Multi-center validation of automated CT-based L1 vertebral Hounsfield unit.
TL;DR
For opportunistic CT screening and diagnosis, automated assessment of L1 trabecular attenuation was generalizable in a multi-center setting.
Abstract
This study assessed variation of automated CT-based L1 trabecular attenuation for opportunistic CT screening across multiple U.S.-based healthcare systems. L1 HU measurements demonstrated similar behavior among the five sites, including relative relationships according to age, sex, race, and CT scanner, and < 100 HU was generalizable for practical opportunistic CT screening and diagnosis.
Purpose
To assess variability of automated population-based L1 trabecular attenuation (HU) measurement at abdominal CT among five US-based medical systems and consider a practical 100 HU threshold for opportunistic CT-based osteoporosis screening and diagnosis.
Methods
Heterogeneous adult patient cohorts undergoing abdominal CT evaluation for any indication at five US medical systems were included for L1 vertebral body trabecular attenuation (L1 HU) assessment using a validated AI pipeline that places a fully automated ROI. Exclusion criteria included IV contrast, non-120 kVp setting, and non-physiologic outliers. All results were normalized to White patient ranges after age and sex matching. Thresholds of 100 HU and 120 HU were considered.
Results
The final multi-center cohort included 123,001 adults (51.1% women, 79.6% White). For White women age 50+, 15% had L1 attenuation values ≤ 100 HU, and 30% ≤ 120 HU. For White men, the equivalent percentages were 12% and 26%; for Black women, 7% and 17%, and for Black men, 5% and 11%. For 20-49-year-olds, only about 1% were below the L1 100 HU threshold. L1 HU measurements showed similar distributions among the five medical systems and similar relative relationships according to age, sex, race, and CT scanner. After Bonferroni correction, none of the differences among different medical systems were significant.
Conclusion
For opportunistic CT screening and diagnosis, automated assessment of L1 trabecular attenuation was generalizable in a multi-center setting. A 100 HU threshold maintains specificity with a prevalence that is age-, sex-, and race-dependent, ranging from 1% in adults under 50 to up to 15% in White women.