Deep learning approach based on a patch residual for pediatric supracondylar subtle fracture detection

Area of Science:

• Pediatric Orthopedics
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Supracondylar humerus fractures are common pediatric elbow injuries.
• Diagnosis is challenging due to pediatric skeletal anatomy and limited labeled fracture data.
• Deep learning models require large datasets, which are scarce for this condition.

Purpose of the Study:

• To develop a deep learning model for automatic detection and localization of pediatric supracondylar fractures.
• To address the challenge of limited labeled data in pediatric fracture diagnosis.
• To improve diagnostic accuracy and efficiency in identifying these fractures.

Main Methods:

• Introduction of a multi-scale patch residual network (MPR) framework.
• Integration of a convolutional neural network (CNN) for feature extraction.
• Utilizing a multi-scale generative adversarial network trained on healthy bone images to model skeletal integrity.

Main Results:

• The MPR model achieved 90.5% accuracy, 89% sensitivity, 92% specificity, and a 0.906 F1 score on an independent test set.
• Performance surpassed emergency medicine physicians and neared pediatric radiologists' accuracy.
• Demonstrated a fast inference speed of 1.1 seconds per image sheet.

Conclusions:

• The proposed deep learning approach effectively detects pediatric supracondylar humerus fractures with high accuracy.
• Leveraging healthy images significantly reduces reliance on labeled fracture data.
• The model shows strong potential for clinical application in improving pediatric fracture diagnosis.