{"id":"60b618f3-65ee-4e8c-8abc-0309ed658065","arxiv_id":"2608.10649","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A three-stage EfficientNetV2 and UNet++ pipeline claims near-perfect colorectal polyp classification and segmentation, but the paper omits results for its subtype, Paris, JNet, and resection stages and does not release code.","lead":"PolypVision is a three-stage deep learning pipeline that classifies, segments, and subtypes colorectal polyps from endoscopy images, reporting high accuracy on public benchmarks. The paper lacks results for most of the claimed stages and appears to conflate polyp detection accuracy with histological classification accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 1 histology AUC is unsupported: Kvasir-SEG, CVC-ClinicDB, and PolypGen provide masks, not adenoma/hyperplasia labels, while ERCPMP, the only listed dataset with pathology annotations, is never evaluated.","rationale":"The reader's weakest assumption identifies essentially the same concern: the Stage 1 AUC is presented as evidence of adenoma-versus-hyperplasia classification even though the evaluated datasets provide only segmentation labels. My independent reading confirms this is the most load-bearing problem. The paper's Table 1 and §3 are internally inconsistent: three datasets are used for evaluation, yet none has histology labels, while the one dataset with pathology annotations (ERCPMP) is listed and described but absent from all reported experiments. This makes the headline result impossible to validate from the stated data. I also note the mAP@50 claim is contradicted by Table 2, where YOLO-LAN reports 96.19% against PolypVision's 94.4%, but that issue is secondary. No code, ablations, or error bars are provided, and no external verification exists; the label-provenance gap alone is sufficient to reject the paper's central claim. Therefore I agree with the reader's REJECT verdict and recommend no change.","tokens_in":7152,"tokens_out":3393,"duration_ms":37916,"concrete_test":"Download the official metadata for Kvasir-SEG and CVC-ClinicDB and verify that they contain only segmentation masks with no adenomatous/hyperplastic labels. Then obtain ERCPMP, the only listed dataset with pathologic annotations, and reproduce Stage 1 using the paper's protocol with a documented train/test split, reporting AUC on held-out images. If the three evaluated datasets contain no histology labels, or if the claimed AUC ~0.99 cannot be reproduced on ERCPMP, the central classification claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires Stage 1 to distinguish adenomatous from hyperplastic polyps with AUC ~0.99, yet the datasets named in the abstract and experiments do not contain histology labels. In §3 and Table 1, PolypGen is described as a segmentation/detection dataset, Kvasir-SEG provides pixel-wise masks, and CVC-ClinicDB provides frame-level annotations; none is listed with adenomatous/hyperplastic class labels. Nevertheless, §3 asserts that \"For Stage 1 (binary classification), images are labeled as adenomatous or hyperplastic,\" with no statement of where these labels came from. Table 1 separately lists ERCPMP (419 images, 37 videos) with \"endoscopic and pathologic annotations,\" and the Data Availability section says ERCPMP is one of the datasets used in this study, but no experiment or result in §5 uses ERCPMP. If ERCPMP supplied the Stage 1 labels, the paper must report the train/test split and results on it; if it did not, the claimed Stage 1 AUC cannot be computed from the three stated public datasets. The ambiguity is load-bearing because Table 3's \"frame classification\" could instead be polyp-versus-background detection, which is a much easier task on mask-annotated datasets and would not support the paper's clinical classification claim. Every downstream component, including Stage 2 transfer and Stage 3 subtype classification, inherits this unsupported Stage 1 premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PolypVision proposes a three-stage hierarchical deep learning framework for colorectal polyp analysis. Stage 1 performs binary classification of polyps as adenomatous or hyperplastic, together with Paris and JNet classification, using EfficientNetV2-M with Focal Loss. Stage 2 performs polyp segmentation with a UNet++ decoder, using the Stage 1 backbone as encoder, and also predicts a recommended resection method. Stage 3 classifies adenomas into tubular, tubulovillous, and villous subtypes using transfer learning from Stage 2. The paper reports an AUC of approximately 0.99 for frame classification and a detection mAP@50 of 94.4% on Kvasir-SEG, and includes a Grad-CAM interpretability analysis. The central technical idea of progressive transfer between classification, segmentation, and subtype classification is plausible, but the reported experiments do not substantiate most of the claimed outputs, and the main classification result appears to be measured on datasets that lack histology labels.","tokens_in":7415,"tokens_out":6830,"duration_ms":71188,"significance":"If the framework truly performed adenoma-versus-hyperplasia classification at AUC ~0.99 while simultaneously providing segmentation, Paris/JNet characterization, resection recommendations, and subtype classification, it would be a clinically valuable contribution. The use of public benchmarks, the shared-backbone transfer-learning design, and the publicly accessible web application are constructive elements. However, the manuscript as written does not provide evidence for the central classification claim, omits results for several advertised stages, and does not report the segmentation metrics promised in the evaluation protocol. The significance of the work therefore depends entirely on whether the authors can supply the missing dataset provenance and experimental results.","major_comments":[{"comment":"The Stage 1 adenoma/hyperplasia claim is unsupported by the reported experiments. The manuscript states in §3 that 'For Stage 1 (binary classification), images are labeled as adenomatous or hyperplastic,' but it never identifies which dataset provides these labels. Kvasir-SEG and CVC-ClinicDB are described as segmentation/detection datasets with pixel-wise and frame-level annotations, respectively, and contain no histopathology labels; PolypGen is likewise a detection/segmentation dataset. Table 1 lists ERCPMP as the only dataset with 'endoscopic and pathologic annotations,' yet no experiment in §5 evaluates ERCPMP. Consequently, the AUC ~0.99 for 'Frame Classification' in Table 3 cannot be interpreted as validation of the adenomatous-vs-hyperplastic classifier; it is at best a frame-level polyp-detection metric on datasets whose labels do not include polyp type. This is a load-bearing gap because Stage 2 and Stage 3 inherit the Stage 1 premise.","section":"§3, Table 1, §5.3"},{"comment":"The segmentation results that are central to Stage 2 are not reported. §5.1 states that Stage 2 is evaluated using Dice Score, IoU, and Precision/Recall, but the only quantitative result in §5.2 is mAP@50 for detection on Kvasir-SEG; no Dice or IoU values appear anywhere, and no CVC-ClinicDB result is given for PolypVision in Table 2. Moreover, the manuscript never defines how mAP@50 is computed for this model (e.g., whether bounding boxes are derived from predicted segmentation masks). Without segmentation metrics, the claimed UNet++ contribution and the subsequent transfer to Stage 3 cannot be assessed.","section":"§5.1, §5.2, Table 2"},{"comment":"Multiple components advertised in the abstract and method sections have no experimental evaluation. No results are reported for the Paris classification, JNet classification, recommended resection method (cold snare polypectomy, EMR, or ESD), or Stage 3 adenoma subtype classification (tubular, tubulovillous, villous). The only classification table (Table 3) reports generic 'Frame Classification' AUC on Kvasir-SEG and CVC-ClinicDB, datasets that contain none of these labels. These omissions are not cosmetic: the paper's title and abstract promise a complete clinical pipeline, yet the evidence covers only a subset of the claimed outputs.","section":"§4.3–§4.5, §5"},{"comment":"Reported performance lacks statistical anchoring. The AUC values in Table 3 are given as '~0.99' without confidence intervals, error bars, or significance tests, and the number of test images per class is not reported. Given that Kvasir-SEG and CVC-ClinicDB are small datasets (1,000 and 612 images, respectively), the headline claim would be more credible with per-class counts and confidence intervals. Additionally, no confusion matrix or class-wise precision/recall is provided for the classification results.","section":"§5.1, §5.3"}],"minor_comments":[{"comment":"The running header on the first page reads 'PolypVision: Ensemble Deep Learning for Colorectal Polyp Classification,' which does not match the main title describing a three-stage hierarchical framework; please harmonize the titles.","section":"Title page"},{"comment":"The sentence 'Stratified train/validation/test splits are applied to preserve class balance across all partitions' is undefined for Stage 1 because the label source is not identified; once the label provenance is clarified, the split process should be described explicitly.","section":"§3"},{"comment":"The Grad-CAM analysis is only qualitative; a quantitative evaluation (e.g., overlap between attention heatmaps and polyp masks, or insertion/deletion metrics) would strengthen the interpretability claim.","section":"§5.4"},{"comment":"The device-independence claim is not supported by any acquisition-system or device metadata in the evaluation; multi-center data alone does not establish device independence.","section":"§6"},{"comment":"The citation for the ResNet-50/VGG16 AUC values is unclear; reference [12] is He et al. on ResNet, and the source of the 0.91–0.98 range should be cited precisely.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The ERCPMP dataset, whose source paper shares the corresponding author, is listed in Table 1 and in the Data Availability statement as a dataset used in this study, but no ERCPMP results appear in §5. The editor may wish to ask the authors to clarify the exact role of ERCPMP, including whether it supplied the Stage 1 histology labels, and to explain why it is absent from the experimental section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the three-stage hierarchical idea is reasonable, but the paper doesn't actually report results for two of the three stages, and the headline 'classification' AUC is probably measuring polyp-vs-background detection rather than adenoma-vs-hyperplasia.\n\nThe architecture and training choices are described clearly: EfficientNetV2-M backbone, UNet++ decoder, Focal Loss for class imbalance, Dice+BCE for segmentation, MixUp at Stage 3. Sharing the backbone across stages is a legitimate way to reduce data requirements, and the Grad-CAM figures at least show the model focuses on the lesion in some cases. The web app is real and makes the work reproducible in the loose sense. These are genuine positives.\n\nBut the central claim does not hold up as written. Kvasir-SEG, CVC-ClinicDB and PolypGen provide masks, not histology labels. The only listed dataset with pathology annotations is ERCPMP, and no experiment uses it. So where do the 'adenomatous vs hyperplastic' labels come from? If they came from ERCPMP, the paper must report the train/test split and results on it. If they came from somewhere else, that source is missing. Table 3's AUC ~0.99 almost certainly reflects a frame-level polyp detection task on mask datasets, not histological classification. The comparison in Table 3 is therefore apples-to-oranges. The mAP@50 comparison also undercuts the 'state-of-the-art' claim: YOLO-LAN reaches 96.19% on Kvasir-SEG, higher than the reported 94.4%.\n\nBeyond that, there are no results at all for Stage 2 segmentation (Dice, IoU), no Stage 3 subtype results, no Paris/JNet results, no resection-method results. No error bars, no ablation isolating the hierarchical transfer contribution, no code or model weights. The paper's own limitation paragraph acknowledges Stage 1 error propagation and calls for prospective validation, which is honest, but the missing evidence is not a minor detail: the system's headline capability is unverified.\n\nWho is this for? A reviewer or researcher wanting to see whether a hierarchical transfer pipeline can beat single-stage baselines. The paper currently can't answer that question, but the design is worth taking seriously. I'd send it to peer review with a clear request for label provenance, all-stage results, and ablations. The authors have built something; they need to show it works.","headline":"A plausible three-stage polyp pipeline that currently only reports detection-level numbers; the histology AUC claim needs label provenance before it can be taken seriously.","tokens_in":7999,"tokens_out":2456,"would_cite":false,"duration_ms":25836,"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":"Three-stage hierarchical deep learning can classify, segment, and subtype colorectal polyps in one pipeline, with about 0.99 AUC and 94.4% mAP@50 on Kvasir-SEG.","keywords":["Colorectal polyps","Deep learning","EfficientNetV2-M","UNet++","Polyp segmentation","Polyp classification","Grad-CAM","Transfer learning"],"falsifier":"Run the trained Stage 1 model on a dataset with explicit per-polyp histology labels—for example the ERCPMP pathology annotations the paper lists but never analyzes—and compute AUC for adenomatous versus hyperplastic on those ground-truth labels. If the AUC drops materially below ~0.99, or if no such histology labels were used in training, the classification claim is not supported by the reported experiments.","tokens_in":6949,"feed_emoji":"🩺","tokens_out":10979,"duration_ms":93535,"temperature":0.7,"pith_summary":"PolypVision is a three-stage hierarchical deep learning framework intended to give a complete colorectal polyp assessment in one pipeline: first distinguishing adenomatous from hyperplastic polyps while also predicting Paris and JNet endoscopic patterns, then producing pixel-level segmentation along with a recommended resection method, and finally subtyping adenomas into tubular, tubulovillous, and villous. The authors report frame-classification AUC around 0.99 on Kvasir-SEG and CVC-ClinicDB, and a detection mAP@50 of 94.4% on Kvasir-SEG, matching or exceeding previous methods. The claim is that sharing one EfficientNetV2-M backbone across stages, with progressive transfer learning and task-specific losses, lets each task benefit from the others and yields clinically useful accuracy without hardware-specific adaptation. A reader should care because the pipeline follows the actual clinical sequence—detect, characterize, segment, resect, subtype—rather than solving each task in isolation, and the authors make it available as a freely usable web application.","feed_headline":"Three-stage AI pipeline hits 99% AUC on colon polyp typing","feed_subtitle":"PolypVision reports 0.99 AUC and 94.4% mAP@50 on Kvasir-SEG by chaining classification, segmentation, and subtyping.","key_machinery":"The load-bearing mechanism is the three-stage hierarchy itself, built on a shared EfficientNetV2-M backbone with 52.8 million parameters. Stage 1 fine-tunes this backbone with Focal Loss ($\\gamma=2$, $\\alpha=0.25$) for adenomatous/hyperplastic classification plus Paris and JNet heads; Stage 2 reuses the Stage 1 backbone as the UNet++ encoder and trains with Dice plus binary cross-entropy for segmentation and cross-entropy for resection recommendation; Stage 3 reinitializes the backbone from Stage 2 and uses Focal Loss with MixUp augmentation for three-way adenoma subtyping. Progressive transfer learning (Stage 1 to Stage 2 to Stage 3), layer-wise learning-rate decay, and early stopping on validation AUC or Dice carry the argument that limited labeled data can be leveraged across related clinical tasks.","core_discovery":"On the paper's own terms, the central discovery is that a single sequential pipeline—Stage 1 adenomatous-versus-hyperplastic classification with simultaneous Paris and JNet outputs, Stage 2 UNet++ segmentation with resection-method recommendation, and Stage 3 histologic subtyping—can reach state-of-the-art accuracy while reusing one shared backbone. The paper reports AUC of about 0.99 for frame classification on both Kvasir-SEG and CVC-ClinicDB and mAP@50 of 94.4% on Kvasir-SEG, and interprets Grad-CAM maps as evidence that the classifier attends to lesion surface texture, pit pattern, and margins. The intended implication is that hierarchical transfer learning across detection, segmentation, and classification is beneficial: each stage supplies the next with learned representations, and the combined framework covers the full clinical workflow in a device-independent way.","pith_inferences":["The headline AUC may be measuring something weaker than histological classification: Kvasir-SEG and CVC-ClinicDB provide only pixel masks and no stated histology labels, so unless the authors had separate pathology labels for those frames, the reported 'frame classification' may be polyp-versus-background detection rather than adenoma-versus-hyperplastic typing.","A direct test of the Stage 1 claim would be to evaluate the same trained model on the ERCPMP dataset's explicit pathology labels, which the paper lists in its datasets and data availability but never describes in the methods or results.","The Paris and JNet outputs and the resection-method recommendations are announced as part of the pipeline, yet no quantitative metrics are reported for them; their accuracy is an open empirical question.","A useful extension would be to report Stage 3 performance conditioned on Stage 1 being correct, isolating the hierarchical error propagation the authors acknowledge as a limitation."],"forward_implications":["If the reported numbers hold, one installed pipeline can give an endoscopist the polyp type, Paris and JNet morphology, a boundary mask, a resection suggestion, and the adenoma subtype from a single image.","The mAP@50 of 94.4% on Kvasir-SEG places the detector between YOLO-LAN's 96.19% and YOLOv8-s's 91.16%, showing a segmentation-driven hierarchical approach can compete with dedicated detectors.","Because Stage 2's encoder is reused to initialize Stage 3, improvements in segmentation quality should flow directly into subtype classification, making progressive transfer a viable path for data-scarce medical imaging.","Device independence would mean colonoscopy units can adopt the framework without recalibrating to proprietary imaging modes or narrow-band systems.","Stage 1 errors propagate downstream by design: any adenoma misclassified as hyperplastic is never passed to the subtyping stage, so overall clinical accuracy is bounded by Stage 1 recall."],"supporting_citations":[{"why":"Supplies the EfficientNetV2-M backbone shared by all three stages.","marker":"[9]"},{"why":"Supplies the UNet++ decoder with nested skip connections used for Stage 2 segmentation.","marker":"[10]"},{"why":"Provides the ResNet/VGG16 baseline reporting 0.91–0.98 AUC on Kvasir-SEG that the paper compares against.","marker":"[12]"},{"why":"Supplies the YOLOv8-s detector baseline reporting 91.16% mAP@50 on Kvasir-SEG.","marker":"[13]"},{"why":"Supplies the YOLO-LAN baseline reporting 96.19% mAP@50 on Kvasir-SEG.","marker":"[14]"},{"why":"Supplies the multi-center PolypGen dataset used for training and evaluation.","marker":"[15]"},{"why":"Supplies the Kvasir-SEG image-mask pairs used for segmentation and the mAP@50 detection benchmark.","marker":"[16]"},{"why":"Supplies the CVC-ClinicDB frames used for segmentation and frame-classification evaluation.","marker":"[17]"}],"fun_headline_variants":["99% AUC polyp AI: classify, segment, subtype in one pipeline","Three-stage polyp AI: 99% AUC plus segmentation and subtyping","PolypVision: 0.99 AUC and 94.4% mAP via three-stage AI","Single neural pipeline classifies, segments, and subtypes polyps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported ~0.99 AUC actually measures adenoma-versus-hyperplastic classification, even though the two datasets used for that result contain pixel masks and no stated histology labels, and the paper never says where those labels come from.","fun_headline_variants_meta":{"raw":{"variants":["99% AUC polyp AI: classify, segment, subtype in one pipeline","Three-stage polyp AI: 99% AUC plus segmentation and subtyping","PolypVision: 0.99 AUC and 94.4% mAP via three-stage AI","Single neural pipeline classifies, segments, and subtypes polyps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000837,"raw_usage":{"total_tokens":3712,"prompt_tokens":1068,"completion_tokens":2644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":2568}},"tokens_in":684,"tokens_out":2644,"duration_ms":20870,"temperature":1.0,"reasoning_tokens":2568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:07:29.331497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained Stage 1 model on a dataset with explicit per-polyp histology labels—for example the ERCPMP pathology annotations the paper lists but never analyzes—and compute AUC for adenomatous versus hyperplastic on those ground-truth labels. If the AUC drops materially below ~0.99, or if no such histology labels were used in training, the classification claim is not supported by the reported experiments.","supporting_citations":[{"cited_title":"UNet++: A nested U -net architecture for medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet++ decoder with nested skip connections used for Stage 2 segmentation."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet/VGG16 baseline reporting 0.91–0.98 AUC on Kvasir-SEG that the paper compares against."},{"cited_title":"YOLOv8 for polyp detection: a comparative study","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv8-s detector baseline reporting 91.16% mAP@50 on Kvasir-SEG."},{"cited_title":"YOLO-LAN: A lightweight YOLO for polyp detection","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLO-LAN baseline reporting 96.19% mAP@50 on Kvasir-SEG."},{"cited_title":"Kvasir-SEG: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the Kvasir-SEG image-mask pairs used for segmentation and the mAP@50 detection benchmark."},{"cited_title":"CVC -ClinicDB: A clinical colonoscopy database for polyp detection","cited_arxiv_id":null,"evidence_quote":"Supplies the CVC-ClinicDB frames used for segmentation and frame-classification evaluation."}],"review_version":1}