Table 1.

Differences between the SkullFix and the SkullBreak dataset tracks.

	SkullFix	SkullBreak
Training/test split	$100/100+10$	$570/100$
Volume size	$512\times 512\times Z$	$512\times 512\times 512$
Voxel size	various	0.4 mm
Preprocessing	acquisition geometry regularization transform	acquisition geometry regularization transform and rigid alignment using the landmarks defining Frankfort-horizontal plane
Skull segmentation	thresholding at 150 HU, noise removal using connected components analysis	convolutional neural network and graph-cut [3]
Defect injection	binary defect shape subtraction from complete skull	binary defect shape subtraction from complete skull and defect border smoothing using morphological operations