Deep Learning Method for Normal, Osteopenia and Osteoporosis Classification on Chest X-rays: A Multinational Investigation

Osteopenia and osteoporosis are prevalent chronic metabolic bone conditions, with the global number of patients on the rise, posing a significant risk of fractures. Our objective was to assess the viability and efficacy of deep learning (DL) models utilizing chest X-rays (CXRs) for screening normal, osteopenia and osteoporosis, leveraging a multinational dataset.

Our DL model evaluates the bone condition of patients on CXR to predict whether it falls within the categories of normal, osteopenia, or osteoporosis. Initially, we established the core training dataset using CXRs from 55,600 patients (54.3% men; mean age, 55.38 ± 7.28 years) obtained from tertiary university hospital A. Additionally, to incorporate bone information from diverse ethnic backgrounds, we gathered 17,577 unlabeled CXRs (57.1% men; mean age, 62.08 ± 8.74 years) from various regions and institutions. Through semi-supervised learning, these unlabeled CXRs were augmented with pseudo labels, enhancing the model's global robustness and generalizability. In total, 73,177 CXRs from patients were utilized for training the model, with 1,989 CXRs (83.9% men; mean age, 58.7 ± 6.76 years) reserved for internal validation. For external validation, CXRs were collected from three institutions representing diverse demographics and healthcare settings. Hospital B (55.5% men; mean age, 59.38 ± 7.31 years) serves as a secondary healthcare facility, the dataset from Hospital C (56.2% men; mean age, 73.64 ± 6.74 years) encompasses varied settings within the healthcare delivery system, and the dataset from D (2.4% men; mean age, 66.37 ± 7.27 years) represents a global medical platform.

In the internal validation at hospital A, our model achieved AUC scores of 0.936 for normal, 0.891 for osteopenia, and 0.965 for osteoporosis. During external validation, at hospital B, the AUC scores were 0.911 for normal, 0.845 for osteopenia, and 0.921 for osteoporosis. At hospital C, the AUC scores were 0.885 for normal, 0.728 for osteopenia, and 0.880 for osteoporosis. For platform D, the AUC scores were 0.812 for normal, 0.630 for osteopenia, and 0.703 for osteoporosis.

At the end of this investigation,  we introduced the DL model for classifying normal, osteopenia and osteoporosis based on CXR using semi-supervised learning. External multinational validation demonstrated that our DL model successfully supports the classification of normal, osteopenia and osteoporosis using CXRs. Furthermore, this DL model indicated significant value by enabling opportunistic screening of osteoporosis and osteopenia patients using CXRs, a widely available and cost-effective imaging modality. With this DL model, timely identification of osteoporosis patients in clinical settings allows for prompt medical treatment, while detecting osteopenia facilitates effective implementation of preventive interventions.