TABLE III.

Example applications of DL into medical image segmentation for radiation therapy planning purposes

Author	Disease Site	Image Type, Number of cases	Algorithm	Outcome
Zhu et al.[1]	Head and Neck	Planning CT, 261	3D U-Net, whole image	-9 OAR contours generated in ~0.12 seconds -Increased DSC ~3%
Lei et al.[2]	Male Pelvis	Cone-beam with synthetic MRI, 100	CycleGAN	-DSC >0.9 and MDA <1.0 mm for bladder, prostate, rectum from ground truth delineations
Dong et al.[3]	Thorax	CT, 35	U-Net GAN	-DSC>0.87 and MDA <1.5 mm for lungs, cord, heart -DSC esophagus ~0.75 -Negligible dosimetric differences
van der Heyden et al.[4]	Brain	Dual-energy CT, 14	2-step 3D U-Net	-Quantitatively and qualitatively outperformed atlas method for all organs at risk but optic nerves
Chen et al.[6]	Abdomen	3T MRI, 102	2D U-Net	-9/10 OARs had DSC 0.87-0.96 -Duodenum DSC ~0.8
Fu et al.[7]	Abdomen	0.35 T, MR-linac, 120	CNN with correction method	-4/5 OARs had DSC >0.85 (liver, kidneys, stomach, bowel) -Duodenum DSC ~0.65