Automated deep learning-based bone mineral density assessment for opportunistic osteoporosis screening using various CT protocols with multi-vendor scanners

Abstract

This retrospective study examined the diagnostic efficacy of automated deep learning-based bone mineral density (DL-BMD) measurements for osteoporosis screening using 422 CT datasets from four vendors in two medical centers, encompassing 159 chest, 156 abdominal, and 107 lumbar spine datasets. DL-BMD values on L1 and L2 vertebral bodies were compared with manual BMD (m-BMD) measurements using Pearson's correlation and intraclass correlation coefficients. Strong agreement was found between m-BMD and DL-BMD in total CT scans (r = 0.953, p < 0.001). The diagnostic performance of DL-BMD was assessed using receiver operating characteristic analysis for osteoporosis and low BMD by dual-energy x-ray absorptiometry (DXA) and m-BMD. Compared to DXA, DL-BMD demonstrated an AUC of 0.790 (95% CI 0.733-0.839) for low BMD and 0.769 (95% CI 0.710-0.820) for osteoporosis, with sensitivity, specificity, and accuracy of 80.8% (95% CI 74.2-86.3%), 56.3% (95% CI 43.4-68.6%), and 74.3% (95% CI 68.3-79.7%) for low BMD and 65.4% (95% CI 50.9-78.0%), 70.9% (95% CI 63.8-77.3%), and 69.7% (95% CI 63.5-75.4%) for osteoporosis, respectively. Compared to m-BMD, DL-BMD showed an AUC of 0.983 (95% CI 0.973-0.993) for low BMD and 0.972 (95% CI 0.958-0.987) for osteoporosis, with sensitivity, specificity, and accuracy of 97.3% (95% CI 94.5-98.9%), 85.2% (95% CI 78.8-90.3%), and 92.7% (95% CI 89.7-95.0%) for low BMD and 94.4% (95% CI 88.3-97.9%), 89.5% (95% CI 85.6-92.7%), and 90.8% (95% CI 87.6-93.4%) for osteoporosis, respectively. The DL-based method can provide accurate and reliable BMD assessments across diverse CT protocols and scanners.

Keywords: Bone density; Deep learning; Osteoporosis; Tomography; X-Ray computed.

Fig. 1

ROC curves of manual BMD (m-BMD) and automated BMD (DL-BMD) for diagnosis of low-BMD and osteoporosis. The total AUC values were evaluated for diagnosing low-BMD (A) and osteoporosis (B) based on the dual-energy X-ray absorptiometry (DXA)-BMD. These analyses were performed using the data from whole vendors in Guro and Ansan Hospital dataset with DXA (N = 241).

Fig. 2

ROC curves of automated BMD (DL-BMD) for diagnosis of low-BMD and osteoporosis. The total AUC values were evaluated for diagnosing low-BMD (A) and osteoporosis (B) based on the manual BMD (m-BMD).

Fig. 3

Flow diagram of the study patients.

Fig. 4

Overview of the thoracolumbar spine segmentation with two different complementary networks (A–E). The spine mask was originally labeled in a way that the labels increased one by one, such as T12 = 19, L1 = 20 and L2 = 21 (A). For T-spine segmentation, lumbar vertebrae were re-labeled as background (B), and conversely for L-spine segmentation thoracic vertebrae were re-labeled as background (C). The final object delineated by the T-spine model was identified as T12 (D), and the initial two consecutive objects in the mask were identified by the L-spine model as L1 and L2 respectively (E).