The study included 249 CT scans from patients aged 71 ± 14 years, including 95 women and 154 men. Among these, 133 cases had at least one bone metastasis. 191 patients had known primary cancers, predominantly breast, lung, or prostate, while 58 had no identified primary cancer at the time of imaging. The dataset included 40 patients with prior cancer treatment, and five with metallic hardware. In positive patients, metastatic burden varied substantially: 61 presented with fewer than five metastases, 50 had between five and 15 metastases and 22 patients had more than 15 metastases. There were 81 patients with a predominance of osteolytic metastases and 52 patients with a majority of sclerotic metastases. The anatomical distribution of metastases showed highest frequency in the spine (n = 444), followed by rib (n = 162), and pelvis (n = 125) (Table 1). 

The analysis focused on lesions with a diameter superior or equal to 20mm. The AI software achieved an Area Under the Curve (AUC) of 0.95 [0.88, 0.99]. The lesion-wise sensitivity of the AI was 0.82 [0.77, 0.87] at 1 false positive per scan, while the radiologists displayed a mean sensitivity of 0.76 [0.74, 0.77], with a mean false positive rate of 1.1 [0.8, 1.3] per scan (Figure 2). The AI system showed equivalent performance in identifying osteolytic and sclerotic metastases with sensitivities of 0.83 [0.78, 0.88] and 0.80 [0.73, 0.88], respectively (p = 0.51). AI  performance also remained consistent for patients with and without any documented primary cancer showing lesion-wise sensitivities of 0.81 [0.75, 0.87] and 0.83 [0.72, 0.91] respectively (p = 0.73). Figure 3 illustrates a case where the AI system detected a sclerotic rib metastasis that was overlooked by all four radiologist readers.