Table 6.

Ninety-five percentile Hausdorff distance (mm) and standard deviation (in parentheses) of the proposed method and the fully-supervised method (FS-model) trained with ground truth labels.

Test set	Method	Mandible	Parotid-R	Parotid-L	Submand-R	Submand-L
TCIA	DIR-majority-vote	2.8 (±0.8)	9.2 (±3.9)	9.3 (±2.8)	8.2 (±1.8)	9.2 (±4.0)
	FS-model	2.6 (±0.7)	10.3 (±5.5)	10.5 (±5.3)	7.7 (±3.5)	5.8 (±1.7)
	Ours(N10M50)	2.7 (±1.3)	8.2 (±3.4)	8.2 (±2.4)	7.1 (±1.7)	7.4 (±3.6)
UTSW	DIR-majority-vote	4.2 (±2.3)	12.1 (±4.2)	21.8 (±25.8)	16.33 (±13.5)	10.3 (±2.4)
	FS-model	3.5 (±1.5)	11.9 (±5.0)	8.2 (±1.7)	16.09 (±15.6)	8.4 (±3.4)
	Ours(N10M50)	3.6 (±2.2)	10.5 (±4.3)	9.4 (±3.0)	14.65 (±12.9)	7.5 (±2.7)
PDDCA	DIR-majority-vote	2.6 (±1.0)	7.8 (±2.9)	8.3 (±3.2)	7.2 (±2.9)	7.3 (±2.0)
	FS-model	2.5 (±0.7)	7.6 (±3.0)	8.1 (±2.1)	6.5 (±1.7)	6.6 (±2.1)
	Ours(N10M50)	2.2 (±0.7)	7.3 (±2.7)	7.8 (±2.6)	6.8 (±2.5)	7.1 (±2.5)