Utilizing heat maps as explainable artificial intelligence for detecting abnormalities on wrist and elbow radiographs

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Radiographs of the wrist and elbow are crucial for diagnosing fractures and degenerative conditions.
• Interpreting these images can be challenging due to complex anatomy and subtle signs.
• Artificial intelligence (AI) and deep learning show promise in improving diagnostic accuracy.

Purpose of the Study:

• To evaluate the performance of transfer-learning models in diagnosing conditions from wrist and elbow radiographs.
• To assess the interpretability of AI models using Gradient-weighted Class Activation Mapping (Grad-CAM).

Main Methods:

• Utilized the MURA-dataset containing musculoskeletal radiographs of wrists and elbows.
• Applied an ensemble of twenty transfer-learning models (e.g., VGG16, ResNet, DenseNet).
• Implemented Grad-CAM for visualizing AI decision-making processes and calculated Dice Similarity Coefficient (DSC).

Main Results:

• Average test accuracies were 0.81 for wrist and 0.60 for elbow radiographs.
• VGG16 achieved the highest accuracy (0.84) for wrist, and DenseNet169 (0.73) for elbow.
• Model agreement was higher for radiographs with metal implants than for those with fractures.

Conclusions:

• Transfer-learning models demonstrate potential for enhancing diagnostic accuracy in wrist and elbow radiography.
• Grad-CAM provides valuable insights into model performance and areas for improvement.
• AI-driven diagnostics, with explainable methods, can boost clinical trust and adoption.