Development and validation of an AI framework for automated vertebral bone quality assessment of lumbar MRI.

OBJECTIVE: This study aimed to develop and validate a fully automated vertebral bone quality (VBQ) pipeline for lumbar MRI studies and to assess its diagnostic performance with dual-energy x-ray absorptiometry (DXA)-defined osteoporosis. METHODS: Lumbar MRI studies from patients being evaluated for low-back pain were collected, deidentified, and assigned to reviewers for manual annotations. Each MRI study was manually annotated by two independent reviewers and by a custom, automated artificial intelligence (AI)-based segmentation algorithm. The measured signal intensity from the regions of interest (ROIs) was then extracted. Agreement between manual and automated methods was assessed using intraclass correlation coefficients (ICCs), root mean square error, and the paired t-test. The automated algorithm was validated on a separate cohort of 100 MRI examinations with paired DXA studies. Diagnostic performance for identifying osteoporosis was evaluated using receiver operating characteristic analysis to calculate the area under the curve (AUC). RESULTS: A total of 1089 patients were included to compare manual and automated VBQ calculations. Correlation between reviewer and automation was strong (r = 0.87, p < 0.001), and agreement was good with an ICC of 0.86 (95% CI 0.81-0.89). Diagnostic capacity was validated in a separate cohort (n = 100) with an AUC of 0.79 (95% CI 0.63-0.92). The Youden-optimized VBQ score of 2.59 increased sensitivity from 0.20-0.50 to 0.66-0.80 while maintaining high specificity at 0.83-0.96. CONCLUSIONS: Automated VBQ measurement closely aligns with manual methods. An optimized threshold of 2.59 improved sensitivity for DXA-defined osteoporosis while maintaining high specificity. The automated algorithm may serve as an adjunct screening tool for osteoporosis in clinical practice and as a scalable research application suitable for integration into clinical workflows and large-scale MRI dataset analysis.