Table 4.
The application of AI in treatment planning.
| Author (Year) | Data Type | Dataset Size (Training/Test) |
Algorithms | Purpose | Performance |
|---|---|---|---|---|---|
| Xie et al. (2010) [121] | Cephalometric variables, cast measurement. | 180/20 | ANN | To predict tooth extraction diagnosis. | Accuracy: 80%. |
| Jung et al. (2016) [122] | Cephalometric variables, dental variable, profile variables, and chief complaint for protrusion. | 64/60 (additional 32 samples than validation set) |
ANN | To predict tooth extraction diagnosis, and extraction patterns. | Success rate: Tooth extraction diagnosis: 93%. Extraction patterns: 84%. |
| Li et al. (2019) [123] | Demographic data, cephalometric data, dental data, and soft tissue data. | A total of 302 samples | MLP (ANN) | To predict tooth extraction diagnosis, extraction patterns and anchorage patterns. | Accuracy: For extraction diagnosis: 94%. For extraction patterns: 84.2%. For anchorage patterns: 92.8%. |
| Suhail et al. (2020) [124] | Diagnosis, feature identification of photos, models and X-rays. | A total of 287 samples | ANN, LR, RF | To predict tooth extraction diagnosis, and extraction patterns. | For extraction diagnosis: LR outperformed the ANN. For extraction patterns: RF outperformed ANN. |
| Etemad et al. (2021) [125] | Demographic data, cephalometric data, dental data, and soft tissue data. | A total of 838 samples | RF, MLP (ANN) | To predict tooth extraction diagnosis. | Accuracy of RF with 22/117/all inputs: 0.75/0.76/0.75. Accuracy of MLP with 22/117/all inputs: 0.79/0.75/0.79. |
| Shojaei et al. (2022) [126] | Medical records, extra and intra oral photos, dental model records, and radiographic images. | A total of 126 samples | LR, SVM, DT, RF, Gaussian NB, KNN Classifier, ANN | To predict tooth extraction diagnosis, extraction patterns and anchorage patterns. | Accuracy for extraction decision: ANN: 93%, LR:86%, SVM:83%, DT: 76%, RF: 83%, Gaussian NB: 72%, KNN Classifier: 72%. Accuracy for extraction pattern: ANN: 89%, RF:40%. Accuracy for extraction and anchorage pattern: ANN: 81%, RF:23%. |
| Real et al. (2022) [127] | Sex, model variables, cephalometric variables, outcome variables. | -/214 | Commercially available software (Auto-WEKA) | To predict tooth extraction diagnosis. | Accuracy: 93.9%: input model and cephalometric data. 87.4%: input only model data. 72.7%: input only cephalometric data. |
| Leavitt et al. (2023) [128] |
Cephalometric variables, dental variables, demographic characteristics. | 256/110 | RF, LR, SVM | To predict tooth extraction patterns. | Overall accuracy: RF: 54.55%, SVM: 52.73%, LR: 49.09%. |
| Ryu et al. (2023) [78] | Intraoral photographs, extraction decision. | 2736/400 | ResNet (ResNet50, ResNet101), VggNet (VGG16, and VGG19) | To predict tooth extraction diagnosis. | Accuracy: Maxilla: VGG19 (0.922) > ResNet101 (0.915) > VGG16 (0.910) > ResNet50 (0.909). Mandible: VGG19 (0.898) = VGG16 (0.898) > ResNet50 (0.895) > ResNet101 (0.890). |
| Prasad et al. (2022) [129] | Clinical data, cephalometric data, cast and photographic data. | A total of 700 samples | RF, XGB, LR, DT, K-Neighbors, Linear SVM, NB | To predict skeletal jaw base, extraction diagnosis for Class 1 jaw base, and functional/camouflage/surgical strategies for Class 2/3 jaw base. | Different algorithms showed different accuracies in different layers. RF performed best in 3 out of 4 layers. |
| Knoops et al. (2019) [130] | 3D face scans | A total of 4261 | SVM for classification LR, RR, LARS, and LASSO for regression |
To predict surgery/non-surgery decision and surgical outcomes. | For surgery/non-surgery decision: Accuracy: 95.4%. Sensitivity: 95.5%. Specificity: 95.2%. For surgical outcomes simulation: Average error: LARS and RR (1.1 ± 0.3 mm). LASSO (1.3 ± 0.3 mm). LR (3.0 ± 1.2 mm). |
| Choi et al. (2019) [131] | Lateral cephalometric variables, dental variable, profile variables, chief complaint for protrusion. | 136/112 (additional 68 samples than validation set) |
ANN | To predict surgery/non-surgery decision, extraction/non-extraction for surgical treatment. | Accuracy for all dataset: Diagnosis of surgery/non-surgery: 96%. Diagnosis of extraction/non-extraction for Class II surgery: 97%. Diagnosis of extraction/non-extraction for Class III surgery: 88%. Diagnosis of extraction/non-extraction for surgery: 91%. |
| Lee et al. (2020) [132] | Lateral cephalograms. | 220/40 (additional 73 samples than validation set) | CNN (Modified-Alexnet, MobileNet, and Resnet50) | To predict the need for orthognathic surgery. | Average accuracy for all dataset: Modified-Alexnet: 96.4%. MobileNet: 95.4%. Resnet50: 95.6%. |
| Jeong et al. (2020) [133] | Facial photos (front and right). | A total of 822 samples. Group 1: 207/204. Group 2: 205/206 |
CNN | To predict the need for orthognathic surgery. | Accuracy: 0.893. Precision: 0.912. Recall: 0.867. F1 scores:0.889. |
| Shin et al. (2021) [134] | Lateral cephalograms and posteroanterior cephalograms. | A total of 840 samples. Group 1: 273/304 (additional 30 samples than validation set). Group 2: 98/109 (additional 11 samples than validation set) |
CNN | To predict the diagnosis of orthognathic surgery. | Accuracy: 0.954. Sensitivity: 0.844. Specificity: 0.993. |
| Kim et al. (2021) [135] | Lateral cephalograms. | 810/150 | CNN (ResNet-18, 34, 50, 101) |
To predict the diagnosis of orthognathic surgery. | Accuracy for test dataset: ResNet-18/34/50/101: 93.80%/93.60%/91.13%/91.33%. |
| Lee et al. (2022) [136] | Cephalometric measurements, demographic characteristics, dental analysis, and chief complaint. | 136/60 | RF, LR | To predict the diagnosis of orthognathic surgery. | Accuracy (RF/LR): 90%/78%. Sensitivity (RF/LR): 84%/89%. Specificity (RF/LR): 93%/73%. |
| Woo et al. (2023) [137] | Intraoral scan data | -/30 | Three commercially available software packages (Autolign, Outcome Simulator Pro, Ortho Simulation) | To evaluate the accuracy of automated digital setup accuracy. | Mean error of three pieces of software: Linear movement: 0.39~1.40 mm. Angular movement: 3.25~7.80°. |
| Park et al. (2021) [138] | Lateral cephalograms | A total of 284 cases | CNN (U-Net) | To predict the cephalometric changes of Class II patients after using modified C-palatal plates. | Total mean error: 1.79 ± 1.77 mm. |
| Tanikawa et al. (2021) [139] | 3D facial images | A total of 72 cases in surgery group and 65 cases in extraction group | Deep learning | To predict facial morphology change after orthodontic or orthognathic surgical treatment. | Average system errors: Surgery group: 0.94 ± 0.43 mm; orthodontic group: 0.69 ± 0.28 mm. Success rates (<1 mm): Surgery group: 54%; orthodontic group: 98%. Success rates (<2 mm): Surgery group: 100%; orthodontic group: 100%. |
| Park et al. (2022) [140] | CBCT | 268/44 | cGAN | To predict post-orthodontic facial changes. | Mean prediction error: 1.2 ± 1.01 mm. Accuracy within 2 mm: 80.8%. |
| Xu et al. (2022) [141] | Total of 17 clinical features | A total of 196 cases | ANN | To predict patient experience of Invisalign treatment. | Predictive success rate: Pain: 87.7%. Anxiety: 93.4%. Quality of life: 92.4%. |
ANN, artificial neural network; DT, decision tree; RF, random forest; LR, logistic regression; SVM, support vector machine; NB, naive bayes; KNN, k-nearest neighbors; MLP, multilayer perceptron; XGB, eXtreme Gradient Boosting; RR, ridge regression; LARS, least-angle regression; LASSO, least absolute shrinkage and selection operator regression; CNN, convolutional neural network; CBCT, cone-beam computed tomography; cGAN, conditional generative adversarial networks.