{"id":"865bbbc3-7a85-4ee2-a267-1e39216ea5b4","arxiv_id":"2411.14385","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid fuzzy clustering plus DeepLabv3+ pipeline is claimed to segment gastric bleeding lesions with 87.95% accuracy, but missing experimental detail and internal inconsistencies leave the claim unsupported.","lead":"This paper proposes a two-stage pipeline that combines fuzzy clustering with a DeepLabv3+ neural network to segment gastric bleeding lesions in endoscopic images. The authors report 87.95% accuracy and 96.33% specificity, but provide no code, data, or detailed experimental protocol to support the claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline accuracy/specificity claim is not testable from the paper: the DuS-KFCM pre-segmentation is generated on the same images later used for training/testing, no split or control ablation is reported, and Eqs. (2)/(8) define sensitivity/specificity incorrectly.","rationale":"The central claim is the quantitative outperformance claim in the abstract. For that claim to hold, the reported numbers must be computed with a valid, controlled evaluation protocol and correct metric definitions. The manuscript satisfies neither: Section 2.2 gives no explicit split; the pre-segmentation is performed on the same images that are then refined, with no ablation isolating the deep stage; and the sensitivity/specificity equations are wrong. These are internal-consistency failures, not merely disagreements with community practice. Even if the two-stage architecture is a plausible idea, the evidence chain does not support the headline. I therefore agree with the reader's REJECT verdict and see no reason to change it. This stress-test adds a more pointed technical basis: the specificity number quoted in the abstract cannot be reproduced from Eq. (8), and the uncontrolled pre-segmentation makes the accuracy gain uninterpretable. The proposed check would settle whether the concern is actual leakage or merely missing documentation. If the authors supplied predictions and the control arm, the claim could be evaluated; without those, the manuscript remains unverifiable.","tokens_in":8824,"tokens_out":8421,"duration_ms":80565,"concrete_test":"Ask the authors to release code, trained models, and per-image predictions, then run a single controlled experiment: split the Kvasir-SEG GB and GRS images at the patient/session level into train/validation/test; train DeepLabv3+ (a) on raw RGB images and (b) on RGB images plus DuS-KFCM coarse masks, with identical hyperparameters and no feature selection on the combined set; evaluate both on the untouched test set using standard definitions (sensitivity=TP/(TP+FN), specificity=TN/(TN+FP), Dice, IoU). If arm (b) does not beat arm (a) by a meaningful margin, or if the recomputed specificity is not approximately 96.33%, the central outperformance claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 is the only protocol description: 'We start by taking each image from the training set and segmenting the gastric bleeding lesions from the image using the DuS-KFCM fuzzy clustering technique... Then, for every testing image, we segment the GB lesion using the same method...'. It never states how training and test sets were formed, whether same endoscopy sessions/images appear in both, or whether CFS feature selection / hyperparameters were tuned on the test data. Because the coarse DuS-KFCM mask is computed from the same image that later enters DeepLabv3+, the reported 87.95% accuracy and 96.33% specificity may be an artifact of giving the deep network a strong per-image prior rather than of the proposed refinement. No ablation separates DeepLabv3+ alone from Deep DuS-KFCM, and no held-out split or cross-validation detail is given. Independently, the metrics are internally wrong: Eq. (2) defines sensitivity as TP/(TN+FN) and Eq. (8) defines specificity as TN/(TN+FN), instead of TP/(TP+FN) and TN/(TN+FP); the abstract's 96.33% specificity does not appear anywhere in the results section. The central quantitative claim is therefore not interpretable as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes Deep DuS-KFCM, a coarse-to-fine segmentation pipeline for gastric bleeding lesions in endoscopic images. The method first applies a fuzzy clustering technique (DuS-KFCM) with GLCM and color features to obtain a coarse lesion mask, then refines that mask with a DeepLabv3+ (ResNet50) network. The paper claims high accuracy (87.95%) and specificity (96.33%) on public gastric bleeding and gastric red spot datasets and states that the method outperforms contemporary segmentation approaches. The central contribution, as presented, is a hybrid neuro-fuzzy segmentation architecture.","tokens_in":9151,"tokens_out":3941,"duration_ms":37113,"significance":"If the reported results were reproducible and the metric definitions correct, the proposed combination of fuzzy clustering priors with a deep refinement network could be a practically relevant contribution to endoscopic lesion segmentation. However, the manuscript as written does not support this claim: the sensitivity and specificity formulas are mathematically wrong, the experimental protocol permits information leakage between the clustering and the deep refinement stages, the reported numbers are inconsistent across sections, and no reproducibility details (splits, hyperparameters, error bars) are provided. The clinical motivation is sound, but the evidence presented is not sufficient to assess the method's actual performance.","major_comments":[{"comment":"The definitions of sensitivity and specificity are incorrect. Eq. (2) defines sensitivity as TP/(TN+FN), which omits true positives from the denominator; the standard definition is TP/(TP+FN). Eq. (8) defines specificity as TN/(TN+FN), using false negatives instead of false positives; the standard definition is TN/(TN+FP). Because all reported performance values in Section 4 rely on these metrics, the quantitative results are not interpretable and any comparison to literature values is invalid.","section":"Section 3, Eqs. (2) and (8)"},{"comment":"The experimental protocol applies the DuS-KFCM clustering algorithm to the same images that are later used for training and testing the deep network, and the test-time pipeline feeds the clustering mask into the final classification stage. The paper does not report an ablation separating DeepLabv3+ alone from the full pipeline, nor does it provide any held-out split or cross-validation details. Consequently, the reported accuracy of 87.95% and specificity of 96.33% cannot be attributed to the proposed refinement; they may be driven by the strong per-image prior supplied by the clustering step.","section":"Section 2.2 and Figure 1(c)"},{"comment":"The quantitative claims are internally inconsistent. The abstract reports an accuracy of 87.95% and a specificity of 96.33%; Section 4 reports 'a remarkable accuracy of 98%' in endoscopy gastric images; and the conclusion reports accuracy values of 87.95% and 79.72% and precision values of 86.69% and 62.82%. The specificity value from the abstract does not appear anywhere in the results section. These contradictions make the central quantitative claim unverifiable as written.","section":"Section 4 and abstract"},{"comment":"The experimental setup is not reproducible. The paper specifies no train/validation/test split sizes, no hyperparameters for DeepLabv3+ training (optimizer, learning rate, batch size, number of epochs), no details on the CFS feature selection procedure, no values for the fuzzy clustering parameters (fuzzifier m, number of clusters), and no error bars or statistical significance tests. Without these details, the reported 'unprecedented accuracy' cannot be independently verified or reproduced.","section":"Section 4"},{"comment":"The claim that the method outperforms 'contemporary segmentation methods' is not supported by the comparisons presented. The baselines used are FKM, GMM, and FCM, which are classical clustering methods, not modern deep-learning segmentation approaches. The paper makes no comparison to U-Net, DeepLab variants, or other published segmentation methods on the same datasets, and it does not cite or benchmark against known Kvasir-SEG results. The superiority claim is therefore unsubstantiated.","section":"Sections 1 and 4"}],"minor_comments":[{"comment":"The method name is inconsistently written as both 'Dus-KFCM' and 'DuS-KFCM'; please standardize the spelling in the abstract, main text, and figures.","section":"Throughout"},{"comment":"Equation (1) uses 'Si ∩ Gi' without defining how the intersection is computed across multiple clusters or how overlapping regions are handled; please provide a precise mathematical definition.","section":"Eq. (1)"},{"comment":"The text states that specificity is 'expressed in (9)' but the equation itself is numbered (8); the cross-reference is incorrect.","section":"Section 3"},{"comment":"Reference [38] is incomplete, giving only a short author list and a DOI to an arXiv preprint without a title or full author list; reference [6] cites a general cancer statistics webpage for a specific survival-rate claim, which should be supported by an epidemiological source.","section":"References"},{"comment":"The caption refers to 'Synthetic Pseudo-color Images,' but the relationship between these synthetic images and the real gastric bleeding/red spot datasets is never explained in the text.","section":"Figure 3 caption"}],"recommendation":"reject","confidential_remarks":"This manuscript reads as an early and internally inconsistent draft. The metric errors and the contradictory accuracy figures are not local issues; they invalidate the central claim. In addition, the experimental protocol as described cannot rule out leakage from the clustering prior into the deep refinement stage, and the absence of any held-out validation details makes the results irreproducible. A revision would require a complete redo of the experiments with proper validation, corrected metrics, and full reproducibility details, which goes beyond normal major revision. If the editor considers a resubmission, the burden of proof should be on the authors to provide code, splits, and hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is not ready for the literature. The clinical problem is real, and the rough idea of combining fuzzy clustering with DeepLabv3+ is not crazy, but the manuscript does not supply the evidence needed to evaluate it: no split, no code, no ablation, and the metric definitions in Section 3 are wrong. The headline numbers cannot be interpreted.\n\nWhat is actually new: the specific combination of SKFCM with GLCM features and DeepLabv3+ refinement applied to gastric bleeding and red spots from Kvasir-SEG. That is an application of known components. There is no new theory, no new dataset, no clinical validation.\n\nThe paper does a few things well: it works with a public dataset, it reports visual comparisons against FKM/GMM/FCM, and it situates the problem in a real clinical context. The qualitative figures suggest the method can find lesion boundaries on some images. That is worth acknowledging.\n\nThe soft spots are structural. First, the protocol in Section 2.2 says DuS-KFCM pre-segmentation is computed on each training image and then used to train DeepLabv3+, and the same clustering is applied to test images before refinement. But the paper never states how training and test sets were formed, whether images from the same endoscopy session appear in both, or whether any feature selection or hyperparameters were tuned on test data. So the reported 87.95% accuracy and 96.33% specificity may simply reflect the per-image prior from clustering, not a learned refinement. There is no ablation separating DeepLabv3+ alone from Deep DuS-KFCM.\n\nSecond, the metrics are defined incorrectly. Eq. (2) gives sensitivity = TP/(TN+FN); Eq. (8) gives specificity = TN/(TN+FN). Those are not sensitivity and specificity. The abstract says 96.33% specificity but the results section never reports that number; instead it says 98% accuracy and then later 87.95% accuracy. The numbers do not cohere.\n\nThird, no code, no hyperparameters, no error bars, and no held-out evaluation. The baselines are classic methods, not recent deep segmentation networks, so the comparison is weak. The stress-test concern is accurate: the central quantitative claim is not testable as written.\n\nWho is this for? Someone doing a survey of fuzzy-clustering-plus-deep approaches might note it, but the paper will not be citable until the experimental protocol and metrics are fixed. A serious editor should probably send it back for major revision at best, but honestly the current form is closer to a desk reject: the missing split and broken metrics are not stylistic, they are load-bearing. I would not spend a referee slot on it as is.","headline":"Short version: plausible clinical application, but the missing split, wrong metric formulas, and inconsistent results make the headline claim untestable as written.","tokens_in":9666,"tokens_out":2742,"would_cite":false,"duration_ms":24950,"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 that a two-stage pipeline—fuzzy clustering pre-segmentation followed by DeepLabv3+ refinement—segments gastric bleeding lesions in endoscopic images with 87.95% accuracy and 96.33% specificity on the gastric bleeding and…","keywords":["gastric bleeding segmentation","fuzzy C-means","DeepLabv3+","ResNet50","GLCM texture features","coarse-to-fine segmentation","endoscopic image analysis","hybrid neuro-fuzzy model"],"falsifier":"Train the DeepLabv3+ refinement stage on the original images alone (no DuS-KFCM masks) using the same train/test split, and compare it on a held-out test set whose images were never used to generate clustering masks; if accuracy and specificity do not improve materially, the coarse-to-fine contribution is not supported. A second check is to recompute the clustering masks on the test set independently and verify the refinement stage is not simply reproducing training-set masks.","tokens_in":8658,"feed_emoji":"🩸","tokens_out":13349,"duration_ms":112046,"temperature":0.7,"pith_summary":"The paper sets out to show that gastric bleeding lesions in endoscopic images can be segmented more accurately by combining fuzzy clustering with a deep refinement network than by either approach alone. It introduces Deep DuS-KFCM, a two-stage pipeline in which a spatial kernelized fuzzy C-means variant (DuS-KFCM) uses color and texture features to produce a coarse lesion mask, and DeepLabv3+ with a ResNet50 backbone refines that mask into the final boundary. On the gastric bleeding and gastric red-spot datasets, the authors report 87.95% accuracy and 96.33% specificity, with higher precision and IoU than FKM, GMM, FCM, and other compared methods. If the result holds, the method would give clinicians a fast, reproducible way to delineate bleeding regions from endoscopy frames without manual tracing.","feed_headline":"Fuzzy-plus-deep model hits 87.95% accuracy on gastric bleeding","feed_subtitle":"A two-stage pipeline pairs fuzzy clustering with deep-learning refinement to spot subtle bleeding lesions in endoscopy images.","key_machinery":"The machinery is a two-stage coarse-to-fine segmentation pipeline. Stage one is DuS-KFCM, a Dual Spatial Kernelized Constrained Fuzzy C-Means algorithm that fuses RGB color features with 22 GLCM texture statistics, selects discriminative features, and clusters pixels into coarse lesion and background regions. Stage two is DeepLabv3+ with a ResNet50 encoder: atrous convolutions and an atrous spatial pyramid pooling module gather multi-scale context, and a decoder fuses that context with low-level spatial features to refine the lesion boundary. The paper's argument is that the fuzzy stage handles the intensity-overlap problem and the deep stage recovers boundary precision, so the two stages carry different parts of the segmentation task.","core_discovery":"The central claim is that the intensity overlap between bleeding tissue and adjacent gastric structures can be resolved by a coarse-to-fine strategy rather than by a single segmentation model. In the paper's telling, DuS-KFCM first separates likely lesion pixels using fuzzy memberships built from spatial, color, and GLCM texture information; DeepLabv3+ then takes those coarse masks and the original image and refines them into sharp, noise-resistant lesion boundaries. The authors report that the combined model reaches 87.95% accuracy and 96.33% specificity on the gastric bleeding and gastric red-spot datasets, and they present visual and quantitative comparisons against FKM, GMM, and FCM showing better boundary detection. They interpret this as evidence that the hybrid neuro-fuzzy design is well suited to subtle bleeding symptoms.","pith_inferences":["The soft membership maps produced by fuzzy clustering need no ground truth, so they could be reused as pseudo-labels or attention priors for semi-supervised segmentation of unlabelled endoscopy video.","The same coarse-to-fine recipe—fuzzy pre-segmentation plus atrous-convolution refinement—could transfer to other endoscopic lesion types with similar intensity overlap, but the transfer should be validated on new datasets.","A direct ablation comparing the full model to DeepLabv3+ trained on raw images would separate the clustering stage's contribution from the deep stage's, giving a cleaner estimate of where the accuracy gain comes from."],"forward_implications":["Automated segmentation could give clinicians a highlighted lesion boundary from an endoscopy frame with less manual effort and less inter-observer variability.","The reported specificity of 96.33% implies few false-positive detections, which matters for screening decisions that should avoid unnecessary follow-up.","The method's applicability to both gastric bleeding and gastric red-spot images suggests the coarse-to-fine design is not tied to one lesion appearance.","Combining fuzzy clustering with a deep decoder may reduce sensitivity to noise and to intensity overlap between lesion and healthy tissue, the problem that motivated the work."],"supporting_citations":[{"why":"Supplies the GLCM texture statistics used to build the DuS-KFCM feature vector in stage one.","marker":"[10]"},{"why":"Supplies the ResNet-50 residual network architecture used as the feature extractor in the DeepLabv3+ refinement stage.","marker":"[33]"},{"why":"Supplies the atrous-convolution semantic segmentation design (DeepLabv3+) used to refine the coarse lesion boundaries.","marker":"[43]"}],"fun_headline_variants":["Hybrid fuzzy-deep model hits 87.95% accuracy on gastric bleeding","Two-stage AI sharpens gastric bleed segmentation to 87.95%","Neuro-fuzzy pipeline beats baselines in gastric bleed detection","Coarse-to-fine fuzzy-deep model excels at subtle gastric bleeds","Dual fuzzy-deep approach lifts gastric bleed accuracy to 87.95%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the public pixel-level annotations are trustworthy and that feeding each image's own fuzzy-clustering mask into the deep refinement stage during training does not leak information or bias the measured improvement; the paper reports no ablation or independent validation that would rule this out.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid fuzzy-deep model hits 87.95% accuracy on gastric bleeding","Two-stage AI sharpens gastric bleed segmentation to 87.95%","Neuro-fuzzy pipeline beats baselines in gastric bleed detection","Coarse-to-fine fuzzy-deep model excels at subtle gastric bleeds","Dual fuzzy-deep approach lifts gastric bleed accuracy to 87.95%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2884,"prompt_tokens":945,"completion_tokens":1939,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1841}},"tokens_in":561,"tokens_out":1939,"duration_ms":11380,"temperature":1.0,"reasoning_tokens":1841,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:13:50.867962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the DeepLabv3+ refinement stage on the original images alone (no DuS-KFCM masks) using the same train/test split, and compare it on a held-out test set whose images were never used to generate clustering masks; if accuracy and specificity do not improve materially, the coarse-to-fine contribution is not supported. A second check is to recompute the clustering masks on the test set independently and verify the refinement stage is not simply reproducing training-set masks.","supporting_citations":[{"cited_title":"M., Shanmugam, K","cited_arxiv_id":null,"evidence_quote":"Supplies the GLCM texture statistics used to build the DuS-KFCM feature vector in stage one."},{"cited_title":"& Sun, J","cited_arxiv_id":null,"evidence_quote":"Supplies the ResNet-50 residual network architecture used as the feature extractor in the DeepLabv3+ refinement stage."}],"review_version":1}