Deep Learning-Based Multi-Class Pediatric Wrist Fracture Subtype Classification: A Pilot Study Comparing...

Area of Science:

• Radiology
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Pediatric wrist fractures are common, with subtypes impacting treatment and outcomes.
• Accurate classification requires specialized radiographic expertise, often unavailable in resource-limited settings.
• Deep learning (DL) offers potential for automated fracture classification.

Purpose of the Study:

• To evaluate a convolutional neural network (CNN) for classifying five types of pediatric wrist fractures.
• To establish a benchmark for DL models in pediatric fracture diagnosis.

Main Methods:

• Utilized the GRAZPEDWRI-DX dataset of 940 pediatric wrist radiographs.
• Applied contrast-limited adaptive histogram equalization (CLAHE) and resizing for image preprocessing.
• Trained three ImageNet-pretrained CNN architectures (DenseNet-169, ResNet-50, EfficientNet-B4) using transfer learning.
• Assessed performance using balanced accuracy, macro F1, macro AUROC, and Cohen's kappa.

Main Results:

• DenseNet-169 achieved the highest performance under initial pilot training.
• Extended training to 50 epochs significantly improved DenseNet-169's balanced accuracy to 0.532 and macro AUROC to 0.815.
• Model sensitivity was highest for no-fracture detection (0.969) and lowest for buckle/torus fractures (0.393).
• Gradient-weighted class activation mapping (Grad-CAM) indicated anatomically relevant feature focus.

Conclusions:

• DenseNet-169 demonstrated the best performance among evaluated CNNs for pediatric wrist fracture classification.
• Extended training improved DL model accuracy, but it remains below clinically usable thresholds.
• The study established a reproducible DL pipeline and benchmark for future research, not a clinical tool.