Table II.

Evaluation of the geometric accuracy of auto-segmenting the left and right parotid gland of the cross-modality learning approach (highlighted in bold). As a benchmark, we also include the geometric accuracy of the CT-trained network.

ROI	Method	$\stackrel{─}{DSC}$	$\stackrel{─}{HD}$ (mm)	$\stackrel{─}{MSD}$ (mm)
Right	Cross-modality learning	0.76 ± 0.06	18.32 ± 10.12	2.66 ± 1.26
Parotid	CT only	0.81 ± 0.07	13.01 ± 5.61	1.87 ± 0.84
	Interobserver variability	0.84 ± 0.04	10.76 ± 4.35	1.40 ± 0.45
Left	Cross-modality learning	0.77 ± 0.04	17.75 ± 7.49	2.36 ± 0.75
Parotid	CT only	0.82 ± 0.05	12.98 ± 5.15	1.74 ± 0.53
	Interobserver variability	0.83 ± 0.04	10.94 ± 3.75	1.59 ± 0.63