Opportunistic Osteoporosis Screening on Chest radiographs with deep learning methods

Abstract data (Mention the category of your abstract)

Cat 1. Clinical

Cat 2. Basic Science

Cat 3. Translational

Abstract Details (limited to 1,950 characters (About 300-350 words)

Recently, deep learning algorithms have been widely recognized as an outperforming and reliable approach to identify clinically useful features directly from medical images. Chest radiography is the most widely available medical imaging modality, providing a wealth of information regarding the cardiovascular and musculoskeletal systems. Selecting the high-risk group for osteoporosis assessment with chest radiographs (CXRs) is worth exploring using deep learning. We aimed to develop and evaluate deep learning approaches for screening osteoporosis using classified CXRs based on central dual-energy X-ray absorptiometry (DXA)-diagnosed osteoporosis. We collected paired data consisting of CXRs and DXA results from four datasets across two hospitals: training, internal, and a weak external test dataset from Asan Medical Center (AMC), and an external test dataset from Veterans Health Service Medical Center (VHSMC), representing diverse settings within the healthcare delivery system. The number of cases in the AMC training dataset and internal test dataset was 11,037 and 1,989 cases, respectively, while the weak external test set comprised 1,087 cases and the VHSMC test dataset had 937 cases. This model was exclusively trained on the AMC training dataset. All participants in this study were over 50 years old, with 83.86%, 88.98%, and 43.86% being women in the internal, weak external, and VHSMC test datasets, respectively. The distributions of osteoporosis versus normal and osteopenia were 33.33%, 29.20%, and 24.87%, respectively. The interval between DXA and CXR was 0 days in the internal dataset, 90 days in the weak external dataset, and 180 days in the VHSMC dataset. The model output was a continuous value between 0 and 1, and its performance metric used the concordance index (C-index) for predicting osteoporosis. The C-index for osteoporosis was 0.939, 0.742, and 0.772 in the internal, weak external, and VHSMC external test datasets, respectively. We developed the osteoporosis risk assessment model and validated it with two external test datasets. The findings strongly support the potential use of this model on CXRs for opportunistic and automated osteoporosis screening in real clinical scenarios.