Deep learning detection of subtle fractures using staged algorithms to mimic radiologist search pattern.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Subtle fractures like triquetral avulsion and Segond fractures are often missed on radiographs.
• Deep convolutional neural networks (CNNs) show promise in medical image analysis.

Purpose of the Study:

• To develop and evaluate a two-stage deep CNN system for detecting triquetral avulsion and Segond fractures.
• To mimic a radiologist's focused search pattern for improved fracture detection.

Main Methods:

• A two-stage CNN system was developed: object detectors for region cropping, followed by classifiers for fracture detection.
• The system was trained and validated on wrist and knee radiographs from public datasets.
• Gradient-class activation mapping was used for interpretability.

Main Results:

• The two-stage system achieved high area under the receiver operating characteristic curve values (0.959 for triquetral, 0.989 for Segond fractures).
• Accuracy for the two-stage system was 90.8% for triquetral and 92.5% for Segond fractures.
• Performance significantly outperformed a one-stage classifier.

Conclusions:

• A two-stage deep learning pipeline enhances the accuracy of detecting subtle fractures compared to a one-stage approach.
• The system demonstrated good generalizability to external test data.
• Focused attention on specific image regions improves the detection of subtle radiographic findings.