Deep learning models in classifying primary bone tumors and bone infections based on radiographs

Area of Science:

• Radiology
• Oncology
• Infectious Diseases
• Artificial Intelligence

Background:

• Primary bone tumors (PBTs) and bone infections share similar radiographic features, complicating diagnosis.
• Accurate differentiation is crucial for timely and appropriate patient management.

Purpose of the Study:

• To develop and evaluate an ensemble deep learning framework for distinguishing PBTs from bone infections.
• To compare the ensemble model's performance against standalone radiograph-based imaging models and human radiologists.

Main Methods:

• An ensemble deep learning model was created integrating multicenter radiographs and clinical data.
• The model was trained and validated on internal datasets and tested on an external dataset.
• Performance was compared to EfficientNet B3/B4, Vision Transformer, and Swin Transformer models.

Main Results:

• The ensemble model achieved high diagnostic performance with AUCs of 0.948 (internal) and 0.963 (external).
• Achieved accuracies were 0.881 (internal) and 0.895 (external), surpassing individual imaging models.
• The model's accuracy exceeded junior and mid-level radiologists and was comparable to senior radiologists.

Conclusions:

• Ensemble deep learning offers a powerful tool for enhancing diagnostic precision in differentiating PBTs and bone infections.
• This AI framework has the potential to support clinical decision-making and improve patient outcomes.