Fig. 1.

(a) Patient positioning possibilities in clinical X-Ray examination, A and B represent adversarial positioning and C represents correct positioning; (b) Simulation setup for generating rotated and non-rotated synthetic X-rays with an angle $\alpha$ in the range of [$-{20}^{\circ }$, ${20}^{\circ }$]; (c) Our proposed workflow: In the CT domain, we generated ground truth segmentation masks using TotalSegmentator. Paired CT image and segmentation volumes were forward projected to X-ray domain as normal and adversarial data for training and testing, and to quantify robustness in segmentation models TorchXRayVision, U-Net(Real), U-Net($\text{Real}+{\text{Syn}}_{\mathit{cen}}$) and U-Net($\text{Real}+{\text{Syn}}_{\mathit{all}}$).