Fig. 2.

Algorithm discriminative power in pooled subgroups (“Algorithm 1,” “Algorithm 2”). (a): “Algorithm 2” outperformed “Algorithm 1” for all subgroups differing in the consideration of smaller PTXs that limit algorithm performance. (b): The presence of a contralateral PTX of minor or equal size improves the algorithm-based identification of suspicious images (especially compared to “Algorithm 1”), compared with corresponding subgroups in a. (c, d): PTX-positive SCXRs with inserted TTs (d) are significantly easier to be detected compared with similar images without inlying TTs (c); this effect is more pronounced for “Algorithm 1.” (a–d): Areas under receiver operating curves are illustrated including the 95% confidence intervals. Subgroup definitions partially based on the pooled inclusion/exclusion of PTX size subgroups (e.g., PTX > 1 cm means the pooled consideration of PTX 1–2 cm and PTX > 2 cm). Therefore, the numbers partially do not add up. PTX-positive cases that do not meet the subgroup PTX size definitions have been excluded from ROC analysis