Deep Learning Approach for Classification of Osteoporosis and Osteopenia on Chest X-ray with a Multinational Study

Purpose: 

We aimed to determine the feasibility and performance of deep learning (DL) models based on CXR with screening osteoporosis and osteopenia with diverse external validation.

Materials and Methods: 

Our DL model assesses the bone status of patients on CXR to predict whether it is normal, osteopenia, or osteoporosis. We initially constructed the fundamental training set using the radiographs of 55,600 patients (54.3% men; mean age, 55.38 ± 7.28 years) from the tertiary university hospital A. Besides, to obtain bone information from other multi-ethnic, we collected 17,577 unlabeled radiographs (57.1% men; mean age, 62.08 ± 8.74 years) from several public datasets. The unlabeled radiographs were appended pseudo labels via semi-supervised learning to achieve global robustness and generalizability. Total radiographs of 73,177 patients were utilized to train the model, with 1,989 (83.9% men; mean age, 58.7 ± 6.76) radiographs employed for internal validation. For external validation, we collected radiographs from 3 institutions, which consist of multi-ethnic/regions. Hospital B (55.5% men; mean age, 59.38 ± 7.31) is a secondary healthcare facility, the dataset of  Hospital C (56.2% men; mean age, 73.64 ± 6.74) represents diverse settings within the healthcare delivery system, and the dataset of D (2.4% men; mean age, 66.37 ± 7.27) denotes global medical platform.

Results: 

Our results of hospital A internal validation AUC scores of normal, osteopenia, and osteoporosis were 0.936, 0.891, and 0.965, respectively. In external validation, hospital B, hospital C, and platform D, AUC scores of normal were 0.911, 0.885, and 0.812, AUC scores of osteopenia were 0.845, 0.728, and 0.630, and AUC scores of osteoporosis were 0.921, 0.880, and 0.703, respectively. 

Conclusions: 

This study presents a DL model for CXR-based classification of Osteopenia and Osteoporosis via semi-supervised learning. Evaluating external multinational validation demonstrated that the proposed DL model is feasible for classification of Osteopenia and osteoporosis with CXR contrary to DXA with limited accessibility. 

Clinical relevance:

This study has the potential to demonstrate its value by facilitating opportunistic screening for osteoporosis and osteopenia patients using CXRs, which are among the most prevalent and cost-effective medical imaging techniques. Furthermore, compared to previous studies, the proposed DL model via semi-supervised learning improved global robustness and generalizability across multinational areas. In clinical practice, timely identification of osteoporosis patients is crucial for initiating appropriate medical treatment, while recognizing osteopenia allows for appropriate preventive interventions.