Automated Opportunistic Trabecular Volumetric Bone Mineral Density Extraction Outperforms Manual Measurements for the Prediction of Vertebral Fractures in Routine CT

Sophia S. Goller, Jon Rischewski, Thomas Liebig, Jens Ricke, Sebastian Siller, Vanessa F. Schmidt, Robert Stahl, Julian Kulozik, Thomas Baum, Jan S. Kirschke, Sarah C. Foreman, Alexandra S. GersingDiagnostics 2023 · DOI

TL;DR

This paper introduces an automated method using Convolutional Neural Networks (CNNs) to extract volumetric bone mineral density (vBMD) from routine CT scans, demonstrating its superior ability to predict future vertebral fractures compared to manual vBMD measurements. This advancement offers a cost-effective and radiation-free way to screen for osteoporosis and identify individuals at high risk of fractures.

What Problem Does This Paper Address?

Osteoporosis is a major public health concern, characterized by reduced bone mass and microarchitectural deterioration, leading to an increased risk of fractures, particularly vertebral fractures (VFs). VFs can significantly impair quality of life and increase the risk of subsequent fractures. While Dual-energy X-ray absorptiometry (DXA) is the standard for osteoporosis screening, it has limitations, and only a fraction of individuals with fractures exhibit low DXA-measured bone density. Quantitative Computed Tomography (QCT) offers a more accurate measure of volumetric bone mineral density (vBMD), but dedicated QCT scans require additional radiation and cost. This paper addresses the challenge of leveraging existing routine CT scans for opportunistic osteoporosis screening by developing and validating an automated method for vBMD extraction and its predictive power for incident VFs.

Methods

The study employed a deep learning approach using a CNN-based framework (Anduin) for automated segmentation and labeling of vertebrae from routine CT scans of the thoracolumbar spine. This automated pipeline then extracted trabecular vBMD using asynchronous calibration and correction for contrast medium. For comparison, manual vBMD measurements were obtained by two radiologists by placing volumetric regions of interest in the anterior trabecular region of vertebrae L1-L4. The study then compared the predictive performance of both automated and manual vBMD measurements for the occurrence of new incidental VFs at a 1.5-year follow-up. Statistical analysis involved multivariable linear regression models, adjusting for age and sex, to assess the association between vBMD and fracture status.

Datasets

The study utilized a retrospective dataset of 157 patients (84 females, 73 males) who underwent at least two CT scans of the thoracolumbar spine between August 2009 and April 2021. Patients were categorized into two groups: those who developed at least one new incidental osteoporotic VF at the 1.5-year follow-up (n=53) and a frequency-matched group of patients who did not develop an incidental VF (n=104). The CT data were acquired on six different MDCT scanners from GE and Siemens Healthineers. No new benchmark datasets were introduced; the study used existing clinical data.

Key Insights and Ideas

  • Automated vBMD is a powerful predictor: The core insight is that automated vBMD extraction from routine CT scans is not only feasible but also significantly better at predicting future vertebral fractures than manual measurements. This suggests a paradigm shift towards leveraging existing imaging data for proactive bone health assessment.
  • CNNs excel at anatomical segmentation: The success of the automated method hinges on the CNN's ability to accurately segment vertebrae, even in routine scans acquired for other clinical indications. This highlights the power of deep learning in extracting clinically relevant information from diverse imaging sources.
  • Opportunistic screening is viable: The study strongly supports the concept of opportunistic osteoporosis screening using readily available CT scans, eliminating the need for additional radiation exposure and costs associated with dedicated bone density scans.
  • Trabecular vBMD is key: The focus on trabecular vBMD, rather than cortical bone, is crucial, as trabecular bone is more metabolically active and may be a better indicator of bone health and fracture risk.

Key Results

  • Patients who developed incidental VFs at follow-up had significantly lower automatically extracted baseline vBMD (83.6 ± 29.4 mg/cm³) compared to those without VFs (102.1 ± 27.7 mg/cm³, p < 0.001).
  • Manual vBMD measurements did not show a significant difference between the VF and no-VF groups (99.2 ± 37.6 mg/cm³ vs. 107.9 ± 33.9 mg/cm³, p = 0.30).
  • After adjusting for age and sex, automated vBMD showed a stronger association with incidental VFs (β = -0.32, p < 0.001) compared to manual vBMD (β = -0.15, p < 0.03).
  • The automated method demonstrated significant associations with VFs across all individual lumbar vertebrae (L1-L4) and combinations thereof, while the manual method showed significant associations only for L1-L2 and L2-L3.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective study design: This limits the ability to correlate laboratory parameters (e.g., Vitamin D, calcium) with vBMD.
  • Scanner variability: CT scans were acquired on different scanners with varying protocols, which, while reflecting clinical reality, could introduce variability.
  • Lack of DXA/QCT comparison: The study did not directly compare automated vBMD with DXA or dedicated QCT measurements.

Future work could involve:

  • Prospective studies: To further validate the findings and establish causal relationships.
  • Correlation with lab parameters: Investigating the relationship between automated vBMD and biochemical markers of bone health.
  • Application to pediatric cohorts: Exploring the utility of automated vBMD in younger populations at risk for conditions like scoliosis.
  • Integration into clinical workflows: Developing seamless integration of automated vBMD assessment into routine CT reporting systems.

Summary updated

Abstract

Opportunistic osteoporosis screening using multidetector CT-scans (MDCT) and convolutional neural network (CNN)-derived segmentations of the spine to generate volumetric bone mineral density (vBMD) bears the potential to improve incidental osteoporotic vertebral fracture (VF) prediction. However, the performance compared to the established manual opportunistic vBMD measures remains unclear. Hence, we investigated patients with a routine MDCT of the spine who had developed a new osteoporotic incidental VF and frequency matched to patients without incidental VFs as assessed on follow-up MDCT images after 1.5 years. Automated vBMD was generated using CNN-generated segmentation masks and asynchronous calibration. Additionally, manual vBMD was sampled by two radiologists. Automated vBMD measurements in patients with incidental VFs at 1.5-years follow-up (<i>n</i> = 53) were significantly lower compared to patients without incidental VFs (<i>n</i> = 104) (83.6 ± 29.4 mg/cm<sup>3</sup> vs. 102.1 ± 27.7 mg/cm<sup>3</sup>, <i>p</i> < 0.001). This comparison was not significant for manually assessed vBMD (99.2 ± 37.6 mg/cm<sup>3</sup> vs. 107.9 ± 33.9 mg/cm<sup>3</sup>, <i>p</i> = 0.30). When adjusting for age and sex, both automated and manual vBMD measurements were significantly associated with incidental VFs at 1.5-year follow-up, however, the associations were stronger for automated measurements (β = -0.32; 95% confidence interval (CI): -20.10, 4.35; <i>p</i> < 0.001) compared to manual measurements (β = -0.15; 95% CI: -11.16, 5.16; <i>p</i> < 0.03). In conclusion, automated opportunistic measurements are feasible and can be useful for bone mineral density assessment in clinical routine.
BibTeX
@inproceedings{goller2023automated,
  author = {Sophia S. Goller and Jon Rischewski and Thomas Liebig and Jens Ricke and Sebastian Siller and Vanessa F. Schmidt and Robert Stahl and Julian Kulozik and Thomas Baum and Jan S. Kirschke and Sarah C. Foreman and Alexandra S. Gersing},
  title = {{Automated Opportunistic Trabecular Volumetric Bone Mineral Density Extraction Outperforms Manual Measurements for the Prediction of Vertebral Fractures in Routine CT}},
  booktitle = {Diagnostics},
  year = {2023},
  doi = {10.3390/diagnostics13122119},
}
Figure