Robust Multi-View Fracture Detection in the Presence of Other Abnormalities Using HAMIL-Net

Area of Science:

• Orthopedic imaging analysis
• Artificial intelligence in medicine
• Deep learning for medical diagnosis

Background:

• Foot and ankle fractures are a common military health concern.
• Automated diagnosis systems can enhance efficiency and resource allocation.
• Distinguishing fractures from other pathologies is critical for accurate diagnosis.

Purpose of the Study:

• To evaluate the performance of HAMIL-Net for detecting foot and ankle fractures.
• To assess HAMIL-Net's robustness in the presence of other orthopedic abnormalities.
• To improve automated diagnosis for musculoskeletal imaging studies.

Main Methods:

• Utilized HAMIL-Net, a deep neural network with hierarchical attention and multiple-instance learning.
• Trained the model on a large dataset of 148K musculoskeletal imaging studies from 51K Veterans.
• Employed a semi-automated annotation pipeline using radiology reports and natural language processing.

Main Results:

• HAMIL-Net achieved an area under the receiver operational curve of 0.87 for fracture detection.
• Performance improved when integrating a fracture-specific model with a general abnormality detection model (accuracy from 0.53 to 0.77, F-score from 0.46 to 0.86).
• Evaluated performance across different patient age groups, including young adults (18-35).

Conclusions:

• Automated fracture detection requires consideration of co-existing abnormalities for clinical deployment.
• The enhanced HAMIL-Net demonstrated improved fracture detection and identification of incidental musculoskeletal findings.
• This AI approach holds potential for improving the diagnosis of foot and ankle injuries in clinical settings.