Table 1.

Included studies

1st author	Year	Country	Imagery	Data source	Architecture/modelling framework	N landmarks	Total sample	Train/validate sample	Reference test on training/validation data	Unification of labels	Test sample	Reference test on test data	Unification of labels
Arik 2017 [20]	2017	USA	Lateral 2D	IEEE Grand Challenge 2015	Custom CNN combined with a shape model for refinement	19	400	150	2 experts	Average	150+100	2 experts	Average
Chen 2019 [21]	2019	China	Lateral 2D	IEEE Grand Challenge 2015	VGG-19, ResNet20, and Inception; custom attentive feature pyramid fusion module	19	400	150	2 experts	Average	150+100	2 experts	Average
Gilmour 2020 [22]	2020	Canada	Lateral 2D	IEEE Grand Challenge 2015	Modified ResNet34 combined with a custom image pyramids approach (spatialized features)	19	400	150	2 experts	Average	150+100	2 experts	Average
Huang 2020 [23]	2020	Germany	Lateral 2D	CQ500 CTs (train) and IEEE Grand Challenge 2015 (test)	LeNet-5 for ROI patches and ResNet50 for landmark location	19	na	491	3 radiologists	Majority	150	2 experts	Average
Hwang 2020 [24]	2020	Korea	Lateral 2D	Own dataset	Customized YOLO V3	80	1311	1028	1 expert	NA	283	1 expert	NA
Kim 2020 [25]	2020	Korea	Lateral 2D	Own dataset+IEEE Grand Challenge 2015	Stacked hourglass-shaped networks	23	2475	1875	2 experts	Unclear	200+225+400	2 experts	Unclear or average (IEEE)
Lee 2020 [26]	2020	Korea	Lateral 2D	IEEE Grand Challenge 2015	Custom CNN for ROI and custom Bayesian CNN for landmark detection	19	400	250	2 experts	Unclear	150	2 experts	Average
Lee 2019 [27]	2019	Japan	Lateral 2D	Own dataset	Combined custom CNNs for ROI classification and point estimation	22	936	835	3 experts	Unclear	100	3 experts	Unclear
Lee 2019 [28]	2019	Korea	3D	Own dataset	VGG-19	7	27	20	2 experts	Average	7	2 experts	average
Ma [29]	2020	Japan	3D	Own dataset	Custom CNNs for classification and regression	13	66	58	1 expert	NA	8	1 expert	NA
Muraev 2020 [30]	2020	Russia	Frontal 2D	Unclear	Multiclass FPN and ResNeXt-50 with Squeeze-and-Excitation blocks	45	330	300	Students, corrected by experts	Consensus	30	students, corrected by experts	Consensus
Noothout 2020 [31]	2019	Netherlands	Lateral 2D	IEEE Grand Challenge 2015	Custom FCNs based on ResNet34	19	400	150	2 experts	Unclear	150+100	2 experts	Average
O'Neil 2018 [32]	2018	UK	Lateral 3D	Own dataset	Custom FCN and Atlas Correction	22	22	201	3 experts	Unclear	20	2 experts	Unclear
Oh 2020 [33]	2020	Korea	Lateral 2D	IEEE Grand Challenge 2015	DACFL, custom FCN combined with a local feature perturbator and the anatomical context loss	19	400	150	2 dental experts	Average	150+100	2 experts	Average
Park 2020 [34]	2019	Korea	Lateral 2D	Own dataset	YOLO V3 and SSD	80	1311	1028	1 expert	NA	283	1 expert	NA
Qian 2020 [35]	2020	China	Lateral 2D	IEEE Grand Challenge 2015	Cepha-NN, combining U-Net-shaped networks, attention mechanism, and region enhancing loss	19	400	150	2 experts	Average	150+100	2 experts	Average
Song 2020[36]	2020	China	Lateral 2D	IEEE Grand Challenge 2015	ROI extraction and ResNet50	19	400	150	2 experts	Average	150+100	2 experts	Average
Yun 2020 [37]	2020	Korea	3D	Own dataset	Custom CNNs, combined skull normalization, and VAE for coarse to fine detection tasks	93	26	22	1 expert	NA	4	1 expert	NA
Zhong 2020 [38]	2019	China	Lateral 2D	IEEE Grand Challenge 2015	2-stage (global and local) U-Net models	19	400	150	2 experts	Average	150+100	2 experts	Average

Abbreviations: FCN, fully convolutional neural network; ROI, region of interest. Single Shot Detector.