{"id":"038a95d8-6047-406a-b877-20e313553e8a","arxiv_id":"1909.02477","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"AFP-Net, an anchor-free polyp detector with a context enhancement module and cosine ground-truth projection, achieves 99.36% precision and 96.44% recall on CVC-Clinic, and 52.6 FPS.","lead":"This paper presents AFP-Net, a real-time deep learning model that detects colorectal polyps in colonoscopy images without using predefined anchor boxes. The model reports state-of-the-art precision and recall on two public polyp datasets, while running at 52.6 frames per second.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III's 'CVC-Clinic-train' rows have inconsistent ground-truth totals (TP+FN = 646 for AFP-Net/CenterNet/SSD but 531 for Faster-RCNN), so the compared methods were not evaluated on the same test set; the SOTA claim is not supported as written.","rationale":"The paper's method is plausible and its ablation study is internally consistent, but the paper's key claim is comparative: it outperforms all previous approaches. Table III is the only evidence for that claim, and a simple arithmetic check shows the rows cannot all come from the same evaluation: AFP-Net, CenterNet, and SSD-baseline each have TP+FN = 646 polyps on 'CVC-Clinic-train', while Faster-RCNN has TP+FN = 531. Since TP+FN is the total number of ground-truth boxes in the test set, the sets differ. The text also alternates between 'all 612 images are used as our testing set' (Section IV.A) and 'CVC-Clinic training dataset' (Section IV.E), and no image IDs are given. The loose centroid-based TP rule (Section IV.C) further makes absolute scores hard to interpret. Because the central comparative claim lacks a common evaluation basis as printed, the current verdict should be REJECT; a corrected re-evaluation on a shared split with an identical metric would be needed to restore it.","tokens_in":12189,"tokens_out":13134,"duration_ms":138721,"concrete_test":"Obtain from the authors the exact image IDs and ground-truth annotation counts for the evaluation set called 'CVC-Clinic-train', along with trained weights for AFP-Net, SSD-baseline, CenterNet, and Faster-RCNN [26]. Write a single evaluation script implementing the paper's centroid-in-ground-truth TP rule, run all four models on that identical image set, and compare TP+FN across methods; if the totals are not equal, the test sets differ. Then compute precision, recall, F1, and F2 for each model on the common set. This check settles the concern: if AFP-Net no longer has the highest F1 and F2, the central claim fails; if it still does, the claim can be restored after relabeling the table.","verdict_should_be":"REJECT","load_bearing_attack":"Section IV.E makes the central claim that AFP-Net outperforms all previous approaches in F1/F2 on both testing datasets, with Table III as the only support. The table is internally inconsistent: on the rows labeled 'CVC-Clinic-train', AFP-Net (TP=623, FN=23), CenterNet (TP=603, FN=43), and SSD-baseline (TP=618, FN=28) each give TP+FN = 646, whereas Faster-RCNN [26] gives TP=523, FN=8, i.e., TP+FN = 531. Under the paper's own evaluation definition, FN is the number of ground-truth polyps not detected and each GT counts once, so TP+FN is the total number of ground-truth boxes in the test set. That total cannot depend on the method. The 115-instance gap means Faster-RCNN's numbers were produced on a different image set, a different set of ground-truth annotations, or under a different counting rule. Section IV.A says all 612 CVC-Clinic images are used as the testing set, while Section IV.E says the test was run on 'CVC-Clinic training dataset', and no image IDs or split files are provided. The TP definition in Section IV.C (centroid inside ground truth) is also unusually loose and ignores localization quality, which further obscures the comparison. With no common evaluation basis and no error bars, the claimed F1/F2 advantage over the closest baselines (0.71 points over SSD-baseline; 1.33 over FRCNNPL) is not established. The central state-of-the-art claim therefore fails as printed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AFP-Net, an anchor-free polyp detector for colonoscopy. The architecture uses a VGG16 backbone, a feature pyramid, a Context Enhancement Module (CEM), a cosine ground-truth projection for multi-scale label assignment, focal loss with a Gaussian penalty, and anchor-free box regression. The authors report state-of-the-art precision/recall/F1/F2 on CVC-Clinic and ETIS-LARIB test sets, with 52.6 FPS inference speed, and support the claims with an ablation study and comparisons against anchor-based and anchor-free baselines.","tokens_in":12594,"tokens_out":3934,"duration_ms":42496,"significance":"If the reported results are reproducible, the paper makes a useful contribution: it is among the first anchor-free detectors for polyp detection, provides detailed ablations of each component, and directly compares with a matched SSD-style anchor-based baseline, showing a real-time speed advantage. The systematic ablation (Table I) and the inclusion of inference-time measurements are strengths. However, the central state-of-the-art claim depends entirely on the comparison in Table III, and that table has internal inconsistencies and ambiguous test-set descriptions. The evaluation protocol, especially the centroid-only true-positive criterion and the absence of error bars, is too weak to support the claimed margins over the closest baselines as currently written.","major_comments":[{"comment":"The rows labeled 'CVC-Clinic-train' in Table III are internally inconsistent. Using the paper's own definitions in Section IV.C, each ground-truth box is either a true positive or a false negative, so TP+FN equals the total number of ground-truth boxes in the test set and cannot depend on the method. The table reports TP+FN = 646 for AFP-Net (623+23), CenterNet (603+43), and SSD-baseline (618+28), but TP+FN = 531 for Faster-RCNN (523+8). The 115-instance gap means Faster-RCNN's numbers were produced on a different image set, a different annotation set, or under a different counting rule. Since the paper's central claim that AFP-Net 'outperforms all previous approaches' is supported only by this table, the claim is not established as printed.","section":"Section IV.E / Table III"},{"comment":"The test-set description is ambiguous and potentially contradictory. Section IV.A states that all 612 CVC-Clinic images are used as the testing set, while Section IV.E says the test was run on 'CVC-Clinic training dataset' and Table III uses the column header 'CVC-Clinic-train'. Because the CVC-Clinic images are also said to be identical to the GIANA polyp-segmentation test set, and no image IDs or split files are provided, a reader cannot determine whether the compared methods were trained on the same images, whether there is overlap between training and testing, or whether the numbers in Table III were all obtained under the same protocol. The authors should specify the exact split, release or cite image IDs, and ensure that the evaluation set is clearly labeled and consistently used for all methods.","section":"Section IV.A / Section IV.E / Table III"},{"comment":"The true-positive criterion is too loose to support the precision/recall comparisons. Section IV.C defines a TP as any detection whose centroid falls inside a ground-truth box, with no IoU threshold, and counts only one TP per ground truth even if multiple boxes overlap. This criterion does not penalize poor localization, and combined with the low NMS threshold of 0.1, it can inflate precision and recall relative to standard detection metrics. The differences against the closest baselines are small (for example, F1 97.88 vs. 97.17 for SSD-baseline on CVC-Clinic-train, and F1 84.63 vs. 83.3 for FRCNNPL on ETIS-LARIB), so a stricter metric or at least error bars across multiple runs is needed to establish that the improvement is meaningful. No error bars, variance estimates, or multiple-seed experiments are reported anywhere.","section":"Section IV.C / Table III"},{"comment":"The comparison with prior work mixes numbers obtained under different conditions. Some rows are taken from previously published papers on datasets that are no longer available (e.g., Y-Net and RYCO on ASU-MAYO*), some report only N/A for TP/FP/FN, and the hardware and input sizes differ across methods. It is not stated whether the prior methods were re-run under the same protocol as AFP-Net or whether the published numbers are quoted as-is. This weakens the 'outperforms all previous approaches' conclusion, especially because the reader cannot verify whether the compared methods used the same test split and the same true-positive definition.","section":"Section IV.E / Table III"}],"minor_comments":[{"comment":"The text says 'we show the ROC curve with respect to precision and recall,' but Figure 5 is a precision-recall curve; the wording should be corrected.","section":"Section IV.D / Figure 5"},{"comment":"There is a missing period in 'Our network uses VGG16 [34] as the backbone Our framework selects k = 6 feature maps'; the sentence break should be fixed.","section":"Section III.A"},{"comment":"The phrase 'we have m2 ... center points' is typeset ambiguously; it should be 'm^2 center points' or 'm-by-m center points' for clarity.","section":"Section III.C"},{"comment":"Section IV.A describes GIANA as a MICCAI2017 challenge while Section IV.C refers to the 'MICCAI2015 [5] challenge' protocol; the distinction between the challenge year and the cited protocol should be clarified to avoid confusion.","section":"Section IV.A / Section IV.C"},{"comment":"No code, trained models, or split files are provided, and the manuscript does not state whether they will be released. Given the ambiguities in the evaluation, releasing the exact test split and the evaluation script would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency in Table III (TP+FN totals differing by method) is a serious correctness issue for the main claim, but it is fixable if the authors re-evaluate the compared methods on a single, clearly defined test set and report consistent metrics. I therefore do not recommend rejection, but the revision must address the evaluation protocol before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies anchor-free detection to polyp detection, with two real innovations: a cosine ground-truth projection that distributes non-negative regions across feature-scale levels, and a context enhancement module using dilated convolutions. The ablation study is actually careful: each component is turned on/off and the precision/recall trade-offs are reported. That is the best part of the paper.\n\nBut the central claim, stated in Section IV.E, that AFP-Net outperforms all previous approaches in F1/F2 on both test sets, is not supported by Table III as printed. The rows labeled 'CVC-Clinic-train' show TP+FN = 646 for AFP-Net, CenterNet, and SSD-baseline, but 531 for Faster-RCNN. Under the paper's own definition, FN is the number of ground-truth polyps not detected, so TP+FN is the total ground-truth count. That total cannot depend on the detector. The Faster-RCNN numbers were clearly produced on a different image set or with a different counting rule. Once you see that, the comparative table loses its basis. There is also the ambiguity that Section IV.E says the test is on 'CVC-Clinic training dataset' while Section IV.A says all 612 CVC-Clinic images are used as testing. And the evaluation metric is generous: a true positive is a predicted box whose centroid falls anywhere inside the ground truth, which ignores localization quality. No error bars or multiple runs are reported, so the 0.7-point F1 edge over SSD-baseline is not statistically grounded.\n\nThe methodology itself seems coherent. The label assignment and the cosine projection are described clearly, and the loss definitions are explicit. I don't see an issue with the math, and the hyperparameters are stated. But the paper ships no code or weights, so the experiments are not independently testable.\n\nWhere does this leave us? If the evaluation can be cleaned up and re-run transparently, the approach might well be competitive. As written, the SOTA claim fails. I would not desk-reject it: the engineering idea has merit and the ablation is a useful reference for anyone working on anchor-free detectors in medical imaging. But it needs a major revision, not a tweak. The authors need to release the exact test split, fix the count inconsistency, and report results with standard IoU thresholds and error bars.\n\nRecommendation: send to peer review, with the expectation of substantial revision. It's not a waste of referee time, but the referee should not accept the numbers at face value.","headline":"Plausible anchor-free polyp detector with a solid ablation study, but the evaluation table is internally inconsistent and the SOTA claim doesn't hold as printed.","tokens_in":13081,"tokens_out":2203,"would_cite":false,"duration_ms":22631,"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":"The paper claims that an anchor-free, center-point detector with a cosine ground-truth projection can detect colon polyps at 99.36% precision and 96.44% recall while running at 52.6 frames per second.","keywords":["anchor-free detection","polyp detection","colonoscopy","center-point detection","cosine ground-truth projection","context enhancement module","feature pyramid network","focal loss"],"falsifier":"Re-run AFP-Net and the compared detectors on a colonoscopy test set that none of them trained on, count a detection as correct only when its predicted box overlaps the ground truth by at least 0.5 IoU, and report scores across multiple train/test splits; if AFP-Net no longer leads in F1 and F2, or the gap falls within run-to-run variation, the state-of-the-art claim would be refuted.","tokens_in":12030,"feed_emoji":"🩺","tokens_out":8141,"duration_ms":76773,"temperature":0.7,"pith_summary":"The paper proposes AFP-Net, a detector that finds colorectal polyps in colonoscopy video without using predefined anchor boxes: it represents each polyp as a center point on a feature map and regresses its box from there. The authors claim that this anchor-free design reaches 99.36% precision and 96.44% recall on the CVC-Clinic test frames, 88.89% precision and 80.77% recall on ETIS-LARIB, and runs at 52.6 frames per second, beating earlier anchor-based and anchor-free detectors on the F1 and F2 scores. The main technical move is a cosine ground-truth projection that spreads each polyp's training labels across neighboring feature-pyramid levels, which they argue recovers the recall that removing anchors would otherwise cost. The motivation is clinical: automated real-time detection could reduce the roughly 25% miss rate of polyps during colonoscopy.","feed_headline":"Real-time anchor-free detector hits 99.36% polyp precision","feed_subtitle":"A center-point detector drops anchor boxes and still beats anchor-based rivals in F1 and F2 scores.","key_machinery":"The load-bearing object is the cosine ground-truth projection: for a ground truth assigned to its best feature level $i$, the non-negative label region at another level $l$ is scaled by $\\phi_l=\\max(\\cos(\\lambda d\\pi/2k),0)$, where $d=|i-l|$, $k$ is the number of feature levels, and $\\lambda=2.5$. This factor decides which center points are positive, ignored, or negative at each feature-pyramid level, so that neighboring scales still supervise keypoints near the object center instead of labeling them negative. Around it sits the Context Enhancement Module, which splits channels into three branches of dilated convolutions and concatenates them to enlarge the receptive field, and the multitask loss that combines Smooth L1 box regression, focal loss on negative samples, and Gaussian-weighted cross-entropy on positive samples.","core_discovery":"On its own terms, the paper's central discovery is that an anchor-free detector can match or beat anchor-based polyp detectors if the label assignment is redesigned rather than copied. Removing anchors eliminates the need for IoU thresholds and anchor hyperparameters, but it also creates 'unnecessary negative' keypoints: near-center points on non-best feature levels are labeled negative even though the enlarged receptive field could cover the object. AFP-Net answers this with a cosine ground-truth projection, which shrinks the non-negative region on each neighboring feature level by a cosine factor based on distance from the best level, turning those points into useful training signals. Combined with a Context Enhancement Module that widens the receptive field and a loss using focal loss on negatives and a Gaussian distance penalty on positives, the model reaches the reported precision/recall numbers and runs in real time.","pith_inferences":["The same cosine label-spreading trick could transfer to other sparse-object detection tasks in medical imaging, such as lesions in CT or capsule endoscopy, where small objects and class imbalance make anchor-free recall loss especially costly.","The reported true-positive rule -- predicted center inside the ground-truth box -- is more permissive than an IoU threshold like 0.5; re-scoring with standard IoU would likely lower absolute numbers and might change the ranking against anchor-based methods.","The test column labeled 'CVC-Clinic-train' raises the possibility that compared models trained on the same frames; a held-out split with no overlap could narrow or reverse the claimed advantage.","Since the detector already runs at 52.6 FPS, a natural extension is to couple it with temporal tracking across video frames, which could suppress the fold-shaped false positives shown in the paper's failure cases."],"forward_implications":["Anchor-based detectors are not required for accurate polyp detection: AFP-Net's center-point head outperforms its direct anchor-based counterpart, SSD-baseline, which shares the same backbone, pyramid, and context module.","Cosine ground-truth projection is what recovers recall after anchors are removed: dropping it (Experiment 4 vs 6) lowers recall by a considerable margin in the ablation.","Context enrichment is a necessary ingredient: the paper reports that removing the Context Enhancement Module degrades both precision and recall.","The full pipeline operates at 52.6 FPS on an RTX 2080 Ti, satisfying real-time colonoscopy video requirements while keeping the reported accuracy.","Because no anchors or IoU assignment thresholds need tuning, the detector can be adapted to new endoscopic datasets without per-dataset anchor design."],"supporting_citations":[{"why":"Supplies the center-keypoint anchor-free formulation that AFP-Net adapts.","marker":"[44]"},{"why":"Motivates context enrichment for anchor-free detectors and the use of corner pooling.","marker":"[18]"},{"why":"Provides the top-down multi-scale path used to enrich semantic features.","marker":"[22]"},{"why":"Provides the single-stage baseline architecture, the feature-map selection, and the anchor-based counterpart for speed comparison.","marker":"[25]"},{"why":"Offers the classification loss suited to the sparse positive/negative ratio in polyp images.","marker":"[24]"},{"why":"Defines the evaluation protocol (precision, recall, F1, F2) and the challenge datasets used for training and testing.","marker":"[5]"},{"why":"Supplies the CVC-Clinic test frames with pixel-level ground truth.","marker":"[4]"},{"why":"Supplies the second test set of high-resolution polyp images.","marker":"[31]"},{"why":"Provides the anchor-based Faster R-CNN baseline that AFP-Net is compared against on CVC-Clinic.","marker":"[26]"},{"why":"Provides the post-learning Faster R-CNN baseline used for the ETIS-LARIB comparison.","marker":"[32]"}],"fun_headline_variants":["Anchor-free polyp detection: 99.36% precision in real time","AFP-Net: real-time polyp spotting without anchor boxes","Cosine trick boosts anchor-free polyp detection to 99%","Real-time colonoscopy AI finds polyps with 99% precision","Detecting polyps in real time: anchor-free wins with cosine design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported evaluation is a fair, unbiased measure, yet the test set is labeled 'CVC-Clinic-train,' the true-positive rule is only that the predicted center falls inside the ground-truth box, and no error bars or split variance are reported.","fun_headline_variants_meta":{"raw":{"variants":["Anchor-free polyp detection: 99.36% precision in real time","AFP-Net: real-time polyp spotting without anchor boxes","Cosine trick boosts anchor-free polyp detection to 99%","Real-time colonoscopy AI finds polyps with 99% precision","Detecting polyps in real time: anchor-free wins with cosine design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1242,"prompt_tokens":949,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":204}},"tokens_in":565,"tokens_out":293,"duration_ms":3066,"temperature":1.0,"reasoning_tokens":204,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:48:31.178087+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run AFP-Net and the compared detectors on a colonoscopy test set that none of them trained on, count a detection as correct only when its predicted box overlaps the ground truth by at least 0.5 IoU, and report scores across multiple train/test splits; if AFP-Net no longer leads in F1 and F2, or the gap falls within run-to-run variation, the state-of-the-art claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the second test set of high-resolution polyp images."},{"cited_title":"Bernal, N","cited_arxiv_id":null,"evidence_quote":"Defines the evaluation protocol (precision, recall, F1, F2) and the challenge datasets used for training and testing."},{"cited_title":"Bernal, F","cited_arxiv_id":null,"evidence_quote":"Supplies the CVC-Clinic test frames with pixel-level ground truth."},{"cited_title":"An Efficient Approach for Polyps Detection in Endoscopic Videos Based on Faster R-CNN","cited_arxiv_id":"1809.01263","evidence_quote":"Provides the anchor-based Faster R-CNN baseline that AFP-Net is compared against on CVC-Clinic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the post-learning Faster R-CNN baseline used for the ETIS-LARIB comparison."}],"review_version":1}