A Novel Machine Learning-Based Semi-Automated Phantom-Less QCT Model for Osteoporosis Screening on 100 kVp
Miao Wei1, Mian Huang2, Jianjun Wu2
1Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China (M.W., J.P., W.L., W.Z., F.L.).
Rationale And Objectives:
Opportunistic osteoporosis screening using chest CT is increasingly explored, yet conventional QCT models are calibrated at 120 kVp and may be inaccurate for ultra-low-dose scans acquired at lower tube voltages. This study aimed to develop and validate a machine learning-assisted phantom-less QCT (PL-QCT) model for BMD quantification at 100 kVp and assess its diagnostic performance.
Methods:
Sixteen repeated European Spine Phantom (ESP) scans were acquired using a 100 kVp ultra-low-dose chest CT protocol. A total of 508 patients were retrospectively included for model training and internal validation. A 100 kVp PL-QCT model was developed, calibrated against ESP reference values, and compared with a conventional 120 kVp QCT model. External validation was performed on an independent CT system using ESP scans and 197 patients under a 100 kVp chest CT protocol. Diagnostic performance was evaluated in 178 individuals with both DXA and chest CT.
Results:
The mean effective dose was 0.82±0.19 mSv. In internal validation, the 100 kVp model showed significantly lower BMD error than the 120 kVp model (2.39±7.12 vs 16.68±8.26 mg/cm³, p<0.0001), with improved accuracy across L1-L3. External validation confirmed lower error for the 100 kVp model (-0.11±4.39 vs 13.78±5.10 mg/cm³). Agreement with DXA was 88.8%.
Conclusion:
The 100 kVp machine learning-assisted PL-QCT model enables accurate BMD quantification from ultra-low-dose chest CT, outperforming conventional 120 kVp models and supporting reliable cross-scanner opportunistic osteoporosis screening.
