Enhancing Osteoporosis Detection in Chest X-Ray with a Large-Scale Vision Foundation Model: A Multi-Institutional Study

Objective(s): 

Osteoporosis often remains undiagnosed until fractures occur, causing significant individual and societal burdens. Although dual-energy X-ray absorptiometry (DXA) is the diagnostic standard, its limited accessibility hinders early detection. To address this, we developed a large-scale chest X-ray (CXR)-based foundation model for robust osteoporosis classification across multiple institutions in diverse clinical environments, aiming to enable early risk identification.

Materials and Methods: We developed a vision foundation model trained on approximately 700,000 open-access frontal CXRs using self-supervised learning. This model captured diverse radiographic features from variations in CXR equipment and institutional practices. Subsequently, we fine-tuned the foundation model using a dataset of paired CXRs and DXA T-scores to classify osteoporosis and evaluated its performance on three datasets (A-C). Datasets A and B came from the same institution, representing health screening and outpatient clinic data, respectively, while Dataset C consisted of screening data from a secondary healthcare center. We compared the performance between the foundation model and a simple CNN model, noting significant improvements.

Results: The AUCs and sensitivities using the foundation model were 0.96 and 0.83 for Dataset A, 0.87 and 0.84 for Dataset B, and 0.91 and 0.95 for Dataset C. In comparison, the performance without foundation model achieved AUCs and sensitivities of 0.96 and 0.68 for Dataset A, 0.88 and 0.68 for Dataset B, and 0.91 and 0.57 for Dataset C. While the AUCs were comparable across models, the foundation model demonstrated substantially higher sensitivities for all datasets.

Conclusion(s): By integrating a vision foundation model trained on large scale, diverse CXR data, we achieved comparable AUCs and higher sensitivities across all datasets compared to a simple CNN model, highlighting its potential to enhance robustness and generalizability in osteoporosis classification.