Fig. 2.

Robustness as Dice score for real data-trained segmentation models in boxplot on well- and adversarial- positioned data. TorchXRayVision is black and U-Net(Real) is red in color. Internal patient rotation $\alpha$ values are divided into five groups. Well-positioned X-ray images are grouped as center. While adversarial data indicated by Moderate Negative (-), Low Negative (-), Low Positive (+), and Moderate Positive (+). The ranges includes $\alpha$ values with a step size of $2^{\circ }$. Ie. center represent $\alpha$ values of $-2^{\circ }$, $0^{\circ }$, and $2^{\circ }$.