Figure 1

Predictive performance comparison of the trained models and the radiologists on the ROC. The radiologists’ performance with PI-RADS \(\ge\) 3 and \(\ge\) 4 are marked to locate clinically informative performance points. The sensitivity in the low and high specificity area is increased using either registration or misalignment augmentation compared to the unregistered dataset. Their performance closely match and slightly exceed the PI-RADS \(\ge\) 3 point, respectively, but none of them improved the sensitivity towards PI-RADS \(\ge\) 4. Combining registration and misalignment augmentation closely matches both PI-RADS performance points with the highest sensitivity improvement over the widest specificity range.