Fig. 3.

Clavicle segmentation results of real-data-trained segmentation models. Three subjects with adversarial-positioned ($\alpha =-2^{\circ }$ and ${20}^{\circ }$) and well-positioned ($\alpha =0^{\circ }$) images are shown.The ground truth (GT) segmentation contours from TotalSegmentator are red, TorchXRayVision are green, and U-Net(Real) are blue in color. The Dice score analysis with respect to GT for each patient and angle are shown below each image, left is TorchXRayVision and right is U-Net(Real).