Accuracy of an artificial intelligence algorithm for detecting moderate-to-severe vertebral compression fractures on abdominal and thoracic computed tomography scans

Abstract

Abstract Objective: To describe the accuracy of HealthVCF, a software product that uses artificial intelligence, in the detection of incidental moderate-to-severe vertebral compression fractures (VCFs) on chest and abdominal computed tomography scans. Materials and Methods: We included a consecutive sample of 899 chest and abdominal computed tomography scans of patients 51–99 years of age. Scans were retrospectively evaluated by the software and by two specialists in musculoskeletal imaging for the presence of VCFs with vertebral body height loss > 25%. We compared the software analysis with that of a general radiologist, using the evaluation of the two specialists as the reference. Results: The software showed a diagnostic accuracy of 89.6% (95% CI: 87.4–91.5%) for moderate-to-severe VCFs, with a sensitivity of 73.8%, a specificity of 92.7%, and a negative predictive value of 94.8%. Among the 145 positive scans detected by the software, the general radiologist failed to report the fractures in 62 (42.8%), and the algorithm detected additional fractures in 38 of those scans. Conclusion: The software has good accuracy for the detection of moderate-to-severe VCFs, with high specificity, and can increase the opportunistic detection rate of VCFs by radiologists who do not specialize in musculoskeletal imaging.