Table 5.

Summary of related studies for AI application in orthodontics.

AI Application	Author, Year (Ref)	Architecture	Data Modality	Dataset Size Split (Train/Val/Test or Train/Test)	Study Factor	Reference Standard (Ground Truth)	Validation Scheme	Results (Performance Metrics/Values)	Conclusion
Disease diagnosis	Chen et al., 2020 [33]	ML-based algorithm based on multisource integration framework	CBCT images	36 CBCT images, Train: 30 images, Test: 6 images 30/6	Assess maxillary structure variation	NA	NA	Dice score of maxilla: 0.80	The method is helpful in assessing maxillary structure variation in unilateral canine impaction
	Nino-Sandoval et al., 2016 [34]	Support Vector Machine (SVM)	cephalograms	229 cephalograms	Sagittal (skeletal) patterns	NA	Ten-fold	Accuracy: 0.741	The non-parametric method has the potential to classify skeletal patterns using craniomaxillary variables
	Yu et al., 2014 [35]	Support Vector Regression (SVR)	Colored Photographs	108 images	Facial attractiveness (most attractive to least attractive)	69 Orthodontists	NA	Accuracy: 0.718	The model was helpful in finding close correlation with facial attractiveness from orthodontic photographs
Treatment planning	Riri et al., 2020 [37]	Tree Based Classification	Extraoral intraoral and mould images	1207 total images, Extraoral: 325 images, Intraoral: 812 images, Mould: 70 images	Facial and skin color features	NA	NA	Accuracy: 0.942, Sensitivity: 0.953, Specificity: 0.996, F1-score: 0.926	The automatic approach was helpful in classification of orthodontic images with encouraging classification performance
	Suhail et al., 2020 [38]	Random Forest Ensemble Learning Method	Patient records (medical charts, x-rays, facial photographs)	287 patient records	Decision making for teeth extraction	Five experts	Five-Fold	Accuracy: 0.944	The RF ensemble classifier was helpful in extraction and treatment planning
Landmark detection in cephalograms	Song et al., 2020 [60]	Pretrained ResNet-50 using transfer learning	X-Ray images	400 cephalograms, Train: 150 images, Testing sets: Test set 1—150 images, Test set 2—100 images	Detect cephalometric landmarks	Two experienced doctors	NA	Successful Detection Rate (SDR): Test 1—0.862, Test 2—0.758	The model was able to achieve satisfying SDR in detecting 19 landmarks
	Kim et al., 2020 [65]	Two stage DNN with stacked hourglass network	Dataset 1: 2075 cephalograms, Dataset 2: 400	cephalograms	Detect cephalometric landmarks	Two experts	NA	SDR: Test set 1—0.883, Test set 2—0.77	The fully automated cephalometric analysis algorithm and web application help in the diagnosis of anatomic landmarks
	Gilmour and Ray, 2020 [66]	Pretrained ResNet-50 with foveated pyramid attention algorithm	cephalograms	400 cephalograms	Detect cephalometric landmarks	NA	Four-fold	SDR: Test set 1—0.883, Test set 2—0.77	The multiresolution approach was useful in learning features across all scales and is promising for large images
	Zhong et al., 2019 [67]	Attention guided deep regression model through 2 stage U-Net	cephalograms	300 cephalograms Train: 150, images Test: 150 images 150/150	Detect cephalometric landmarks	Two experienced doctors	Four-fold	SDR: 86.74%	The attention-guided mechanism ensures that smaller searching scopes and high data resolution with minimum information redundancy and the model is generalizable to other landmarks
	Park et al., 2019 [61]	YOLOv3	cephalograms	1311 cephalograms, Train: 1028 images, Test: 283 images 1028/283	Detect cephalometric landmarks	One examiner	NA	SDR: 0.804 Computational time: 0.05 s	The model was effective in identifying 80 landmarks with 5% higher accuracy compared to top benchmarks
Disease diagnosis	Makaremi et al., 2019 [68]	Customized CNN	cephalograms	1870 cephalograms	Cervical vertebra maturation (CVM) degree	Experts	Three-fold	Accuracy: 0.95	The model was helpful in determining the degree of maturation of CVM and has the potential to be implemented in real world scenario
	Yu et al., 2020 [69]	Modified DenseNet pretrained with ImageNet weights	Lateral cephalograms	5890 cephalograms	skeletal classification	Five orthodontic specialists	NA	Sagittal Accuracy: 0.957, Vertical Accuracy: 0.964	The model shows potential in skeletal orthodontic diagnosis using lateral cephalograms
Treatment plannimng	Lee et al., 2020 [62]	Modified AlexNet	cephalograms	333 cephalograms, Train: 220 images, Valid: 73 images, Test: 40 images	Differential orthodontic diagnosis	NA	Four-fold	Accuracy: 0.919, Sensitivity: 0.852, Specificity: 0.973, AUC: 0.969	The study indicates that the DCNNs-based model can be applied for differential diagnosis in orthodontic surgery
Disease diagnosis	Amasya et al., 2020 [21]	ANN	cephalograms	647 cephalograms, Train: 498 images, Test: 149 images 498/149	Cervical vertebra maturation degree	Two independent observers	NA	Accuracy: 0.869, Sagittal Sensitivity: 0.935, Vertical Specificity: 0.945	The model was helpful in CVM staging and cervical vertebral morphology classification
	Kok et al., 2019 [70]	ANN	cephalograms	300 cephalograms	Cervical vertebrae stages	Orthodontists	Five-fold	AUC: CV1—0.99, CV2—0.96, CV3—0.94, CV4—0.90, CV5—0.91, CV6—0.96	Compared with machine learning algorithms, ANN provides the most stable results with 2.17 average rank on hand-wrist radiographs
	Budiman et al., 2013 [22]	ANN	Orthodontic scans	190 scanned dental casts	Shape of arch form	Three orthodonotics	NA	Accuracy: 0.763	ANN has the potential to identify arch forms with encouraging accuracy
Treatment planning and prognosis	Choi et al., 2019 [24]	ANN	cephalogram	316 cephalograms, Train: 136 images, Valid: 68 images, Test: 112 images	Surgery type and extraction decision	One otthodontist	NA	Accuracy Surgery decision: 0.96, Surgery type and extraction decision: 0.91	The model was helpful in diagnosing and making surgery type and extraction decision effectively and can be used as auxiliary reference when clinicians make a decision