Fig. 2.

Dice score (larger values are better) and Robust Hausdorff Distances (HD95, smaller values are better) on validation cases for comparison of baseline, individual method contributions (not cumulative) and CerebNet. CerebNet combines multiple data augmentations with pre-training on a representative cross-study dataset. Deformation-250/500 indicates the number of realistic deformation fields used for static augmentation. The baseline model is our network without augmentation or pre-training. Error bars indicate 95% confidence intervals.