Table 7.

The existing DL models on landmark localization and their function and performance.

Authors	DL Models	Year	Training Dataset	Validation/Test Dataset	Functions	Best Performance of DL	Time- Consuming
Bagci et al. [77]	Long short-term memory network	2019	20,480	5120	Mandible segmentation and 9 automatic landmarks	DSC: 0.9382 95HD: 5.47 IoU: 1 Sensitivity: 0.9342 Specificity: 0.9997	No
Shen et al. [78]	Multi-task dynamic transformer network	2020	no	no	64 CMF landmarks	DSC: 0.9395 ± 0.0130	No
Shen et al. [79]	U-Net, graph convolution network	2020	20	5 for validation 10 for test	60 CMF landmarks	Accuracy: 1.69 mm	1~3 min for DL
Yap et al. [80]	3D faster R-CNN, 3D MS-UNet	2021	60	60	18 CMF landmarks	Accuracy: 0.79 ± 0.62 mm	26.6 s for DL
Wang et al. [81]	3D Mask R-CNN	2022	25	25	105 CMF landmarks	Accuracy: 1.38 ± 0.95 mm	No
Yoon et al. [82]	Mask R-CNN	2022	170	30	23 CMF landmarks	mean absolute value of deviation length: 1 mm angle: <2°	25~35 min for manual 17 s for DL