Figure 3.

Generic inference scheme for all tasks: (1) a conditioning image is passed to the trained generative model, which guides its synthesis with information such as the orientation of the knee in voxel space; (2) the trained generative model synthesizes a fake, healthy version of the target image; (3) a voxel-wise subtraction of the real and fake target images generates an anomaly map; (4) deterministic postprocessing is applied to the anomaly map to produce the final model prediction: a binary segmentation map of the BMEL. Thumbnails exemplify the bone translation task outlined in Section 2.4.