Fig. 3: Violin plots showing centroid error distributions.

Centroid errors of the conditional Generative Adversarial Network segmentation compared to the ground truth, shown for each patient in the anterior–posterior/lateral (AP/LAT, red) and superior–inferior (SI, yellow) directions for the masked dataset (a) and the markerless dataset (b). Each violin plot represents data from n = 1000 images, with the bars indicating the median (thick) and quartiles (thin).