Automatic Measurement of Subregional Vertebral Bone Mineral Density via Deep Learning of Quantitative Computed Tomography Images

(Pages 1-11)
Chentian Li1,2,3, Chi Ma1,2, Xianglong Zhuo4, Wei Wang1, Li Li2,4, Wing-Yuk Ip2, Bing Li4, Tao Li4, Songjian Li3, Feng Zhu1,2 and William W. Lu1,2,5

1Department of Orthopaedics, & Department of Radiology, the University of Hong Kong-Shenzhen Hospital, Shenzhen, Guangdong, P.R. China; 2Department of Orthopaedics & Traumatology, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong SAR, P.R. China; 3Department of Orthopaedics, Zhujiang Hospital, Southern Medical University, Guangzhou, Guangdong, P.R. China; 4Department of Orthopaedics, & Department of Radiology, Liuzhou Worker’s Hospital, Guangxi Medical University, Liuzhou, Guangxi, P.R. China; 5Center for Human Tissues and Organs Degeneration, Shenzhen Institutes of Advanced 5 Technology, Chinese Academy of Science, Shenzhen, Guangdong, P.R. China

DOI: http://dx.doi.org/10.12974/2313-0954.2020.07.1

Download

Abstract: Background: Measurement of subregional Bone Mineral Density (BMD) of the vertebral body has been shown to hold a critical role in osteoporotic fracture risk analysis. The reproducibility and precision of the measurement rely highly on the vertebral body region of interest segmentation accuracy, which requires expert-level experience in medical image preprocessing and is time-consuming work. The establishment of a reliable automatic method could enhance the efficiency and precision of these measurements in clinical practice.
Purpose: To develop and validate a deep learning-based segmentation approach for subregional vertebral BMD measurement with quantitative CT scans.
Materials and Methods: Quantitative CT images from 115 subjects (62 women and 53 men with a mean age of 66.4 ± 13.4 years) were retrospectively collected. A deep learning-based segmentation pipeline was trained on a total of 403 manual segmented lumbar vertebral bodies. The performance was evaluated by its accuracy, Dice Score, and Intersection over Union (IoU) score. A scan-rescan test was performed to evaluate the subregional BMD measurement reliability and reproducibility by analyzing the intraclass correlation coefficient and Bland-Altman analysis.
Results: This automatic approach achieved high segmentation performance for the entire vertebral body segmentation (accuracy 0.98 ± 0.02, dice coefficient 0.92 ± 0.06, and IoU 0.87 ± 0.09), cortical bone segmentation (accuracy 0.95 ± 0.02, dice coefficient 0.92 ± 0.03, and IoU 0.85 ± 0.05), and endplate segmentation (accuracy 0.89 ± 0.05 and Dice coefficient 0.75 ± 0.09, IoU 0.61 ± 0.12). The scan-rescan test further showed the automatic measurement is highly reproducible (r = 0.96, limit of agreement [LoA] = −20.4~17.9 mg/cm3 for entire region; r = 0.95, LoA = −39.5~33.3 mg/cm3 for cortical region; r = 0.89, LoA = −23.4~20.9 mg/cm3 for cancellous region; r = 0.82, LoA = −44.9~58.9 mg/cm3 for superior endplate; r = 0.63, LoA = −81.6~106.5 mg/cm3, respectively).
Conclusion: The deep learning-based approach is feasible for vertebral body subregions segmentation, which ensures the precision and reproducibility of BMD measurement. The cortical and cancellous BMD can be separately measured by the deep learning-based approach, providing an automatic and reliable framework for the investigation of subregional osteoporosis changes with Quantitative Computed Tomography (QCT) spine scans.

Keywords: Bone mineral density, Spine imaging, Deep learning, Atlas-based segmentation, Quantitative computed tomography.