Table 2:

Evaluation criteria for comparing quality of generated segmentation masks for organs-at-risk compared to manual segmentations for real week1 CBCT input images. Setting anatomy labels for esophagus, spinal cord, heart and lungs as 4, 3, 2, 1 respectively (instead of 1, 2, 3, 4 respectively) helps to substantially improve the segmentation results for esophagus and spinal cord without any noticeable degradation in heart and lungs results. The stabilized version has GAN stabilization techniques added in the optimization phase and produces better results compared to the unstabilized version.

Settings	Anatomy	DICE	MSD (mm)	HD95 (mm)
		mean ± std	mean ± std	mean ± std
msk+aapm+stabilized+eso4	Lungs	0.96 ± 0.01	0.96 ± 0.20	3.44 ± 0.93
	Heart	0.88 ± 0.08	2.28 ± 0.94	8.07 ± 4.80
	Spinal Cord	0.83 ± 0.03	1.12 ± 0.32	3.45 ± 3.39
	Esophagus	0.66 ± 0.06	2.22 ± 0.40	6.50 ± 1.91
msk+aapm+unstabilized+eso4	Lungs	0.95 ± 0.01	1.38 ± 0.55	6.34 ± 4.83
	Heart	0.81 ± 0.09	4.24 ± 1.32	19.08 ± 7.19
	Spinal Cord	0.80 ± 0.03	1.25 ± 0.39	3.81 ± 3.97
	Esophagus	0.58 ± 0.11	4.01 ± 2.96	17.83 ± 20.73
msk+stabilized+eso4	Lungs	0.95 ± 0.01	1.01 ± 0.21	3.71 ± 1.08
	Heart	0.88 ± 0.10	2.13 ± 0.82	7.38 ± 3.27
	Spinal Cord	0.82 ± 0.03	1.16 ± 0.36	3.61 ± 3.68
	Esophagus	0.63 ± 0.06	2.51 ± 0.45	8.27 ± 4.08
msk+aapm+stabilized+eso1	Lungs	0.96 ± 0.01	1.03 ± 0.24	3.73 ± 1.08
	Heart	0.87 ± 0.08	2.31 ± 0.78	7.43 ± 2.21
	Spinal Cord	0.78 ± 0.04	1.32 ± 0.36	3.81 ± 2.87
	Esophagus	0.56 ± 0.12	3.61 ± 2.26	15.22 ± 15.76