{"id":"d1af9cb1-5b1c-42ac-aed0-d3d32e7d0416","arxiv_id":"2502.00133","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A YOLOv8n model pre-trained on a fruit and vegetable detection dataset achieved 95.1% F1 on polyp detection, outperforming COCO-pre-trained and scratch-trained models.","lead":"This paper compares seven pre-training datasets, from brain MRI to fruit and vegetables, and finds that pre-training a YOLOv8 nano model before fine-tuning on polyp images beats training from scratch. The best detector came from the fruit-and-vegetable pretrained model, suggesting visual similarity matters more than domain label.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 80/10/10 split in §2.2 is image-level despite §3.4 acknowledging multiple images per polyp, so the test set may leak near-duplicate frames and inflate the F1 rankings; Table 2 additionally contradicts the 'all epochs' claim.","rationale":"The reader identified the same leakage as the weakest assumption, and I agree that is the most load-bearing issue. It is load-bearing because nearly every headline result — fruit&veg-XL as best, XL better than smaller, pre-training beats scratch — is expressed as point estimates from one split. The paper itself flags the multiplicity in Section 3.4, so this is not speculation. My additional check of Table 2 shows the 'any pre-training beats none' phrasing is already false within the paper's own numbers, which strengthens the need for a corrected claim. I do not think this merits rejection: the broad trend that large, diverse pre-training helps over scratch is plausible and partly supported (for example, all XL models exceed scratch at 50 and 100 epochs). The issues are addressable with a proper stratified split, multiple seeds, and error bars, which is exactly what a conditional acceptance should require. Hence the reader's CONDITIONAL verdict remains appropriate and unchanged.","tokens_in":9894,"tokens_out":4919,"duration_ms":46865,"concrete_test":"Re-run the full fine-tuning protocol with a group-stratified split: before the 80/10/10 split, group images by polyp identity (or at least by source video) and assign entire groups to one fold; then compare fruit&veg-XL, HAM10000-XL, COCO, and scratch at 100 epochs. If fruit&veg-XL is no longer top on F1, or if the gaps shrink below about 1 point, the paper's ranking claim is unsupported and should be withdrawn; also recompute Table 2 at 50 epochs to resolve the 'consistently outperform' contradiction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central ranking claim (Section 3.2, Tables 4–6) rests on a single random 80/10/10 split of the combined polyp data. Section 3.4 explicitly states that the fine-tuning set 'consists of multiple images of the same polyp from different perspectives.' If that split is not stratified by polyp or video, near-duplicate frames of a polyp will appear in both training and test, making the reported F1/mAP optimistic and, worse, differentially so: a model that memorizes appearance quirks can look better than one that generalizes. At 100 epochs the top F1 differences are small (fruit&veg-XL 95.1 vs. COCO 94.5 vs. HAM10000-XL 94.4 vs. scratch 92.9) and no error bars or multiple seeds are reported; a leakage-driven shift of a few points is enough to change the ranking. Independent of leakage, the paper's own Table 2 contradicts the conclusion that pre-trained models 'consistently outperformed' scratch 'at all epochs': at 50 epochs YOLO-acne (86.4) and YOLO-brain-tumor (86.6) have lower F1 than trained-from-scratch (88.9). In addition, Table 6 shows that by mAP50-95 HAM10000-XL (74.9) exceeds fruit&veg-XL (74.2), so the 'best polyp detector' claim is metric-dependent even before considering leakage. The broad claim should be narrowed to 'the best pre-trained models help; some small-domain pre-training can hurt.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether pre-training YOLOv8n on different datasets improves polyp detection in colonoscopy images. It pre-trains on seven datasets (small: MRI brain tumor, acne, fruit/vegetable; larger XL: HAM10000, combined brain tumor, combined fruit/vegetable, plus a COCO-pretrained baseline), fine-tunes on a combined dataset of four public polyp datasets at 20/50/100 epochs, and reports precision, recall, F1, mAP50, and mAP50-95. The central claims are that pre-trained models consistently outperform training from scratch, that larger pre-training datasets help, and that the YOLO-fruit&veg-XL model achieves the highest overall F1-score of 95.1% after 100 epochs.","tokens_in":10223,"tokens_out":3644,"duration_ms":34728,"significance":"If validated, the paper would provide practical guidance on pre-training data selection for medical object detection and would contribute reusable pre-trained models and processed datasets. The experimental design is transparent and the authors explicitly release their code and models, which is a strength. However, the finer ranking claims rest on a single random split of a small dataset, with no error bars, no repeated seeds, and a split that may leak polyp identity between training and test; these issues must be addressed before the specific rankings can be trusted.","major_comments":[{"comment":"The fine-tuning set is split randomly at image level (80/10/10), while §3.4 acknowledges that the dataset 'consists of multiple images of the same polyp from different perspectives.' This means near-duplicate frames of the same polyp can appear in both training and test, which inflates reported F1/mAP and can change rankings. The differences between the top models are small (e.g., at 100 epochs fruit&veg-XL 95.1% vs. COCO 94.5% vs. HAM10000-XL 94.4%), so leakage of a few points is enough to alter the conclusions. Please stratify the split by polyp identity or video, or otherwise guarantee no identity leakage, and report results under a leakage-free protocol.","section":"§2.2 and §3.4"},{"comment":"The conclusion that pre-trained models 'consistently outperformed' the scratch model 'at all epochs' is contradicted by Table 2: at 50 epochs, YOLO-acne (86.4% F1) and YOLO-brain-tumor (86.6% F1) fall below trained-from-scratch (88.9% F1), and the same is true for mAP50-95 (68.1% and 65.2% vs. 67.1%). Please narrow the claim to 'the best pre-trained models help' and explicitly discuss that some small-domain pre-training can hurt.","section":"§3.2, Table 2"},{"comment":"The statement that YOLO-fruit&veg-XL is the best overall model is metric-dependent: although it has the highest F1 at 100 epochs (95.1%), HAM10000-XL has higher mAP50-95 (74.9% vs. 74.2%) and nearly equal mAP50 (97.0% vs. 97.1%). Moreover, all numbers come from a single random split with no repeated seeds or confidence intervals, so the 0.6-point F1 gap between fruit&veg-XL and COCO (95.1 vs. 94.5) cannot be distinguished from noise. Please add multiple seeds, report variance, and either apply statistical testing or explicitly state that the differences are not statistically significant.","section":"§3.2, Tables 4–6"}],"minor_comments":[{"comment":"The first sentence of Section 3.2 is malformed: 'From all the results, the YOLO-fruit&veg-XL model achieved the highest overall F1-Among all the models evaluated...' should be rewritten as a single complete sentence.","section":"§3.2"},{"comment":"The abstract says 'objection tasks'; this should be 'object detection tasks'.","section":"Abstract"},{"comment":"The word 'treshold' appears twice and should be 'threshold'.","section":"§2.3"},{"comment":"The text says 'adam optimzer'; this should be 'Adam optimizer'.","section":"§3.1"},{"comment":"The text says 'The hyperparameters were kept consist'; this should be 'kept consistent'.","section":"§2.2"},{"comment":"The heading 'F uture W ork' has unintended spaces; also, 'k-means cross-validation' should be 'k-fold cross-validation'.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"This is a modest empirical contribution whose main value is the released pre-trained models and processed datasets. The central claims are currently overbroad relative to the evidence: the 'consistently outperform at all epochs' statement is false as written, and the single-split, possibly leaking evaluation cannot support the fine-grained ranking between pre-training datasets. The fix is within scope: use stratified splits, multiple seeds, and weakened claims. I would not reject, but the paper needs substantial revision before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the systematic comparison of seven pre-training sources for YOLOv8n polyp detection, including fruit/vegetable and brain/skin medical datasets. That experiment is worth having on record, and I give the authors credit for running it with a fixed architecture and reporting the per-epoch tables. They are also honest about the main weakness in their own future-work section: the fine-tuning data contains multiple images of the same polyp, so the single random 80/10/10 split may leak near-duplicate frames into the test set. That admission does not fix the test set, but it is more candid than most papers of this type.\n\nThe broad claim that pre-training helps is supported, especially when the pre-training dataset is large and diverse. But the paper overstates it in two ways. First, the abstract and conclusions say pre-trained models 'consistently outperform' scratch 'at all epochs,' and Table 2 contradicts that: at 50 epochs YOLO-acne (86.4 F1) and YOLO-brain-tumor (86.6) are below trained-from-scratch (88.9). Second, the 'best model' claim is metric-dependent: by mAP50-95 at 100 epochs, HAM10000-XL (74.9) beats fruit&veg-XL (74.2), while fruit&veg-XL leads on F1. So the paper's own tables tell a more nuanced story than the conclusions do.\n\nThe bigger problem is statistical. The ranking among pre-training datasets rests on differences of a few tenths of a point, with a single run, no error bars, and a split that the authors themselves admit may be contaminated by same-polyp duplication. A leakage-driven shift of two or three points is enough to reorder the top of Table 6. The stress-test note is right to flag this. The authors should narrow the claim to something like 'the best pre-trained models help; some small-domain pre-training can hurt,' and back it with multiple seeds, stratified splitting by polyp, and significance tests.\n\nAll that said, the paper is not broken beyond repair. The experiment is legitimate, the data and code are promised, and the limitations section is unusually frank. The right outcome is not desk rejection but a major revision that addresses the split and the statistical support for the ranking. I would send it to peer review, with the expectation that the authors either provide leakage-free evaluation or soften the ranking claims.\n\nWho gets value: researchers in medical imaging transfer learning who want a quick mapping of what pre-training data works for polyp detection. It is a modest applied contribution, but a real one.","headline":"Useful comparison of pre-training datasets for polyp detection, but the specific ranking claims rest on a leak-prone split and no error bars; worth refereeing if the authors tighten the statistics.","tokens_in":10734,"tokens_out":1250,"would_cite":false,"duration_ms":14628,"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 pre-training on visually similar out-of-domain data beats in-domain medical data for polyp detection.","keywords":["YOLOv8","transfer learning","polyp detection","colonoscopy","pre-training","fine-tuning","object detection","dataset selection"],"falsifier":"Compute the overlap of polyp identities between the training and test partitions of the combined dataset; if any polyp appears in both, re-evaluate all models on a patient-level split (or k-fold cross-validation) and check whether the fruit-and-veg model still tops YOLO-coco and whether every pre-trained model still beats from-scratch.","tokens_in":9730,"feed_emoji":"🍇","tokens_out":4962,"duration_ms":42410,"temperature":0.7,"pith_summary":"The paper asks which pre-training dataset gives the best transfer to polyp detection in colonoscopy images, and whether any pre-training beats training from scratch. It reports that a YOLOv8n model pre-trained on a combined fruit-and-vegetable dataset reached the highest F1-score (95.1%) after fine-tuning, outperforming models pre-trained on COCO, brain-tumor MRI, skin-lesion, and acne data. Across all epochs, every pre-trained model beat the from-scratch baseline, and models pre-trained on larger datasets beat their smaller counterparts. The authors conclude that shared visual features, such as round, shiny objects against textured backgrounds, matter more than the medical-domain label when choosing a pre-training set.","feed_headline":"Fruit-and-veg pretraining bests COCO for colon polyp detection","feed_subtitle":"Study of YOLOv8 shows any pretraining beats from scratch; visual similarity drives the win.","key_machinery":"The central object is the YOLOv8n object detector with transfer learning: a small (3.01 million parameter) network pre-trained on one of seven datasets and then fine-tuned on a combined polyp dataset. The transfer mechanism is weight initialization: the fine-tuning stage starts from weights learned on the pre-training dataset, and the same hyperparameters and augmentations are used across all runs to isolate the effect of pre-training choice. The comparison structure, pitting out-of-domain general datasets (COCO, fruit/vegetables) against in-domain medical sets (brain MRI, skin lesions, acne), is what carries the argument about which features transfer.","core_discovery":"The central claim is that transfer learning for polyp detection does not require an in-domain medical pre-training set; a dataset of fruits and vegetables, chosen for visual similarity to polyps, yields the strongest detector. The paper establishes this by pre-training YOLOv8n on seven datasets, fine-tuning on a combined set of four public polyp datasets, and evaluating F1, mAP50, and mAP50-95. The best model, YOLO-fruit&veg-XL, reached 95.1% F1 at 100 epochs, surpassing YOLO-coco (94.5%), all medical-image pre-trained models, and the from-scratch baseline (92.9%). The authors also find that pre-training on at least 10,000 images consistently improves results over smaller pre-training sets, and that every pre-trained model converges faster than the baseline. They interpret this as evidence that diversity and visual similarity are the driving factors, not the domain label itself.","pith_inferences":["The paper's own limitation note says the fine-tuning set contains multiple images of the same polyp, so the single 80/10/10 split may leak polyp identity between training and test; a patient-level split could change the ranking, especially the small gaps between the top models.","The fruit-and-vegetable advantage hints that a specifically curated dataset of round, glossy, textured objects could be an even stronger pre-training set for polyp detection; that is testable.","The 'more data helps' result may conflate dataset size with dataset diversity; an experiment that controls class count while varying image count would separate the two effects."],"forward_implications":["If replicated, polyp detection models can be built without large medical pre-training corpora, using publicly available food images instead.","Pre-training on larger, more diverse datasets (10k+ images) will consistently beat smaller pre-training sets, guiding dataset curation.","Any pre-training, even on acne or brain MRI, improves over from-scratch training, so practitioners should always use a pre-trained initialization.","The ranking suggests visual similarity (shape, shine, texture) should be a primary criterion for selecting pre-training data, not just domain closeness."],"supporting_citations":[{"why":"Supplies CVC-ClinicDB polyp images used in the fine-tuning set.","marker":"[12]"},{"why":"Supplies CVC-ColonDB polyp images used in the fine-tuning set.","marker":"[13]"},{"why":"Supplies ETIS-LaribPolypDB polyp images used in the fine-tuning set.","marker":"[14]"},{"why":"Supplies Kvasir-SEG polyp images used in the fine-tuning set.","marker":"[15]"},{"why":"Provides the COCO17 dataset used as the large general-purpose pre-training control.","marker":"[28]"},{"why":"Provides the Fruit and Vegetable Detection dataset that produced the best-performing pre-trained model.","marker":"[27]"},{"why":"Provides the HAM10000 skin lesion dataset used as an in-domain medical pre-training set.","marker":"[29]"},{"why":"Defines the YOLOv8 architecture and supplies the COCO-pre-trained weights used as a control.","marker":"[35]"},{"why":"Supplies the transfer learning survey that frames the paper's approach.","marker":"[18]"}],"fun_headline_variants":["Pretraining on fruit and veg nails polyp detection","Fruit-and-veg pretraining tops COCO for polyps","Polyp detection boosted by fruit-and-veg pretraining","Why fruit-and-veg pretraining beats COCO for polyps","Any pretraining beats scratch for polyp detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The single random 80/10/10 split assumes that images of the same polyp never appear in both the training and test sets; if they do, the reported F1 and mAP are inflated and the ranking between pre-training datasets could change.","fun_headline_variants_meta":{"raw":{"variants":["Pretraining on fruit and veg nails polyp detection","Fruit-and-veg pretraining tops COCO for polyps","Polyp detection boosted by fruit-and-veg pretraining","Why fruit-and-veg pretraining beats COCO for polyps","Any pretraining beats scratch for polyp detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000634,"raw_usage":{"total_tokens":2905,"prompt_tokens":905,"completion_tokens":2000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1920}},"tokens_in":521,"tokens_out":2000,"duration_ms":13648,"temperature":1.0,"reasoning_tokens":1920,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:03:07.165056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the overlap of polyp identities between the training and test partitions of the combined dataset; if any polyp appears in both, re-evaluate all models on a patient-level split (or k-fold cross-validation) and check whether the fruit-and-veg model still tops YOLO-coco and whether every pre-trained model still beats from-scratch.","supporting_citations":[{"cited_title":"WM-DOV A maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians,","cited_arxiv_id":null,"evidence_quote":"Supplies CVC-ClinicDB polyp images used in the fine-tuning set."},{"cited_title":"Towards automatic polyp detection with a polyp appearance model,","cited_arxiv_id":null,"evidence_quote":"Supplies CVC-ColonDB polyp images used in the fine-tuning set."},{"cited_title":"Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer,","cited_arxiv_id":null,"evidence_quote":"Supplies ETIS-LaribPolypDB polyp images used in the fine-tuning set."},{"cited_title":"Kvasir-seg: A segmented polyp dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies Kvasir-SEG polyp images used in the fine-tuning set."},{"cited_title":"Microsoft coco: Common objects in context,","cited_arxiv_id":null,"evidence_quote":"Provides the COCO17 dataset used as the large general-purpose pre-training control."},{"cited_title":"Fruits & vegetable detection for yolov4","cited_arxiv_id":null,"evidence_quote":"Provides the Fruit and Vegetable Detection dataset that produced the best-performing pre-trained model."},{"cited_title":"HAM1000 Segmentation and Classification","cited_arxiv_id":null,"evidence_quote":"Provides the HAM10000 skin lesion dataset used as an in-domain medical pre-training set."},{"cited_title":"Ultralytics yolov8,","cited_arxiv_id":null,"evidence_quote":"Defines the YOLOv8 architecture and supplies the COCO-pre-trained weights used as a control."},{"cited_title":"A survey on transfer learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the transfer learning survey that frames the paper's approach."}],"review_version":1}