{"id":"5994a986-0a33-4d3f-a9cd-00afabcf0ea6","arxiv_id":"2412.01372","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A YOLOv5 variant augmented with Swin-Transformer, GAM attention, multi-scale fusion, and EIoU loss reports 92.6% mAP@0.5 on p16/Ki-67 dual-stained cervical cytology patches, a claimed improvement over baseline YOLOv5s.","lead":"This paper builds a cervical cancer screening model called DSIR-YOLO by adding attention, transformer, and multi-scale fusion modules to the YOLOv5 object detector, and reports higher detection accuracy on p16/Ki-67 dual-stained cell images. It also shows that cleaning up annotation boxes substantially improves the model, but the results come from a small private dataset with no code or data released.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patch-level train/validation splits across only 11 whole-slide images allow slide-level leakage, so the reported accuracy gains and p-values may not reflect generalization to unseen slides.","rationale":"The reader's weakest assumption identifies exactly the load-bearing threat to the central claim: random patch-level splitting of only 11 whole-slide images cannot support claims of improved detection on new cervical cytology slides. This is not merely a hypothetical concern; medical image patches from the same whole-slide image are strongly correlated, so validation patches have near-duplicates in training. That correlation inflates both absolute performance and the apparent statistical significance of the comparison against YOLOv5s. The paper's own description in Section 2.1 and Section 3.3.4 supports this reading. I agree with the reader that this is the most important unresolved issue. The architecture itself is plausible, the ablation results are internally coherent, and the annotation-quality study is a useful practical observation, but none of these compensate for the lack of slide-level generalization evidence. The reader's conditional verdict is therefore appropriate: accept only if the evaluation protocol is corrected or explicitly reframed as patch-level performance without generalization claims. No additional objection is needed beyond the slide-leakage concern, because all other issues—inconsistent baseline mAP values, the final ablation row, and missing code/data—are secondary or stem from the same evaluation weakness.","tokens_in":9652,"tokens_out":2127,"duration_ms":21161,"concrete_test":"Obtain or reconstruct the source-slide identifier for every patch and rerun the comparison with slide-level separation: e.g., leave-one-slide-out cross-validation (train on patches from 10 whole-slide images, test on all patches from the remaining slide) or a five-fold cross-validation in which all patches from a given slide are assigned to the same fold. If DSIR-YOLO's margin over YOLOv5s shrinks, reverses, or loses statistical significance under this split, the reported patch-level improvements are attributable to slide-level leakage rather than to the architectural modifications. Report per-slide mAP values and confidence intervals, not only fold-level means.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that DSIR-YOLO significantly outperforms YOLOv5s on p16/Ki-67 dual-stained cervical cytology images, with the five-fold cross-validation shown in Table 5 and p-values presented as statistical support. Section 2.1 states that 7,662 patches were derived from 11 whole-slide images and randomly split at an 8:2 ratio; Section 3.3.4 describes a random five-fold partition of the patches. Neither procedure separates patches by source slide or patient. Because patches from the same whole-slide image share staining conditions, background texture, cellular populations, and annotation noise, random patch-level partitioning creates near-duplicate content between training and validation folds. A model can exploit slide-specific correlations rather than learn generalizable cell morphology, inflating mAP@0.5, mAP@0.5:0.95, and the apparent stability of results. The p-values in Table 5 are computed from only five paired fold-level observations and do not correct for clustering by slide; they therefore do not establish that the gains transfer to new slides or new patients. The reported improvements over YOLOv5s may be real on patch-level holdouts, but the paper does not demonstrate the stronger claim implied by the abstract and conclusion that the algorithm improves detection on unseen dual-stained cervical cell images.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DSIR-YOLO, a modified YOLOv5 model for detecting p16/Ki-67 dual-stained positive cells in cervical cytology images. The modifications are a Swin-Transformer backbone, GAM attention, BiFPN-style multi-scale feature fusion, and EIoU loss. Using a private dataset of 7,662 patches derived from 11 whole-slide images, the authors compare DSIR-YOLO with YOLOv5 variants and other detectors, report ablation results, and analyze how four annotation-quality refinements affect detection. They report mAP@0.5 of 92.6% and mAP@0.5:0.95 of 70.5% and claim consistent gains over YOLOv5s in five-fold cross-validation.","tokens_in":9928,"tokens_out":5142,"duration_ms":42815,"significance":"If the results are valid, the paper offers a practically useful detector for cervical cancer screening and a useful demonstration that annotation quality substantially affects object-detection performance in cytology. The strengths are the systematic study of annotation quality (pixel sealing, scale discrepancies, unlabeled cells, and large-angle clusters), the comparison with several YOLO and non-YOLO baselines, and the explicit attempt at cross-validation. However, the central comparison rests on patch-level splits of only 11 slides, which risks slide-level leakage, and several reported numbers are internally inconsistent. The architectural novelty is modest, but the empirical contribution could be valuable for a clinical engineering audience if these issues are resolved.","major_comments":[{"comment":"The 7,662 patches are randomly split at an 8:2 ratio and randomly partitioned into five folds without any separation by source whole-slide image or patient. Because only 11 whole-slide images are used, patches from the same slide can appear in both training and validation, sharing staining conditions, background texture, and annotation noise. This near-duplicate content can inflate mAP values and the p-values in Table 5, so the reported results do not yet establish generalization to unseen slides or patients. Please report slide-level or patient-level cross-validation, or explicitly restrict the claims to patch-level recognition within the same slides.","section":"Section 2.1 and Section 3.3.4"},{"comment":"The mean row for YOLOv5s and for 'This Study' in Table 5 is identical to the Fold1 row, while the other folds differ, which is almost certainly a copy-paste error. In addition, the YOLOv5s mAP@0.5 mean in Table 5 (88.0%) contradicts the YOLOv5s mAP@0.5 in Table 2 (84.9%), and the abstract's 'reaching 92.6% and 70.5%' corresponds to the single split in Table 2 while the subsequent 'increased by 2.3%, 4.1%, 4.3%, and 8.0%' corresponds to the cross-validation means in Table 5. The manuscript must state which evaluation is primary and reconcile these numbers, because the central claim depends on a consistent baseline.","section":"Table 5 and Table 2"},{"comment":"The final ablation row combines all modules with 'Data Augmentation,' but Section 2.2 states that Mosaic augmentation, MixUp augmentation, and random horizontal flipping were already used to expand the training dataset. The paper does not define what 'Data Augmentation' adds in this ablation, and no no-augmentation baseline is provided. Consequently, the marginal contribution of the architectural modules is not cleanly isolated, and the attribution of the final gains to the proposed modules is undermined. Please specify the augmentation protocol for every row and hold augmentation fixed across the ablation comparisons.","section":"Section 3.3.2, Table 3"},{"comment":"The p-values are reported for five folds, but no statistical test is named, no test statistic is given, and the effective sample size is not discussed. On five paired folds, most nonparametric tests cannot reach p<0.006, and the p-values do not account for clustering by source slide. Please state the test, report exact p-values, and provide a justification for the sample size and independence assumptions.","section":"Section 3.3.4, Table 5"}],"minor_comments":[{"comment":"Equations (1), (2), and (3) are rendered with placeholder '??' characters and are not readable; they need proper typesetting and complete definitions of all symbols.","section":"Equations (1)-(3)"},{"comment":"The text refers to 'Table 6-3' and 'Figure 6-1,' which should be 'Table 3' and 'Figure 11,' respectively.","section":"Section 3.3.2 and Section 3.4"},{"comment":"The abstract contains 'This paper Proposes' with an uppercase 'P,' and the model acronym DSIR is not defined anywhere in the text; please define it and proofread for capitalization and grammar.","section":"Abstract and Section 1"},{"comment":"The narrative sentence 'Finally, adding the GAM module...' does not match the table structure, because the all-four-modules row already includes GAM and the final row adds data augmentation; please align the description with the rows in Table 3.","section":"Section 3.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a domain application with a private dataset and no public code or benchmark. The main risk is that the evaluation protocol, rather than the architecture, drives the reported gains; the patch-level leakage needs to be addressed directly. The report to the authors should require slide-level or patient-level cross-validation, reconciliation of Tables 2 and 5, and a clarified ablation protocol before the claims about clinical generalization can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a routine architecture-assembly paper with a genuinely useful dataset-quality study bolted on, and the evaluation doesn't support the headline generalization claim.\n\nWhat's new: applying Swin-Transformer + GAM + BiFPN-style fusion + EIoU to p16/Ki-67 dual-stained cytology is a new combination, but each piece is published and the assembly is exactly what you'd expect. The annotation-quality analysis is the real contribution: they show that tightening bounding boxes, splitting large clusters, labeling isolated positives, and fixing diagonal boxes lifts YOLOv5s mAP@0.5:0.95 from 38.3% to 68.8% on their data. That's a practical result for anyone building cytology datasets.\n\nWhat's solid: the experiments are fairly extensive for a private dataset. They run ablations, compare with YOLOv5 variants, YOLOX, YOLOv8, Faster R-CNN, and SSD, and do five-fold CV. The reported differences between DSIR-YOLO and YOLOv5s are consistent across folds in Table 5.\n\nSoft spots: the biggest problem is the patch-level split. 7,662 patches come from only 11 whole-slide images, and the split is random at patch level. Patches from the same slide share staining, background, and annotation noise, so train and validation contain near-duplicates. That leaks slide-specific information and inflates all the metrics. The p-values in Table 5 are computed from five paired folds but don't account for clustering by slide, so they're not evidence of generalization to new slides or patients.\n\nThere's also an unexplained inconsistency: YOLOv5s mAP@0.5 is 84.9% in Table 2 and 88.0% in the five-fold mean in Table 5. Same model, same dataset, different numbers. The ablation table's last row adds data augmentation on top of all architectural changes, so the final gain mixes augmentation with architecture. And there's no code or data, so the numbers can't be checked.\n\nBottom line: the paper is a competent applied effort, the dataset-quality observations are worth publishing, but the central claim—that DSIR-YOLO significantly outperforms YOLOv5s on unseen dual-stained images—is not demonstrated. Who's it for: applied researchers who want a baseline recipe and a reminder that annotation quality matters more than architecture tricks. I'd send it to peer review because the application and the dataset study are legitimate, but the authors need to release data/code and redo the evaluation with slide-level splits before I'd trust the numbers.","headline":"Routine architecture stacking with a useful annotation-quality study; the evaluation leaks slide information, so the headline gains are likely overstated.","tokens_in":10465,"tokens_out":1992,"would_cite":false,"duration_ms":17580,"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":"An improved YOLOv5 detector, DSIR-YOLO, recognizes p16/Ki-67 dual-positive cervical cells at 92.6% mAP@0.5 and 70.5% mAP@0.5:0.95, and the paper reports that cleaning annotation boxes contributes more to the gain than the architectural…","keywords":["cervical cancer screening","p16/Ki-67 dual staining","immunocytochemistry","YOLOv5 object detection","Swin-Transformer","global attention mechanism","EIoU loss","dataset annotation quality"],"falsifier":"Run the same five-fold protocol but split patches by whole-slide image, so every patch of a given slide goes into exactly one fold, and retrain DSIR-YOLO and YOLOv5s; if the mAP@0.5 gap of 4.3 points and the mAP@0.5:0.95 gap of 8.0 points shrink substantially or reverse, the claimed architectural advantage is an artifact of patch-level leakage rather than generalization.","tokens_in":9429,"feed_emoji":"🔬","tokens_out":12048,"duration_ms":91023,"temperature":0.7,"pith_summary":"This paper tries to establish that a specific modification of the YOLOv5 object detector, called DSIR-YOLO, can reliably find cells that show both p16 and Ki-67 staining in cervical cytology images, and that careful cleanup of training annotations contributes as much to performance as the network changes. The authors replace the YOLOv5s backbone with a Swin-Transformer, add a global attention mechanism, multi-scale feature fusion with spatial pyramid pooling, and a Focal-EIoU loss, reporting 92.6% mean average precision at a 0.5 overlap threshold and 70.5% across thresholds 0.5 to 0.95 on their dual-stained dataset. They also report that four annotation-quality fixes alone lift unmodified YOLOv5s by large margins, from 72.4% to 85.7% precision, 69.0% to 84.3% recall, and 38.3% to 68.8% in mAP@0.5:0.95. If these results hold on new slides, they would support computer-assisted reading of p16/Ki-67 dual stains for cervical cancer screening where pathologists are scarce.","feed_headline":"92.6% mAP: improved YOLO reads cervical dual-stained cells","feed_subtitle":"Swin-Transformer plus attention plus cleaner annotations lift YOLOv5s by 4.3 and 8.0 mAP points in five-fold tests","key_machinery":"The carrying object is the assembled DSIR-YOLO detector: a lightweight YOLOv5 architecture whose backbone is replaced by a Swin-Transformer module, a visual transformer built from windowed and shifted-window self-attention; augmented with the GAM attention mechanism to preserve spatial and channel interactions; a BiFPN-like multi-scale feature fusion path with spatial pyramid pooling after downsampling; and the Focal-EIoU loss for bounding-box regression. The argument is carried by ablation experiments that add each component to the baseline and show stepwise gains, with the full configuration reaching 91.5% mAP@0.5 and 69.1% mAP@0.5:0.95 before data augmentation and 92.6% and 70.5% with augmentation.","core_discovery":"The central claim is that the DSIR-YOLO architecture, an enhanced YOLOv5 with a Swin-Transformer backbone, GAM attention, multi-scale feature fusion with spatial pyramid pooling, and Focal-EIoU loss, detects p16/Ki-67 dual-positive cervical cells better than the original and better than several comparators on this dataset. In five-fold cross-validation the improved model's mean precision, recall, mAP@0.5, and mAP@0.5:0.95 are 86.9%, 84.6%, 92.3%, and 69.2%, against 84.6%, 80.5%, 88.0%, and 61.2% for YOLOv5s, with smaller variances and $p<0.006$. The paper also claims that dataset quality is a major lever: after four annotation fixes, unmodified YOLOv5s improves by 13.3 percentage points in precision, 15.3 in recall, 18.3 in mAP@0.5, and 30.5 in mAP@0.5:0.95.","pith_inferences":["Editorial extension: The reported gains are measured on a random patch split, and because the 7,662 patches come from only 11 whole-slide images, a slide-stratified split would reveal how much of the gain is genuine generalization to new slides rather than memorization of slide-specific appearance.","Editorial extension: The large annotation-quality effect suggests a transferable protocol for small-object cytology detection: compare models only on tightly audited labels, because label noise can swamp architectural improvements; this protocol could be tested on other liquid-based cytology datasets.","Editorial extension: The paper notes that the model sacrifices speed for accuracy, so quantifying inference time per whole slide or per patch would let a clinic decide whether DSIR-YOLO can run at screening throughput, a question the paper leaves open."],"forward_implications":["A screening tool could use DSIR-YOLO to flag dual-positive cells automatically, with the reported 92.3% mean mAP@0.5 across folds reducing the manual search burden on pathologists.","Standardizing annotation boxes, by tightening them, splitting clusters, labelling small positive cells, and avoiding oversized boxes on angled clusters, is itself a performance lever, since the paper shows it lifts YOLOv5s by double-digit percentage points before any architectural change.","The architectural improvements concentrate on small-cell detection and false-positive suppression around clusters, which are precisely the error modes that matter in p16/Ki-67 dual-stain reading.","The improved model shows smaller cross-validation variance than YOLOv5s, so a deployed system should be more consistent when retrained on different clinic samples, provided the split measures slide-level generalization."],"supporting_citations":[{"why":"Provides the Swin-Transformer-based visual architecture the paper adopts as its backbone.","marker":"[4]"},{"why":"Defines the Global Attention Mechanism used to retain channel-spatial interactions.","marker":"[5]"},{"why":"Supplies the multi-scale feature fusion concept the paper adapts into a BiFPN-like neck.","marker":"[6]"},{"why":"Introduces spatial pyramid pooling, which the paper inserts after downsampling to enlarge the receptive field.","marker":"[8]"},{"why":"Defines the Focal-EIoU loss the paper substitutes for CIoU in bounding-box regression.","marker":"[9]"},{"why":"Shows that labeling instructions affect biomedical image analysis outcomes, motivating the dataset-quality experiments.","marker":"[11]"}],"fun_headline_variants":["DSIR-YOLO: 92.6% mAP for cervical dual-stain","Swin-Transformer sharpens cervical dual-stain detection","Data cleanup adds 30.5 mAP to cervical screening","Attention boosts YOLO cervical cell recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that randomly splitting 7,662 image patches from only 11 whole-slide images into training and validation sets measures how the model will perform on new slides, even though patches from the same slide can appear on both sides of the split.","fun_headline_variants_meta":{"raw":{"variants":["DSIR-YOLO: 92.6% mAP for cervical dual-stain","Swin-Transformer sharpens cervical dual-stain detection","Data cleanup adds 30.5 mAP to cervical screening","Attention boosts YOLO cervical cell recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001569,"raw_usage":{"total_tokens":6347,"prompt_tokens":1110,"completion_tokens":5237,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":5165}},"tokens_in":726,"tokens_out":5237,"duration_ms":32273,"temperature":1.0,"reasoning_tokens":5165,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:24:50.493549+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five-fold protocol but split patches by whole-slide image, so every patch of a given slide goes into exactly one fold, and retrain DSIR-YOLO and YOLOv5s; if the mAP@0.5 gap of 4.3 points and the mAP@0.5:0.95 gap of 8.0 points shrink substantially or reverse, the claimed architectural advantage is an artifact of patch-level leakage rather than generalization.","supporting_citations":[{"cited_title":"Cell-Vision Fusion: A Swin Transformer-based Approach to Predicting Kinase Inhibitor Mechanism of Action from Cell Painting Data[J]","cited_arxiv_id":null,"evidence_quote":"Provides the Swin-Transformer-based visual architecture the paper adopts as its backbone."},{"cited_title":"Revisiting Multi-Scale Feature Fusion for Semantic Segmentation","cited_arxiv_id":"2203.12683","evidence_quote":"Supplies the multi-scale feature fusion concept the paper adapts into a BiFPN-like neck."},{"cited_title":"Spatial pyramid pooling in deep convolutional networks for visual recognition [J]","cited_arxiv_id":null,"evidence_quote":"Introduces spatial pyramid pooling, which the paper inserts after downsampling to enlarge the receptive field."},{"cited_title":"Focal and efficient IOU loss for accurate bounding box regression [J]","cited_arxiv_id":null,"evidence_quote":"Defines the Focal-EIoU loss the paper substitutes for CIoU in bounding-box regression."},{"cited_title":"Labelling instructions matter in biomedical image analysis [J]","cited_arxiv_id":null,"evidence_quote":"Shows that labeling instructions affect biomedical image analysis outcomes, motivating the dataset-quality experiments."}],"review_version":1}