AI-supported opportunistic detection of vertebral fractures on routine CT scans: Diagnostic performance and clinical relevance

Background: Vertebral fractures (VFs) are among the most common osteoporotic fractures, yet they are frequently underdiagnosed and left untreated. The use of an artificial intelligence (AI) tool may support improved detection rates. This study aimed to evaluate the diagnostic performance of the AI-based software (IB Lab FLAMINGO) in identifying VFs on thoracic and abdominal CT scans, using radiologist assessment as the reference standard. Methods: This was a monocentric, retrospective cross-sectional study. 205 patients with CT scans performed for non-skeletal indications were randomly selected. Sensitivity, specificity, and accuracy were calculated at the vertebra and patient level. We examined the proportion of false positive AI-identified fractures that might represent overlooked fractures upon re-evaluation. Results: Among 205 patients (59 % male; mean age 67.9 +/- 9.5 years), VFs were initially identified in 11.2 % by the radiologist, most frequently at the thoracolumbar junction. Females showed more thoracic (T4-T7) fractures, while males more commonly had lumbar (L1-L4) fractures. The AI analyzed 190 patients (92.7 %), detecting fractures in 24.7 %. At the patient level, IB Lab FLAMINGO showed 81 % accuracy, 74 % sensitivity and 82 % specificity. Vertebra-level performance (N = 3040) demonstrated high accuracy (97 %) and specificity (97 %). As a result of re-evaluation, fractures were confirmed in 29 of 30 AI-positively flagged patients, increasing sensitivity to 88.5 %, specificity to 99.3 %, and the overall presence of VFs to 25 %. Discussion: The performance metrics support potential use of AI IB Lab FLAMINGO as a screening aid and as a quality assurance tool, taking into account the proportion of missed diagnoses by the radiologist.