Fig. 6.

Sample segmentations from the first experiment. Major segmentation mistakes are indicated with yellow arrows. SynthSeg produces very accurate segmentations for all contrasts and resolutions. The T1 baseline makes small errors outside its training domain and cannot be applied to other modalities. While the TTA approach yields very good segmentations for T1mix, its results degrade for larger domain gaps, where it is outperformed by SIFA. Finally, SAMSEG yields coherent results for scans at 1 mm resolution, but is heavily affected by PV effects at low resolution.