Abstract
Wrist fractures are highly prevalent among children and can significantly impact their daily activities, such as attending school, participating in sports, and performing basic self-care tasks. If not treated properly, these fractures can result in chronic pain, reduced wrist functionality, and other long-term complications. Recently, advancements in object detection have shown promise in enhancing fracture detection, with systems achieving accuracy comparable to, or even surpassing, that of human radiologists.
The YOLO series, in particular, has demonstrated notable success in this domain. This study is the first to provide a thorough evaluation of various YOLOv10 variants to assess their performance in detecting pediatric wrist fractures using the GRAZPEDWRI-DX dataset. It investigates how changes in model complexity, scaling the architecture, and implementing a dual-label assignment strategy can enhance detection performance.
Key Achievement: Our trained model achieved mean average precision (mAP@50-95) of 51.9% surpassing the current YOLOv9 benchmark of 43.3% on this dataset. This represents an improvement of 8.6%.
YOLOv10 Architecture
.png)
YOLOv10 Model Architecture: Overview of the YOLOv10 architecture used for pediatric wrist fracture detection
Key Architecture Features:
- Advanced Backbone: Improved feature extraction capabilities
- Dual Label Assignment: Enhanced training strategy for better performance
- Multiple Variants: N, S, M, L, X variants with different complexity levels
- Optimized Detection Head: Specialized for medical imaging applications
Performance Comparison: YOLOv9 vs YOLOv10
Comprehensive Evaluation: Our study provides the first comprehensive evaluation of YOLOv10 variants on pediatric wrist fracture detection, demonstrating significant improvements over YOLOv9 benchmarks.
| Variant | mAP@50 (%) | mAP@50-95 (%) | F1 (%) | Params (M) | FLOPs (G) |
|---|---|---|---|---|---|
| YOLOv9 Baselines | |||||
| YOLOv9-C | 65.3 | 42.7 | 64.0 | 51.0 | 239.0 |
| YOLOv9-E | 65.5 | 43.3 | 64.0 | 69.4 | 244.9 |
| YOLOv9-C' | 66.2 | 45.2 | 66.7 | 25.3 | 102.4 |
| YOLOv9-E' | 67.0 | 44.9 | 70.9 | 57.4 | 189.2 |
| YOLOv10 Results (Our Work) | |||||
| YOLOv10-N | 59.5 | 39.1 | 63.0 | 2.7 | 8.2 |
| YOLOv10-S | 76.1 | 51.7 | 67.5 | 8.0 | 24.5 |
| YOLOv10-M | 75.9 | 51.9 | 67.8 | 16.5 | 63.5 |
| YOLOv10-L | 70.9 | 46.6 | 68.7 | 25.7 | 126.4 |
| YOLOv10-X | 76.2 | 48.2 | 69.8 | 31.6 | 169.9 |
Key Performance Highlights:
🎯 Best mAP@50-95
YOLOv10-M: 51.9%
+8.6% improvement over YOLOv9-E
⚡ Most Efficient
YOLOv10-N: 2.7M params
Extremely lightweight for deployment
🏆 Best mAP@50
YOLOv10-X: 76.2%
Superior detection accuracy
⚖️ Best Balance
YOLOv10-S: 76.1% mAP@50
Optimal performance-efficiency trade-off
Dataset & Methodology
GRAZPEDWRI-DX Dataset:
📊 Dataset Split
Training: 15,245 images (75%)
Validation: 4,066 images (20%)
Testing: 1,016 images (5%)
🏷️ Detection Classes
9 Classes:
Fracture, Bone Anomaly, Metal, Foreign Body, Soft Tissue, etc.
🔬 Medical Focus
Pediatric Wrist X-rays
Specialized dataset for child healthcare
📈 Dual Label Assignment
Novel Training Strategy
Enhanced label assignment for better performance
Training Configuration:
- Input Size: 640×640 pixels
- Batch Size: 32
- Epochs: 100
- Workers: 8 (data loading threads)
- Hardware: NVIDIA GPU + CUDA CuDNN
- Environment: Python 3.12, PyTorch 2.3
Research Impact & Applications
Clinical Significance: This research addresses a critical healthcare need in pediatric medicine, where accurate and rapid fracture detection can significantly impact treatment outcomes and long-term child development.
🏥 Clinical Applications
- Emergency department screening
- Radiologist assistance tools
- Telemedicine support
- Second opinion systems
⚡ Performance Benefits
- Real-time detection capability
- Reduced diagnostic time
- Improved accuracy over previous methods
- Scalable deployment options
🌟 Innovation Highlights
- First comprehensive YOLOv10 evaluation for fractures
- Dual label assignment strategy
- Multiple model variants analyzed
- Open-source implementation
📊 Papers with Code Rankings
- State-of-the-art on GRAZPEDWRI-DX
- Fracture detection leaderboard
- Object detection benchmark
- Medical imaging applications
Implementation & Usage
Easy to Use: Our implementation provides a simple interface for both research and clinical applications. The system includes pre-trained weights and a Gradio-based web interface for immediate deployment.
Quick Start:
# Install dependencies
pip install -r requirements.txt
# Train a model
from ultralytics import YOLO
model = YOLO("yolov10x.pt")
results = model.train(data='dataset/meta.yaml', epochs=100, imgsz=640, batch=32, name='x')
# Run inference
python src/app.pyFuture Directions
This work establishes a new baseline for automated pediatric fracture detection and opens several avenues for future research and clinical implementation:
- Multi-modal Integration: Combining X-ray data with clinical information and patient history
- Real-time Clinical Deployment: Integration with hospital PACS systems and clinical workflows
- Extended Anatomical Coverage: Expansion to other pediatric bone fracture types and anatomical regions
- Uncertainty Quantification: Adding confidence estimation and uncertainty measures for clinical decision support
- Federated Learning: Collaborative training across multiple medical institutions while preserving privacy
- Mobile Applications: Lightweight models for point-of-care diagnosis on mobile devices
Citation & Contact
Authors
Ammar Ahmed - Lead Author
Abdul Manaf - Co-Author
Publication: arXiv:2407.15689 [eess.IV]
Submitted: July 22, 2024
Last Revised: July 31, 2024
If you find our work useful in your research, please consider citing our paper. This research contributes to advancing AI-assisted medical diagnosis in pediatric healthcare.