Fracture-YOLO: Pediatric Fracture Detection

• Fracture-YOLO is a family of YOLO-based systems that localize pediatric wrist fractures by adapting architectures with attention mechanisms, context injection, and tailored augmentation pipelines.
• The models evolve from YOLOv8 to YOLOv11 with innovations like CRSelector, Scale-Aware heads, PGI, and GELAN, enhancing detection sensitivity and addressing class imbalance.
• These systems are designed for practical deployment with low latency and robust performance, serving as decision-support tools in clinical settings for emergency fracture triage.

Fracture-YOLO denotes a family of YOLO-based computer-assisted diagnosis systems for fracture localization in X-ray images, most prominently pediatric wrist trauma radiographs from the GRAZPEDWRI-DX benchmark. In the cited literature, the term is used in two related senses: as an informal label for successive YOLO adaptations specialized for pediatric wrist fracture detection, and as the formal name of a later YOLOv8-based model that adds Critical-Region-Selector Attention and a Scale-Aware head. Across these works, the task is framed primarily as object detection rather than image-level classification, with the emphasis shifting over time from baseline YOLO transfer, to attention and context modeling, to end-to-end NMS-free detection, and finally to lightweight deployment-oriented variants (Ju et al., 2023, Chien et al., 2024, Ahmed et al., 2024, Sun et al., 27 Sep 2025).

1. Terminology and problem formulation

In this literature, Fracture-YOLO is not a single fixed architecture. Rather, it is a recurring designation for YOLO-family detectors adapted to fracture CAD, especially for pediatric wrist trauma. The common clinical target is rapid identification of fractures and related abnormalities in wrist radiographs, motivated by the high prevalence of pediatric wrist trauma and the consequences of missed or delayed diagnosis, including deformity, restricted motion, chronic pain, and long-term complications. Several papers explicitly position these systems as decision-support tools for radiologists and surgeons rather than autonomous replacements (Chien et al., 2024, Ju et al., 2024, Chien et al., 2024).

A defining property of this line of work is the formulation of fracture analysis as localization of abnormal findings in radiographs. The detectors therefore predict bounding boxes over abnormalities such as fracture, bone lesion, metal, or pronator sign, instead of returning only a study-level fracture label. This distinguishes Fracture-YOLO from weakly supervised fracture-classification pipelines that infer probable regions of interest without explicit YOLO-style box regression. A useful contrast is the two-stage weakly supervised pelvic X-ray framework of "Weakly Supervised Universal Fracture Detection in Pelvic X-rays," which uses DenseNet-121 in a multiple instance learning setting followed by ROI classification, and is explicitly not a YOLO-like detector (Wang et al., 2019).

2. Benchmark dataset, labels, and evaluation practice

The dominant benchmark for Fracture-YOLO studies is GRAZPEDWRI-DX, a public pediatric wrist trauma X-ray dataset from the Medical University of Graz. Reported dataset facts include 20,327 X-ray images, 6,091 patients, and 10,643 studies, with 74,459 image labels and 67,771 annotated objects in the broader annotation pool. The dataset contains nine object/pathology categories: bone anomaly, bone lesion, foreign body, fracture, metal, periosteal reaction, pronator sign, soft tissue, and text (Chien et al., 2024, Ju et al., 2024).

No official split is provided, so papers construct their own partitions. A large subset of the literature uses a 70/20/10 split, for example 14,204 training, 4,094 validation, and 2,029 test images, often with training-set doubling through brightness and contrast adjustment using OpenCV addWeighted. A separate YOLOv10 study instead uses a 75/20/5 split with 15,245 training, 4,066 validation, and 1,016 test images, specifically to permit same-split retraining of YOLOv9 for fair comparison (Ju et al., 2023, Ju et al., 2024, Ahmed et al., 2024).

Class imbalance is a persistent difficulty. One YOLOv9 study reports “text” (23,722 instances) and “fracture” (18,090 instances) as the most frequent labels, with “bone anomaly” (276) and “soft tissue” (464) as rare classes. The YOLOv10 study reports that fracture is the dominant category in its evaluation setting, with 13,550 instances (66.6%). This imbalance directly shapes the literature’s error analyses: fracture, metal, and text are usually learned well, whereas bone anomaly and soft tissue remain weak categories (Chien et al., 2024, Ahmed et al., 2024).

The evaluation protocol is standard object detection. Commonly reported metrics are mAP@50, mAP@50-95, F1 score, precision, recall/sensitivity, Params, FLOPs, and inference time. In the reviewed papers, $mAP@50$ denotes mean average precision at $\mathrm{IoU}=0.50$, while $mAP@50$-$95$ averages over IoU thresholds from $0.50$ to $0.95$ in steps of $0.05$ (Ahmed et al., 2024, Ju et al., 2024).

3. Architectural evolution of Fracture-YOLO

The earliest pediatric wrist Fracture-YOLO systems in this sequence are YOLOv8-based. The 2023 YOLOv8 study uses a standard YOLOv8 detector with backbone, neck, head, and loss, emphasizing C2f, SPPF, FP-PAN, a decoupled and anchor-free head, Task Aligned Assigner (TOOD), and BCE, DFL, and CIoU losses. Its main adaptation is not a new module but a fracture-specific training and augmentation pipeline, later packaged into a PySide6/Qt application with ONNX export for surgeon-facing use (Ju et al., 2023).

Subsequent work concentrates on feature recalibration and context injection within YOLOv8. YOLOv8-AM inserts CBAM, GAM, ECA, and SA into the network, with ResCBAM becoming the strongest variant. Closely related papers then isolate specific modules: YOLOv8+GC inserts one GC block after each of the four C2f modules in the neck, while FCE-YOLOv8 generalizes the idea through four context-excitation modules—SE, GC, GE, and GCT—combined with three insertion strategies: M1 after SPPF, M2 in the final C2f of the head, and M3 after each of the four C2f modules in the head (Chien et al., 2024, Ju et al., 2024, Ju et al., 2024).

YOLOv9-based Fracture-YOLO shifts attention from explicit attention blocks to information preservation during training. Its key architectural arguments are centered on Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN). PGI is described as comprising a main branch, an auxiliary reversible branch, and multi-level auxiliary information, while GELAN combines ideas from CSPNet and ELAN to improve layer aggregation with reduced information loss. The explicit claim is that these mechanisms are well suited to low-feature, low-contrast pediatric radiographs (Chien et al., 2024).

The YOLOv10 study introduces a different inflection point: end-to-end detection without non-maximum suppression. Its central mechanism is the dual-label assignment strategy, combining a one-to-many head and a one-to-one head. The paper associates this design with NMS removal, lower latency, and better handling of overlapping or small pathology regions. The same study emphasizes Compact Inverted Block (CIB), a lightweight classification head, spatial-channel decoupled downsampling, rank-guided block design, selective large-kernel convolutions, and Partial Self-Attention (PSA) applied only after Stage 4 (Ahmed et al., 2024).

A parallel line addresses efficiency directly. G-YOLOv11 keeps the standard YOLOv11 four-part structure—input, backbone, neck, and head—but replaces Conv with GhostConv and C3k2 with C3Ghost, while reducing filters to about half of the original YOLOv11 configurations. The explicitly named Fracture-YOLO model of 2025 instead returns to YOLOv8 and adds Critical-Region-Selector Attention (CRSelector) in the neck and a Scale-Aware (ScA) head in the prediction stage. CRSelector uses global texture information, offsets, a key-mask, and self-attention to emphasize critical fracture regions, whereas ScA dynamically weights multi-scale features (Ferdi, 2024, Sun et al., 27 Sep 2025).

4. Reported empirical performance

Direct comparison across papers requires caution because the literature mixes different splits, input resolutions, augmentation policies, and optimization settings. The strongest claims are therefore most reliable within each paper’s own controlled comparison, especially where same-split retraining is performed.

Method	Representative reported result	Distinguishing feature
YOLOv8-based detector (Ju et al., 2023)	mAP@50 = 0.638	Brightness/contrast augmentation on YOLOv8
YOLOv8-ResCBAM-L (Ju et al., 2024)	[email protected] = 65.8%, [email protected]:0.95 = 42.2%	ResCBAM after four C2f modules
YOLOv8+GC-L (Ju et al., 2024)	[email protected] = 66.32%, [email protected]:0.95 = 42.9%, 7.9 ms	Global Context blocks in the neck
YOLOv9-E (Chien et al., 2024)	[email protected] = 65.62%, [email protected]:0.95 = 43.73%	PGI + GELAN
YOLOv10-M (Ahmed et al., 2024)	mAP@50-95 = 51.9%, fracture sensitivity = 92.5%	Dual-label assignment; NMS-free inference
YOLOv8+SE-M3-L (Ju et al., 2024)	mAP@50 = 67.07%, 15.3 ms	SE context excitation, M3 insertion
G-YOLOv11l (Ferdi, 2024)	[email protected] = 0.535, 2.4 ms inference time	GhostConv + C3Ghost; efficiency-oriented
Fracture-YOLO (Sun et al., 27 Sep 2025)	mAP50 = 65.3%, mAP50-95 = 40.0%	CRSelector + ScA

Several model-selection trends recur. In the broad single-stage comparison of YOLOv5/6/7/8 against Faster R-CNN, YOLOv8x attains the highest all-class mAP of 0.77, while YOLOv8m reaches fracture detection sensitivity = 0.92 and mAP = 0.95 for the fracture class. The same study reports that all YOLO variants outperform Faster R-CNN on GRAZPEDWRI-DX, supporting single-stage detection as the dominant Fracture-YOLO regime for this benchmark (Ahmed et al., 2024).

Within the YOLOv8 attention/context branch, the metric leader depends on whether one prioritizes absolute accuracy or latency. YOLOv8+SE-M3-L records the highest mAP@50 value of 67.07%, whereas YOLOv8+GC-M3-L achieves 66.32% with 7.9 ms inference, outperforming the stated SOTA baseline YOLOv8+ResCBAM-L in both accuracy and speed. This suggests that context excitation and global-context modeling are competitive but induce different speed-accuracy tradeoffs (Ju et al., 2024).

The YOLOv10 results are especially important because they improve the stricter localization metric. On the paper’s 75/20/5 split, YOLOv10-M reaches 51.9% mAP@50-95, surpassing the reported YOLOv9 benchmark of 43.3% by 8.6 percentage points. The same paper also finds that scaling is non-monotonic: performance improves strongly from N → S → M, but moving to L and X does not improve mAP@50-95, making YOLOv10-M the reported sweet spot at 16.5M parameters and 63.5 GFLOPs (Ahmed et al., 2024).

5. Efficiency, deployment, and observed limitations

A major axis of Fracture-YOLO research is deployability. The literature repeatedly frames these systems as suitable for emergency triage, web-based or mobile CAD, and resource-limited settings. YOLOv8-based work culminated in a GUI tool, “Fracture Detection Using YOLOv8 App,” implemented with PySide6/Qt and an ONNX-exported detector. Later papers increasingly treat latency, memory, FLOPs, and model size as primary design targets rather than secondary reporting variables (Ju et al., 2023, Chien et al., 2024).

This deployment emphasis is most explicit in the YOLOv10 and G-YOLOv11 lines. The former removes NMS through dual-label assignment and argues that lower-latency direct inference is especially relevant when subtle findings are small or overlapping. The latter makes a more radical efficiency trade: G-YOLOv11l is reported as 68.7% smaller than YOLOv11l and achieves 2.4 ms inference time on an NVIDIA A10 GPU, but its [email protected] = 0.535 is lower than stronger accuracy-oriented baselines such as YOLOv8 and YOLOv9-E. The paper therefore presents state of the art in efficiency rather than in absolute detection accuracy (Ahmed et al., 2024, Ferdi, 2024).

The limitations reported across studies are consistent. First, severe class imbalance persists. Minority classes such as bone anomaly, bone lesion, and soft tissue remain difficult even when fracture performance is strong. For example, one YOLOv9 paper reports only 11.3% for bone anomaly and 28.2% for soft tissue, despite accuracy above 90% for fracture, metal, and text. ResCBAM and GC variants improve rare-class behavior somewhat, but do not eliminate the imbalance problem (Chien et al., 2024, Ju et al., 2024, Ju et al., 2024).

Second, even the strongest detectors exhibit clinically relevant failure cases. The YOLOv10 paper explicitly includes an example in which a fracture is missed. The YOLOv8 paper notes decreasing accuracy in cases with metal puncture and dense multiple fractures. The G-YOLOv11 paper reports occasional false positives, including a pronator-sign false positive in one qualitative example. These observations constrain the interpretation of Fracture-YOLO as assistive CAD rather than a fully autonomous reader (Ju et al., 2023, Ahmed et al., 2024, Ferdi, 2024).

Third, multiple papers release code publicly, which is significant for reproducibility and benchmarking. This suggests a maturing experimental culture around GRAZPEDWRI-DX, although the coexistence of different train/validation/test splits means that reproducibility does not automatically imply strict comparability (Ju et al., 2023, Ahmed et al., 2024, Ju et al., 2024, Ferdi, 2024).

6. Position within the broader fracture-detection literature

Fracture-YOLO occupies the single-stage detection branch of fracture CAD. On GRAZPEDWRI-DX, this branch has been repeatedly benchmarked against two-stage and alternative detection pipelines. The comparative study of YOLOv5/6/7/8 versus Faster R-CNN concludes that the YOLO families outperform the two-stage baseline both for fracture detection and for overall multi-class detection. This is one reason later work treats real-time single-stage detection as the default design space for pediatric wrist X-ray analysis (Ahmed et al., 2024).

At the same time, the label “state-of-the-art” in this area is metric-dependent and protocol-dependent. Different papers optimize different targets: mAP@50, mAP@50-95, fracture sensitivity, or efficiency. A model can therefore be strongest on one axis and not on another. YOLOv8+SE-M3-L is reported as the best mAP@50 model in its setting, YOLOv10-M as the best mAP@50-95 model in its same-split comparison, and G-YOLOv11l as the strongest efficiency result. This suggests that Fracture-YOLO should be understood as a moving Pareto frontier rather than a single dominant detector (Ju et al., 2024, Ahmed et al., 2024, Ferdi, 2024).

A common misconception is to treat Fracture-YOLO as architecturally uniform. The surveyed papers instead show a broad methodological spectrum: baseline YOLOv8 transfer, attention-enhanced YOLOv8, global-context and feature-context excitation variants, PGI/GELAN-based YOLOv9, NMS-free YOLOv10, ghost-convolution YOLOv11, and the later explicitly named Fracture-YOLO with CRSelector and ScA. A second misconception is to equate all fracture-AI systems with YOLO-like object detectors; the weakly supervised pelvic X-ray work based on MIL and ROI mining demonstrates that fracture CAD also includes non-YOLO paradigms with different supervision assumptions and outputs (Chien et al., 2024, Chien et al., 2024, Ahmed et al., 2024, Wang et al., 2019).

Taken together, the Fracture-YOLO literature establishes pediatric wrist fracture detection as a mature object-detection benchmark with clear technical themes: preservation of weak radiographic cues, exploitation of local and global context, careful handling of class imbalance, and increasingly explicit optimization for deployment. The most robust conclusion is not that one architecture has permanently solved the task, but that YOLO-based detectors have become the central experimental platform for pediatric wrist fracture localization on GRAZPEDWRI-DX.