Fig. 7.

A: SynthStrip variability across time-series data, measured by percent of discordant voxel locations (DV) across diffusion-encoded directions, relative to the brain mask volume. The ROBEX median % DV extends beyond the chart axis, as indicated by the black arrow. B: Effect of SDT- and Dice-based loss functions during training. A SynthStrip model trained using ${ℒ}_{sdt}$ predicts substantially smoother brain masks (boundaries indicated in orange) than a model trained with ${ℒ}_{dice}$, resulting in considerably lower maximum surface distance (MD) to ground truth masks and percent of exposed boundary voxels (EBV).