Radiograph-based Deep Learning Algorithm for Assisting Bone Tumor Risk Assessment: A Multicenter Study

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence in Medicine

Background:

• Accurate radiograph-based risk stratification is crucial for effective bone tumor management.
• The multicenter performance of the Bone Reporting and Data System (Bone-RADS) and the utility of deep learning (DL) decision support require further investigation.

Purpose of the Study:

• To develop and evaluate a DL model for binary risk stratification of bone tumors (benign vs. potentially malignant) using radiographs.
• To compare the DL model's performance against radiologist assessments and the Bone-RADS.

Main Methods:

• A multicenter retrospective dataset of 1777 bone tumor radiographs was used for model development.
• A prospective dataset of 152 participants was used for external validation.
• The DL-Clinic-RADS model integrated image, clinical, and semantic features using a two-view cross-attention vision transformer.

Main Results:

• The DL-Clinic-RADS achieved high performance with an AUC of 0.96 in the external set and 0.99 in the prospective set.
• The DL model significantly outperformed radiologist assessments (AUC 0.68-0.82) and Bone-RADS (AUC 0.75-0.81).
• AI assistance improved mean reader AUC by 0.07, enhanced interreader agreement, and reduced reading time by 3 seconds.

Conclusions:

• The DL-Clinic-RADS model demonstrates superior performance in radiographic risk stratification of bone tumors compared to current methods.
• The DL model significantly improves clinical decision-making and workflow efficiency in bone tumor management.
• This AI-driven approach holds promise for advancing bone tumor diagnosis and patient care.