We retrospectively and consecutively collected all chest-abdominal-pelvis (CAP) CT scans, with and without contrast, performed for various medical reasons other than suspicion of VCF (cancer, HIV, neoplasia, infection, etc.) at Sainte-Marguerite Hospital (Marseille, France). All the patients were more than 50 years old. The cases were acquired from January 2019 to August 2020. Non-inclusion criteria corresponded to material in the vertebrae and significant artifacts.

In order to establish the ground truth (GT), two board-certified radiologists analyzed the scans and defined by consensus the presence of VCF employing the Genant’s semiquantitative method [3]. A case was considered as positive if at least one vertebra had a grade 2 or 3 compression fracture. In addition, they labeled all the visible vertebrae from T1 to L5. The same data was processed by a DL-based prototype (CINA-VCF, Avicenna.AI, La Ciotat, France) intended to quantify vertebral height loss, label the vertebrae and passively notify positive cases. The algorithm results for VCF screening were compared to the GT and the sensitivity, specificity and accuracy were calculated at a per-case level. Similarly, the overall percentage agreement for vertebral labeling between the software and the GT was assessed at a per-vertebra level.