Deep Learning-based Automated Opportunistic Osteoporosis Screening Using Chest LDCT and Lumbar CT: A Multicenter Cohort Study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Peking University Third Hospital, 49 Huayuan N Rd, Haidian District, Beijing, China (Y.L., Y.Z., Y.Z., S.L., D.J., C.J., M.N., H.Y.).
- The Institute of Intelligent Diagnostics, Beijing United-Imaging Research Institute of Intelligent Imaging, Building 3-4F, 9 Yongteng N. Road, Beijing 100080 China (M.Z., Z.Q.).
- Department of Radiology, the First Affiliated Hospital of Zhengzhou University, 1 Jianshe East Road, Zhengzhou 450052 China (J.H., Y.W.).
- Department of Radiology, Hebei Medical University Third Hospital, 139 Ziqiang Rd, Qiaoxi District, Shijiazhuang, Shijiazhuang, Hebei, China, 050052 China (L.G., J.Z.).
- Department of Radiology, Peking University Third Hospital, 49 Huayuan N Rd, Haidian District, Beijing, China (Y.L., Y.Z., Y.Z., S.L., D.J., C.J., M.N., H.Y.). Electronic address: [email protected].
Abstract
To propose a deep learning (DL) method for automatically measuring bone mineral density (BMD) and validating its diagnostic performance for opportunistic osteoporosis screening using low-dose chest computed tomography (LDCT) and lumbar CT scans across CT scanners from different manufacturers and hospitals. This retrospective study involved 4305 patients who underwent chest LDCT and lumbar CT scans at five hospitals. Patient data were collected from nine CT scanners and divided into training, validation, and test sets 1-5 (1891, 806, 229, 418, 508, 229, and 224). Four convolutional neural networks (CNNs) were employed for automated vertebral body (VB) segmentation (3D VB-Net and SCN), region of interest extraction (3D VB-Net), and BMD calculation (DenseNet and ResNet) of the target VBs (T12-L2). The BMD values derived from quantitative CT (QCT) were identified as the reference standard. Linear regression and Bland-Altman analyses were used to compare BMD values between 120 kV QCT and 120 kV CNNs. Receiver operating characteristic curve analysis was performed to assess the diagnostic accuracy of the 120 kV CNNs in detecting osteoporosis and low BMD from normal BMD. The BMD values of 120 kV CNNs showed an excellent correlation (R² = 0.970-0.997, P < 0.001) and good agreement (mean error, 0.57-2.45 mg/cm³; 95% limits of agreement, -3.70-5.68 mg/cm³) with 120 kV QCT. The areas under the curve of the 120 kV CNNs in diagnosing osteoporosis and low BMD from normal BMD were 0.993-1.000 and 0.994-1.000, with a sensitivity of 96.55-100.00% and 92.11-100.00%, and a specificity of 92.75-100.00% and 93.83-100.00%, respectively. This DL method achieved high diagnostic performance for automatic osteoporosis screening using chest LDCT and lumbar CT scans and performed well across CT scanners from different manufacturers and hospitals.