{"id":"bee8fd0a-13e4-4926-8fce-da8e250f23e3","arxiv_id":"2501.00647","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"G-YOLOv11l, a ghost-convolution and halved-filter YOLOv11 variant, reaches mAP@0.5 of 0.535 at 2.4 ms per image, 68.7% smaller than YOLOv11l but 13.6% less accurate.","lead":"A YOLOv11 detector reworked with ghost convolutions and halved filters finds wrist fractures in children's X-rays, cutting model size by 68.7% at the cost of a 13.6% drop in detection accuracy. The trade-off is measured on a public pediatric wrist dataset, with the smaller model running at 2.4 ms per image.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported mAP rests on an image-level split of GRAZPEDWRI-DX (20,327 images from 6,091 patients), so images from the same patient are likely in both train and test; until a patient-stratified split is tested, the 0.535 mAP and the 13.6% drop vs YOLOv11l are not reliable.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I see: the accuracy side of the central efficiency-accuracy claim is evaluated on a test set that is not patient-disjoint. This is the most fundamental threat because the abstract's headline numbers (mAP@0.5 of 0.535, 13.6% reduction relative to YOLOv11l) are only meaningful if the test set measures generalization to new patients, which is the intended clinical use case. I also considered two secondary issues: the efficiency gain is confounded with halving the number of filters, and the speed numbers are reported without batch size or precision details. These are real but less decisive: filter halving is transparently described, and even if ghost convolution contributes little, the full system is still smaller and faster. The patient-leakage concern, by contrast, threatens the validity of the accuracy metric itself and is easily testable with the available data and code. Because the reader already issued a CONDITIONAL verdict that captures exactly this risk, my read does not change the verdict.","tokens_in":10718,"tokens_out":7021,"duration_ms":67366,"concrete_test":"Use GroupShuffleSplit or GroupKFold with patient_id as the grouping key to create 70/20/10 splits that keep all images of each patient in a single partition. Retrain G-YOLOv11l and YOLOv11l with identical hyperparameters, augmentation, and epoch count, then compare mAP@0.5 on this patient-disjoint test set against the image-level split results in Table 3. If mAP@0.5 drops by more than ~0.03-0.05, or if the relative gap between G-YOLOv11l and YOLOv11l widens materially, the reported efficiency-accuracy trade-off is inflated by patient leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 describes a random 70/20/10 split of individual images, not patients. GRAZPEDWRI-DX has 20,327 images from 6,091 patients, so on average roughly 3.3 images per patient; a random split will place many patients' images in both training and test sets. Because serial X-rays of the same wrist are near-duplicates in anatomy, positioning, and often fracture appearance, the model can memorize patient-specific features rather than learn fracture-relevant, generalizable cues. This inflates mAP@0.5 and can also distort the comparison with YOLOv11l if the two models memorize at different rates. Since the central claim is an efficiency-accuracy trade-off (0.535 mAP at 68.7% size reduction, 'new state-of-the-art benchmark'), a test set that does not isolate patients cannot support the benchmark claim. The provided code and public dataset make this concern directly checkable, but the paper gives no indication that a patient-level split was used.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes G-YOLOv11, a lightweight variant of YOLOv11 in which standard convolutions and C3k2 modules are replaced by GhostConv and C3Ghost modules, and the number of filters is halved. The authors report experiments on the GRAZPEDWRI-DX pediatric wrist X-ray dataset, claiming an mAP@0.5 of 0.535 for G-YOLOv11l at 2.4 ms inference time on an NVIDIA A10 GPU, a 68.7% size reduction relative to YOLOv11l, and describe G-YOLOv11l as a new state-of-the-art benchmark in efficiency. The paper includes parameter/FLOP comparisons, accuracy tables, qualitative detection examples, and a comparison with prior YOLOv8/YOLOv9/YOLOv10/YOLOv11 baselines reproduced under the same data split.","tokens_in":10920,"tokens_out":3918,"duration_ms":34519,"significance":"If the reported measurements are reliable, the paper demonstrates that a heavily compressed YOLOv11 retains moderate detection accuracy on a pediatric wrist radiograph dataset while offering substantial reductions in parameters, FLOPs, model size, and inference time. The availability of code and a public dataset is a strength, and the reproduction of baselines under the same split is good practice. However, the central efficiency-accuracy trade-off is not characterized statistically, the architectural comparison is confounded, and the test split may mix images from the same patient, so the benchmark claim should be treated as conditional until these points are addressed.","major_comments":[{"comment":"The random 70/20/10 split is performed at the image level, while GRAZPEDWRI-DX contains 20,327 images from 6,091 patients. With roughly 3.3 images per patient, serial radiographs of the same wrist can appear in both the training and test sets. Because such near-duplicate images are highly similar in anatomy and appearance, the model can memorize patient-specific features rather than learn generalizable fracture cues, which would inflate the reported mAP@0.5 of 0.535 and bias the comparisons in Tables 3 and 5. Please re-split the data at the patient level (all images of a patient in one fold) and report the metrics on that split; this is a necessary condition for the claimed benchmark.","section":"Section 2.1"},{"comment":"The efficiency gain is confounded: the proposed model changes the convolution type (Conv to GhostConv) and simultaneously halves the number of filters (e.g., YOLOv11l filters [64,128,256,512,512] become [32,64,128,256,512]). As a result, the measured reductions in parameters, FLOPs, and inference time cannot be attributed to ghost convolution specifically. To support the architectural claim, the paper needs an ablation that (a) applies ghost convolution while keeping the filter counts unchanged, and (b) applies standard convolution with halved filters. Without such an ablation, the efficiency numbers are real but their cause is undetermined.","section":"Section 2.2.5 and Table 1"},{"comment":"The inference-time claim is internally inconsistent. The abstract and Table 5 report 2.4 ms for G-YOLOv11l, while Table 3 lists a speed of 3 ms for the same model; similarly, YOLOv11l is listed as 6 ms in Table 3 but 5.5 ms in Table 5. Since the central contribution is an efficiency benchmark, this discrepancy must be resolved, and the measurement protocol (including whether preprocessing and post-processing are included) should be stated explicitly.","section":"Abstract, Table 3, and Table 5"},{"comment":"The claim of a 'new state-of-the-art benchmark in terms of efficiency' is not supported by the metrics as presented. G-YOLOv11l has the lowest detection accuracy of all compared detectors (mAP@0.5 of 0.535 vs. 0.568–0.672 for the others), and no efficiency-accuracy trade-off metric (e.g., mAP per FLOP or a Pareto-frontier analysis) is provided to justify calling this a benchmark. Moreover, all numbers come from a single training run; no error bars or repeated-seed statistics are reported, so statements such as 'reductions of 13.6%' have no demonstrated statistical significance.","section":"Section 4.3 and Table 5"}],"minor_comments":[{"comment":"The citation [20] is given to Khanam and Hussain's YOLOv11 overview, but the text credits Redmon et al. with introducing YOLO in 2015; please correct the citation to the original YOLO paper.","section":"Section 2.2"},{"comment":"The table lists four filter values for C3k2/C3Ghost modules that are described as eight modules (e.g., YOLOv11l and G-YOLOv11l rows); please clarify how the values map to the module list or provide the repeated list explicitly.","section":"Table 1"},{"comment":"The phrase 'reductions of 13.6% in mAP@0.5' is ambiguous because a lower mAP is a degradation rather than a reduction in resource use; consider rewording to 'a 13.6% relative decrease in mAP@0.5'.","section":"Abstract"},{"comment":"Figure 1 is not referenced in the text, and it is unclear whether the plotted points come from Tables 2–5 or from separate runs; please add a cross-reference and describe the data source in the caption.","section":"Figure 1"},{"comment":"Key training hyperparameters (initial learning rate, momentum, weight decay, learning-rate schedule, mosaic/mixup settings, and image size) are omitted; please state them or point to the configuration file in the repository to make the experiments reproducible.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central efficiency claim is plausible but currently rests on a test split that may leak patient information, a confounded architectural comparison (ghost convolution + filter halving changed simultaneously), and an inference-time inconsistency between tables. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The self-citations [15,16] on pruning are not load-bearing for the fracture-detection claim and could be trimmed or replaced with more directly relevant references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The architecture is a straightforward, well-documented combination: replace standard convs with ghost convolutions from GhostNet and C3k2 with C3Ghost in YOLOv11, and halve filter counts. Code and pretrained weights are public. The reported resource reductions are real and reproducible. But the headline mAP numbers are not reliable, because the test set is a random image-level split of GRAZPEDWRI-DX, not patient-stratified.\n\nWhat is actually new: none of the components are new; the novelty is the specific configuration for pediatric wrist fracture detection. That's fine for an application paper, but the 'state-of-the-art efficiency benchmark' claim is overreach.\n\nWhat it does well: the paper includes detailed per-module filter counts, loss curves, PR curves, and qualitative examples that show false positives. It reproduces comparison detectors under the same split. The writing is transparent about the accuracy drop (13.6% mAP@0.5 vs YOLOv11l).\n\nSoft spots: the split issue is load-bearing. GRAZPEDWRI-DX has 20,327 images from 6,091 patients; a random 70/20/10 split puts many patients' images in both train and test. Since serial wrist images are near-duplicates, the model can memorize patient-level features, inflating mAP and possibly biasing relative comparisons. The paper follows [10]'s split, but that doesn't address the problem. Second, the efficiency gain is confounded: ghost convolution and filter halving change together, so you can't attribute the savings to ghost convs without an ablation. Third, SOTA efficiency is asserted after comparing only to large detectors; no lightweight baselines. Also no error bars or significance tests; a couple of speed numbers differ between tables, which suggests measurement variability.\n\nBottom line: this is an honest engineering note with a fixable methodological flaw. The fix (patient-stratified split) is easy given the public dataset and code. I'd send it to review, mainly so the authors are asked to do that rerun and tone down the SOTA claim. The paper would then be a useful data point for practitioners, not a breakthrough.","headline":"A clean but conceptually derivative lightweight YOLOv11 for wrist fractures; the accuracy numbers are unreliable because the data split leaks patients.","tokens_in":11468,"tokens_out":2561,"would_cite":false,"duration_ms":24072,"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":"G-YOLOv11, a YOLOv11 variant with ghost convolutions and halved filter counts, reports wrist-fracture mAP@0.5 of 0.535 at 2.4 ms inference on an A10 GPU, reducing model size by 68.7% versus YOLOv11l.","keywords":["pediatric wrist fracture detection","YOLOv11","ghost convolution","lightweight object detection","computer-aided diagnosis","X-ray imaging","GRAZPEDWRI-DX","model compression"],"falsifier":"Retrain G-YOLOv11l with the same hyperparameters but on a patient-disjoint split of GRAZPEDWRI-DX, meaning all images from a given patient are confined to a single fold, and compare mAP@0.5 on the held-out patients; if the score falls well below the reported 0.535, the published accuracy is inflated by images of the same patients appearing in both training and test sets.","tokens_in":10489,"feed_emoji":"🦴","tokens_out":6655,"duration_ms":58680,"temperature":0.7,"pith_summary":"This paper tries to show that a pediatric wrist fracture detector can become much cheaper to run without giving up too much accuracy. The proposed detector, G-YOLOv11, takes YOLOv11 and replaces ordinary convolutions and C3k2 blocks with ghost convolution and C3Ghost blocks, and it halves the number of filters in each stage. On the GRAZPEDWRI-DX wrist X-ray dataset, the large variant reports an mAP@0.5 of 0.535 with a 2.4 ms inference time on an NVIDIA A10 GPU, using 7.8 million parameters and 21.0 GFLOPs. Compared with the standard large YOLOv11, this is a 13.6% relative drop in mAP@0.5 while reducing model size by 68.7%. The authors claim this establishes a new efficiency benchmark, arguing that modest accuracy loss is an acceptable trade for real-time clinical use.","feed_headline":"Ghost-convolution YOLOv11 cuts wrist-fracture detector size by 69%","feed_subtitle":"Holds 0.535 mAP while using 7.8 million parameters and 21 GFLOPs, easing clinical deployment.","key_machinery":"The load-bearing mechanism is ghost convolution, which produces the same number of feature maps as a standard convolution by first computing a small set of intrinsic feature maps with a 1x1 convolution and then generating the remaining maps with cheap 5x5 linear operations, concatenating both outputs. The paper inserts this operation in place of every Conv module and packages it into C3Ghost blocks, which combine two ghost convolutions around a depthwise convolution with a shortcut path, replacing YOLOv11's C3k2 blocks. On top of this substitution, each stage's filter count is halved relative to the corresponding YOLOv11 variant. These two changes together cut parameters, FLOPs, memory, and inference time, and they carry the efficiency-accuracy trade-off reported in the paper's experimental tables.","core_discovery":"The paper's central claim is that the standard YOLOv11 detector, which is too heavy for clinical computer-aided diagnosis, can be turned into a lightweight detector by replacing its convolution and C3k2 modules with ghost convolution and C3Ghost modules and halving the filter counts across the network. On the GRAZPEDWRI-DX test set, G-YOLOv11l reaches an mAP@0.5 of 0.535 and an mAP@0.5:0.95 of 0.341 with an inference time of 2.4 ms on an NVIDIA A10 GPU, using 7.794 million parameters and 21.0 GFLOPs. In contrast, YOLOv11l achieves 0.619 mAP@0.5, 25.317 million parameters, 87.3 GFLOPs, and 6 ms inference, so the proposed model trades a 13.6% relative accuracy reduction for a 68.7% smaller model and roughly 2.5x faster inference. The paper interprets this as a new efficiency-oriented benchmark, because it outperforms existing detectors on resource usage while remaining within a moderate accuracy range.","pith_inferences":["If the train/test split is not patient-disjoint, the reported mAP values likely overstate performance on new children; a patient-disjoint split would be a fairer test of whether the efficiency gains come with the accuracy cost shown.","The design decision to halve filters is entangled with the ghost-convolution replacement, so ablating each change separately, ghost only, halving only, then both, would show which one drives the efficiency gain.","The same lightweight recipe could be applied to other YOLO versions or to other small-object medical detection tasks, but the accuracy transfer would need revalidation on datasets with class imbalance similar to wrist X-rays.","For bedside deployment, the 2.4 ms timing on an A10 GPU should be re-measured on lower-power clinical hardware or with integer quantization, since deployed hardware is unlikely to match that GPU."],"forward_implications":["The large G-YOLOv11 model runs at 2.4 ms per image on an NVIDIA A10 GPU, making real-time fracture screening plausible on hardware of that class.","Deploying the detector in a CAD system would require roughly one-third of the storage (15.3 MB versus 48.8 MB) and far fewer FLOPs, easing integration into clinical workstations.","The same recipe, ghost convolution plus halved filters, scales across nano, small, medium, and extra-large configurations with proportional savings in parameters and FLOPs as reported in Table 2.","The accuracy cost on this dataset is bounded: no more than an 8.4% relative mAP@0.5 difference from the corresponding YOLOv11 configuration, according to Table 3.","Compared with the large YOLOv8, YOLOv9, YOLOv10, and YOLOv11 baselines in Table 5, G-YOLOv11l has the lowest parameter count, FLOPs, and inference time among the compared models."],"supporting_citations":[{"why":"Supplies the GRAZPEDWRI-DX dataset, its 20,327 X-ray images, patient counts, and bounding-box annotations used in all experiments.","marker":"[8]"},{"why":"Provides the YOLOv8 baseline on this dataset and the mAP@0.638 result that the paper reuses as a comparison point.","marker":"[9]"},{"why":"Supplies the data division and augmentation protocol (70/20/10 split with contrast and luminance augmentation) and the YOLOv9 baselines used for fair comparison.","marker":"[10]"},{"why":"Provides the YOLOv10 baseline and dual label assignment method that appears in the comparison table.","marker":"[11]"},{"why":"Represents the Enhanced-YOLO11 baseline that the paper compares against as the standard YOLOv11 detector.","marker":"[12]"},{"why":"Defines ghost convolution and the observation of redundant feature maps, which is the core mechanism the proposed detector adapts.","marker":"[17]"},{"why":"Provides the architectural overview of YOLOv11 that the paper relies on for describing the baseline structure being modified.","marker":"[20]"},{"why":"Supplies the official YOLOv11 implementation and pretrained weights used for initialization and for all standard YOLOv11 baseline measurements.","marker":"[21]"}],"fun_headline_variants":["Lighter YOLOv11 detects wrist fractures in 2.4 ms","69% smaller wrist-fracture detector runs in 2.4 ms","G-YOLOv11 trades 14% accuracy for 69% smaller size","Pediatric wrist-fracture detector shrinks 69% with ghost convs","Ghost-conv YOLOv11 cuts size 69% for pediatric X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The data split is a random 70/20/10 split of individual images, not of patients, so images from the same child can appear in both training and test sets; if that happens, the reported mAP values are inflated and the comparisons with baselines become biased.","fun_headline_variants_meta":{"raw":{"variants":["Lighter YOLOv11 detects wrist fractures in 2.4 ms","69% smaller wrist-fracture detector runs in 2.4 ms","G-YOLOv11 trades 14% accuracy for 69% smaller size","Pediatric wrist-fracture detector shrinks 69% with ghost convs","Ghost-conv YOLOv11 cuts size 69% for pediatric X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001151,"raw_usage":{"total_tokens":4828,"prompt_tokens":1060,"completion_tokens":3768,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":3664}},"tokens_in":676,"tokens_out":3768,"duration_ms":27078,"temperature":1.0,"reasoning_tokens":3664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:45:28.655823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain G-YOLOv11l with the same hyperparameters but on a patient-disjoint split of GRAZPEDWRI-DX, meaning all images from a given patient are confined to a single fold, and compare mAP@0.5 on the held-out patients; if the score falls well below the reported 0.535, the published accuracy is inflated by images of the same patients appearing in both training and test sets.","supporting_citations":[{"cited_title":"A pediatric wrist trauma x-ray dataset (grazpedwri-dx) for machine learning","cited_arxiv_id":null,"evidence_quote":"Supplies the GRAZPEDWRI-DX dataset, its 20,327 X-ray images, patient counts, and bounding-box annotations used in all experiments."},{"cited_title":"Fracture detection in pediatric wrist trauma x-ray images using yolov8 algorithm","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv8 baseline on this dataset and the mAP@0.638 result that the paper reuses as a comparison point."},{"cited_title":"Yolov9 for fracture detection in pediatric wrist trauma x-ray images","cited_arxiv_id":null,"evidence_quote":"Supplies the data division and augmentation protocol (70/20/10 split with contrast and luminance augmentation) and the YOLOv9 baselines used for fair comparison."},{"cited_title":"Pediatric Wrist Fracture Detection in X-rays via YOLOv10 Algorithm and Dual Label Assignment System","cited_arxiv_id":"2407.15689","evidence_quote":"Provides the YOLOv10 baseline and dual label assignment method that appears in the comparison table."},{"cited_title":"Detection of bone fractures along with other abnormali- ties in wrist x-ray images using enhanced-yolo11","cited_arxiv_id":null,"evidence_quote":"Represents the Enhanced-YOLO11 baseline that the paper compares against as the standard YOLOv11 detector."},{"cited_title":"Ultralytics yolo11, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the official YOLOv11 implementation and pretrained weights used for initialization and for all standard YOLOv11 baseline measurements."}],"review_version":1}