A large-scale vision foundation model for musculoskeletal radiographs

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Current AI models for musculoskeletal radiographs are often task-specific and require extensive annotations.
• Limited diversity and size of public datasets hinder the development of comprehensive AI models.
• A need exists for adaptable AI frameworks capable of analyzing a wide range of musculoskeletal conditions and anatomical sites.

Purpose of the Study:

• To develop and evaluate SKELEX, a large-scale foundation model for musculoskeletal radiographs using self-supervised learning.
• To assess the model's performance on various downstream diagnostic tasks.
• To demonstrate the model's capability for unsupervised anomaly localization and its potential for clinical translation.

Main Methods:

• Trained a foundation model (SKELEX) on 1.2 million diverse musculoskeletal radiographs using self-supervised learning.
• Evaluated the model on 12 downstream diagnostic tasks, including fracture detection, osteoarthritis grading, and bone tumor classification.
• Utilized reconstruction-based anomaly localization for identifying pathologic regions and developed an interpretable bone tumor classifier.

Main Results:

• SKELEX outperformed baseline models in fracture detection, osteoarthritis grading, and bone tumor classification.
• The model demonstrated unsupervised anomaly localization, generating error maps to pinpoint pathologic areas.
• An interpretable bone tumor classifier built on SKELEX showed robust performance on external datasets and was deployed as a web application.

Conclusions:

• SKELEX offers a scalable, label-efficient, and broadly applicable AI framework for musculoskeletal radiographs.
• The model's unsupervised anomaly detection and classification capabilities show significant potential for clinical translation.
• The developed bone tumor classifier serves as a proof of concept for real-world AI implementation in musculoskeletal radiology.