Table 2.

Quantitative evaluation results of different inputs and loss functions for segmentor S in whole glioma segmentation. $x_{t^{\prime }}$ denotes our synthesized FLAIR image, $x_s$ is the input multi-modal image with absence of FLAIR, and $y_c$ is the coarse segmentation. The last two sections are based on a concatenation of these images as input. ${}^{†}$ denotes significant improvement from $x_s$ as input based on a paired t-test (p-value $<$0.05).

Input	Training loss			Segmentation Performance
	${\mathcal{L}}_{\mathit{fine}}$	${\mathcal{L}}_{\mathit{err}}$	${\mathcal{L}}_{\mathit{con}}$	Dice (%)	ASSD (mm)	HD95 (mm)
$x_s$	✓			84.54$\pm$9.62	2.32$\pm$2.12	10.75$\pm$15.86
$x_{t^{\prime }}$	✓			86.33$\pm$6.85	1.70$\pm$0.91	6.40$\pm$5.13
$x_t^{\prime },y_c$	✓			86.76$\pm$8.11	1.72$\pm$1.15	6.68$\pm$6.69
		✓		86.62$\pm$8.39	1.70$\pm$1.11	6.57$\pm$6.93
	✓	✓		86.81$\pm$7.86	1.70$\pm$1.12	6.81$\pm$7.86
	✓	✓	✓	87.15$\pm$7.41	1.60$\pm$0.84	5.83$\pm$4.75
$x_t^{\prime },x_s,y_c$	✓			86.95$\pm$8.02	1.67$\pm$1.07	6.36$\pm$6.07
		✓		86.81$\pm$8.75	1.69$\pm$1.27	6.55$\pm$6.88
	✓	✓		87.17$\pm$8.56	1.63$\pm$1.16	6.34$\pm$6.78
	✓	✓	✓	87.55$\pm$7.62${}^{†}$	1.53$\pm$0.85${}^{†}$	5.67$\pm$4.92${}^{†}$