Clinical Validation of Commercial AI Software for the Detection of Incidental Vertebral Compression Fractures in CT Scans of the Chest and Abdomen

Abstract

Background/Objectives: The objective of this study was to clinically validate the performance of the Nanox.AI HealthOST software in detecting incidental vertebral compression fractures (VCFs) on outpatient chest and abdomen CT scans using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). A secondary aim was to assess the rate of missed VCFs using initial radiologist reports. Methods: A retrospective analysis was performed on 590 outpatient CT scans. HealthOST, an artificial intelligence solution from Nanox.AI that allows for automated spine analysis using CT images was evaluated against a consensus ground truth established by two radiologists, including a senior musculoskeletal radiologist. Two vertebral body height reduction thresholds were tested: mild (>20%) and moderate (>25%). Original radiologist reports were reviewed to identify missed VCFs. Results: At the 20% threshold, the AI achieved a sensitivity of 92.0%, a specificity of 52.7%, a PPV of 16.5%, and an NPV of 98.5%. At the 25% threshold, sensitivity decreased to 78.0%, while specificity improved to 94.2%, with a PPV of 51.1% and an NPV of 98.2%. The AI identified 88% and 92% of fractures missed by radiologists at the 20% and 25% thresholds, respectively. Conclusions: The Nanox HealthOST AI solution demonstrates potential as an effective screening tool, with threshold selection adaptable to clinical needs with a secondary review by a radiologist that is advisable to ensure diagnostic accuracy. The study further indicates that radiologists often overlook VCFs in reporting non-indicated cases and that AI has a role in enhancing the detection and reporting of vertebral compression fractures in routine clinical practice.

Keywords: artificial intelligence; osteoporosis; spine imaging; vertebral compression fractures.

Figure 1

AI-based L2 vertebral compression fracture calculation and attenuation value of L4 low bone density: representative example.

Figure 2

Sagittal CT images (A–H) in 8 different patients with AI software calling false positive fracture. (A) the white arrows show osteoarthritic wedging deformity involving vertebral bodies T8–10; (B) the white arrows show physiological wedging deformity involving vertebral bodies T12–L1; (C) the white arrows show endplate irregularities denoting Scheuermann’s disease, noted in multiple lower thoracic vertebral bodies, namely T8–11; (D) the white arrow shows cupids bow deformity noted in lumbar vertebral bodies L4–L5; (E) the white arrows show concavity/ballooned disk spaces noted in lumbar vertebral bodies L1–L4; (F) the white arrows show Schmorl’s node involving vertebral bodies T11–12. Fractures of T7 and T9 vertebral bodies were accurately identified; (G) the white arrows show edge of field of view overcalls involving the T1 vertebral body that appears normal; (H) the white arrows show T6 fracture overcall in a patient with scoliosis.