Abstract
BACKGROUND: Despite being the gold standard for diagnosing osteoporosis, dual-energy X-ray absorptiometry (DXA) is an underutilized screening tool for osteoporosis. OBJECTIVES: This study proposed and validated a controllable feature layer of a convolutional neural network (CNN) model with a preprocessing image algorithm to classify osteoporosis and predict T-score on the proximal hip region via simple hip radiographs. DESIGN: This was a single-center, retrospective study. METHODS: An image dataset of 3460 unilateral hip images from 1730 patients (age ⩾50 years) was retrospectively collected with matched DXA assessment for T-score for the targeted proximal hip regions to train (2473 unilateral hip images from 1430 patients) and test (497 unilateral hip images from 300 patients) the proposed CNN model. All images were processed with a fully automated CNN model, X1AI-Osteo. RESULTS: The proposed screening tool illustrated a better performance (sensitivity: 97.2%; specificity: 95.6%; positive predictive value: 95.7%; negative predictive value: 97.1%; area under the curve: 0.96) than the open-sourced CNN models in predicting osteoporosis. Moreover, when combining variables, including age, body mass index, and sex as features in the training metric, there was high consistency in the T-score on the targeted hip regions between the proposed CNN model and the DXA (r = 0.996, p < 0.001). CONCLUSION: The proposed CNN model may identify osteoporosis and predict T-scores on the targeted hip regions from simple hip radiographs with high accuracy, highlighting the future application for population-based opportunistic osteoporosis screening with low cost and high adaptability for a broader population at risk. TRIAL REGISTRATION: TMU-JIRB N201909036.