Pediatric radius torus fractures in x-rays-how computer vision could render lateral projections obsolete

Area of Science:

• Radiology
• Artificial Intelligence
• Pediatric Orthopedics

Background:

• Standard practice for extremity radiography, including pediatric wrist fractures, mandates at least two complementary X-ray projections.
• This approach, while standard, may lead to unnecessary radiation exposure and patient discomfort, particularly for minor injuries.

Purpose of the Study:

• To investigate the potential of computer vision, specifically convolutional neural networks (CNNs), to accurately diagnose distal radius fractures in children from single posteroanterior (PA) radiographs.
• To evaluate if CNN performance can match or exceed human expert capabilities in fracture detection, potentially eliminating the need for a second projection.

Main Methods:

• Trained three state-of-the-art CNNs on a dataset of 2,474 pediatric wrist radiographs, comprising 1,237 images of distal radius torus fractures and 1,237 normal controls.
• The CNNs were tasked with classifying images as either fractured or non-fractured.
• Evaluated CNN performance on a separate test set of 200 unseen images (100 fractured, 100 normal) using area under the curve (AUC) metrics.

Main Results:

• CNN predictions achieved a maximum AUC of 98% (95% CI 96.6%-99.5%), demonstrating high accuracy in classifying wrist radiographs.
• The CNNs consistently outperformed human expert ratings, which yielded a mean AUC of 93.5% (95% CI 89.9%-97.2%).

Conclusions:

• CNNs show significant potential for accurately ruling out distal radius fractures in children using single PA radiographs.
• Successful implementation could allow clinicians to forgo the routine lateral projection, improving radiation hygiene and patient comfort in pediatric radiography.