Precise proximal femur fracture classification for interactive training and surgical planning

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Proximal femur fractures require accurate classification for effective treatment.
• Computer-aided diagnosis (CAD) tools can assist in medical image analysis.
• Deep learning models offer potential for automated fracture detection and classification.

Purpose of the Study:

• To demonstrate the feasibility of a deep learning-based CAD tool for localizing and classifying proximal femur fractures.
• To classify fractures using the AO classification system on X-ray images.
• To improve patient treatment planning and support trauma surgeon resident training.

Main Methods:

• A dataset of 1347 radiographic studies was utilized.
• Fractures were annotated with bounding boxes and classified using the AO standard.
• ResNet-50 and AlexNet architectures were employed for classification and localization.
• The dataset was split into training (70%), validation (10%), and testing (20%) sets.

Main Results:

• The CAD tool achieved an 87% F1-score and 0.95 AUC for classifying fracture types A, B, and non-fractured.
• Classification of fractured versus non-fractured cases reached 94% accuracy and 0.98 AUC.
• The localization model achieved 100% accuracy in predicting the region of interest within manually provided bounding boxes.

Conclusions:

• The developed CAD system accurately localizes, detects, and classifies proximal femur fractures.
• Performance is comparable to expert-level and state-of-the-art results.
• The auxiliary localization model demonstrated high accuracy, supporting clinical adoption and use cases like image retrieval.