Table 1.

Parameters of training and test datasets and segmentation scores on all four imaging modalities using all available training samples.

Modality	Ntrain	Nval	Ntest	Volume shape	Average ${\mathcal{D}}_A$	Weighted ${\mathcal{D}}_V$	$\mathcal{C}\left({\mathcal{D}}_A,CV\right)$	ASSD [mm]
MPRAGE	465	51	6	[128, 128, 128]	(86.7 ± 4.1)%	(88.7 ± 0.8)%	−0.91	(0.67 ± 0.23)
FLAIR	107	11	6	[128, 128, 128]	(81.9 ± 6.7)%	(83.7 ± 1.7)%	−0.87	(0.82 ± 0.26)
DWI	142	15	6	[160, 160, 32]	(80.8 ± 6.6)%	(82.0 ± 5.9)%	−0.87	(0.70 ± 0.27)
CT	34	5	5	[96, 128, 96]	(80.7 ± 8.2)%	(77.9 ± 5.2)%	−0.97	(1.12 ± 0.40)

Each dataset contains N_train training volumes, N_val validation volumes and N_test test samples. The voxel dimension of all volumes in the training and test dataset is fixed to the reported volume shape. Average and weighted Dice scores are reported according to Equations (4) and (5), respectively. The Pearson correlation coefficient between average Dice score and the uncertainty metric CV, obtained from dropout sampling, is listed for each imaging modality. The average symmetric surface distance (ASSD) in millimeters is reported as mean±s.d., where the mean and standard deviation (s.d.) are computed over all samples from the test dataset and averaged over all anatomical structures.