracy, a step toward catching flat polyps during colonoscopy.","key_machinery":"The central object is the staged hybrid encoder-decoder: a Swin Transformer backbone emitting feature maps $F_1,\\dots,F_4$ at resolutions $H/4$, $H/8$, $H/16$, $H/32$ with channel depths $C,2C,4C,8C$; a decoder $D_4,\\dots,D_1$ that refines each level with a $3\\times3$ conv-BN-ReLU block, bilinear 2x upsampling, and concatenative skip connections; and a final $1\\times1$ convolution plus bilinear resize to full resolution. The paper names adaptive fusion modules, context-preserving skip connections, and a cross-attention bridge as the three innovations that balance transformer and CNN features. The mechanism's job is to join long-range global context with pixel-level boundary detail, which is","core_discovery":"Central claim: a hybrid (Transformer + CNN) architecture can outperform established polyp-segmentation models on the Kvasir-SEG benchmark. The paper reports recall 0.9555 and accuracy 0.9849, improving on DUCK-Net's recall (0.9379) and accuracy (0.9842), and states that this comes with the highest FPS among the models compared. The design pairs a pre-trained Swin Transformer encoder—which produces four hierarchical feature maps from $H/4$ to $H/32$ using shifted-window self-attention—with a CNN decoder that restores spatial detail through conv-BN-ReLU blocks, bilinear upsampling, and concatenative skip connections. The author's explanation is that this division of labor handles both ill-defi","pith_inferences":["The paper's own Table 2 leaves mDice and mIoU blank for the hybrid model while DUCK-Net reports 0.9051 mIoU; filling those cells with the same test set would show whether the 'superior performance' claim is purely a recall-and-accuracy result or holds across overlap metrics.","The abstract credits boundary-aware attention, but Section 3's equations specify only conv-BN-ReLU blocks and bilinear upsampling; an ablation that turns off the cross-attention bridge would reveal whether the recall gain actually comes from explicit boundary modeling or from the multi-scale skip connections alone.","A natural extension is to run the same 900/100 Kvasir-SEG split with early stopping on a validation split and to publish per-dataset results for CVC-ClinicDB, ETIS, CVC-ColonDB, and Endotect; that would separate architecture gains from test-set tuning."],"forward_implications":["On Kvasir-SEG, the hybrid model would detect a larger share of polyp pixels than DUCK-Net (recall 0.9555 vs 0.9379), raising the chance that small or flat lesions are visible to the endoscopist.","The reported frames-per-second means the accuracy gain would not necessarily block live 1080p/30fps use, making the model a candidate for real-time assistance.","If the multi-dataset tests on CVC-ClinicDB, ETIS, CVC-ColonDB, and Endotect are as claimed, the design transfers across different scopes, resolutions, and polyp appearances.","The same encoder-decoder split—transformer for context, CNN for detail—would apply to other medical segmentation problems where boundaries are weak."],"supporting_citations":[{"why":"DUCK-Net, the state-of-the-art baseline whose recall and accuracy the hybrid model claims to beat on Kvasir-SEG.","marker":"[59]"},{"why":"Kvasir-SEG, the 1000-image benchmark dataset providing the training and test split for the main comparison.","marker":"[26]"},{"why":"Swin Transformer, the encoder backbone supplying shifted-window self-attention and the four-stage feature hierarchy.","marker":"[35]"},{"why":"U-Net, the canonical encoder-decoder with skip connections that the decoder design and comparisons extend.","marker":"[47]"},{"why":"ResUNet++, a real-time convolutional baseline the paper targets for speed-accuracy trade-off.","marker":"[27]"},{"why":"PraNet, a reverse-attention baseline that set the polyp-segmentation performance target before DUCK-Net.","marker":"[17]"}],"fun_headline_variants":["Hybrid transformer-CNN beats DUCK-Net on polyps","Swin+CNN hybrid tops polyp recall at 0.9555","Hybrid model for polyps: higher recall, faster FPS","Transformer-CNN hybrid edges out SOTA on colon polyps","Hybrid architecture lifts polyp recall and accuracy"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The reported recall and accuracy gains would still appear on genuinely unseen colonoscopy images, even though training was stopped by watching the Dice score on the same test set.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid transformer-CNN beats DUCK-Net on polyps","Swin+CNN hybrid tops polyp recall at 0.9555","Hybrid model for polyps: higher recall, faster FPS","Transformer-CNN hybrid edges out SOTA on colon polyps","Hybrid architecture lifts polyp recall and accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1481,"prompt_tokens":776,"completion_tokens":705,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":520,"tokens_out":705,"duration_ms":7715,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:02:20.629301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same 900/100 Kvasir-SEG training recipe but stop learning based on a held-out validation set rather than the test set, then recompute recall and accuracy on the test images. If recall falls below DUCK-Net's 0.9379 or accuracy below 0.9842, the paper's central superiority claim fails; a quicker check is to compute the missing mDice/mIoU cells in Table 2 and see whether they also exceed DUCK-Net.","supporting_citations":[{"cited_title":"Duck-net: Dense u-shaped convolutional ker- nel network for polyp segmentation","cited_arxiv_id":null,"evidence_quote":"DUCK-Net, the state-of-the-art baseline whose recall and accuracy the hybrid model claims to beat on Kvasir-SEG."},{"cited_title":"Kvasir-SEG: A Segmented Polyp Dataset","cited_arxiv_id":"1911.07069","evidence_quote":"Kvasir-SEG, the 1000-image benchmark dataset providing the training and test split for the main comparison."},{"cited_title":"Swin trans- former: Hierarchical vision transformer using shifted win- dows","cited_arxiv_id":null,"evidence_quote":"Swin Transformer, the encoder backbone supplying shifted-window self-attention and the four-stage feature hierarchy."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"U-Net, the canonical encoder-decoder with skip connections that the decoder design and comparisons extend."},{"cited_title":"Smedsrud, Michael A","cited_arxiv_id":null,"evidence_quote":"ResUNet++, a real-time convolutional baseline the paper targets for speed-accuracy trade-off."},{"cited_title":"Pranet: Parallel reverse attention network for polyp segmentation","cited_arxiv_id":null,"evidence_quote":"PraNet, a reverse-attention baseline that set the polyp-segmentation performance target before DUCK-Net."}],"review_version":1}