Deep learning opportunistic screening for osteoporosis and osteopenia using radiographs of the foot or ankle – A pilot study

Abstract

Background: The gold standard method for diagnosing low bone mineral density (BMD) is using dual-energy X-ray absorptiometry (DXA) however, most patients with low BMD are often not screened. We aimed to create a deep learning (DL) model to screen for osteoporosis/osteopenia in patients by using radiographs of the foot or ankle. Method: This is a retrospective cross-sectional study of patients aged ≥50 years who received X-rays of the foot or ankle and DXA (gold-standard) within 12 months. The 907 (White (96.7 %), Black (1.8 %) and Asian (0.4 %)) patients (3109 radiographs) were randomized (80:20) into training/validation and test datasets, and results were assessed by patient. We developed a novel DL model that extracted deep features from the radiographs of the foot and ankle using a customized architecture. The diagnostic performance of this DL model to predict if a patient had low BMD (osteopenic/osteoporotic) or normal BMD based on DXA, was evaluated with the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). Results: Mean patient age (standard deviation) was 66.9(9.0) years, and 84.6 % were female. 81.3 % and 18.7 % of patients in the training/validation dataset, and 81.5 % and 18.5 % of patients in the test datasets were osteopenic/osteoporotic and normal respectively. The DL model had an AUC of 0.87 (95% CI (0.85, 0.94), sensitivity of 89.89 %, specificity of 83.65 %, PPV of 90.78 % and NPV of 74.14 % in the test dataset. The model had an accuracy of 94.65 % in the training/validation dataset and 89.89 % in the test datasets. Conclusion: Our DL model has the potential to identify patients with osteopenia/osteoporosis using foot or ankle radiographs with high accuracy.