Table 2.
The application of AI in skeletal maturation assessment using lateral cephalograms.
| Author (Year) | Data Type | Dataset Size (Training/Test) |
Algorithm | Performance |
|---|---|---|---|---|
| Kök et al. (2019) [99] | Lateral cephalograms | 240/60 | k-NN, NB, DT, ANN, SVM, RF, LR | Mean rank of accuracy: k-NN: 4.67, NB: 4.50, DT: 3.67, ANN: 2.17, SVM: 2.50, RF: 4.33, LR: 5.83. |
| Makaremi et al. (2019) [100] | Lateral cephalograms | Training: 360/600/900/1870 Evaluation: 300 Testing: 300 |
CNN | Performance varied depending on image numbers and pre-processing method. |
| Amasya et al. (2020) [101] | Lateral cephalograms | 498/149 | LR, SVM, RF, ANN, DT | Accuracy: LR: 78.69%, SVM: 81.08%, RF: 82.38%, ANN: 86.93%, DT: 85.89%. |
| Amasya et al. (2020) [102] | Lateral cephalograms | -/72 | ANN | Average of 58.3% agreement with four human observers. |
| Kök et al. (2021) [91] | Lateral cephalograms | A total of 419 | Total of 24 different ANN models | The highest accuracy was 0.9427. |
| Seo et al. (2021) [103] |
Lateral cephalograms | A total of 600 | ResNet-18, MobileNet-v2, ResNet-50, ResNet-101, Inception-v3, Inception-ResNet-v2 | Accuracy/Precision/Recall/F1 score: ResNet-18: 0.927 ± 0.025/0.808 ± 0.094/0.808 ± 0.065/0.807 ± 0.074. MobileNet-v2: 0.912 ± 0.022/0.775 ± 0.111/0.773 ± 0.040/0.772 ± 0.070. ResNet-50: 0.927 ± 0.025/0.807 ± 0.096/0.808 ± 0.068/0.806 ± 0.075. ResNet-101: 0.934 ± 0.020/0.823 ± 0.113/0.837 ± 0.096/0.822 ± 0.054. Inception-v3: 0.933 ± 0.027/0.822 ± 0.119/0.833 ± 0.100/0.821 ± 0.082. Inception-ResNet-v2: 0.941 ± 0.018/0.840 ± 0.064/0.843 ± 0.061/0.840 ± 0.051. |
| Zhou et al. (2021) [104] |
Lateral cephalograms | 980/100 | CNN | Mean labeling error: 0.36 ± 0.09 mm. Accuracy: 71%. |
| Kim et al. (2021) [105] |
Lateral cephalograms | 480/120 | CNN | Three-step model obtained the highest accuracy at 62.5%. |
| Rahimi et al. (2022) [106] |
Lateral cephalograms | 692/99 (additional 99 images than validation set). |
ResNet-18, ResNet-50, ResNet-101, ResNet-152, VGG19, DenseNet, ResNeXt-50, ResNeXt-101, MobileNetV2, InceptionV3. | ResNeXt-101 showed the best test accuracy: Six-class: 61.62%, three-class: 82.83%. |
| Radwan et al. (2023) [107] | lateral cephalograms | 1201/150 (additional 150 images than validation set). |
U-Net, Alex-Net | Segmentation network: Global accuracy: 0.99. Average dice score: 0.93. Classification network: Accuracy: 0.802. Sensitivity (pre-pubertal/pubertal/post-pubertal): 0.78/0.45/0.98. Specificity (pre-pubertal/pubertal/post-pubertal): 0.94/0.94/0.75. F1 score (pre-pubertal/pubertal/post-pubertal): 0.76/0.57/0.90. |
| Akay et al. (2023) [98] |
lateral cephalograms | 352/141 (additional 94 images than validation set). |
CNN | Classification accuracy: 58.66%. Precision (stage 1/2/3/4/5/6): 0.82/0.47/0.64/0.52/0.55/0.52. Recall (stage 1/2/3/4/5/6): 0.70/0.74/0.58/0.54/0.37/0.60. F1 score (stage 1/2/3/4/5/6): 0.76/0.57/0.61/0.53/0.44/0.56. |
k-NN, k-nearest neighbors; NB, Naive Bayes; LR, logistic regression; CNN, convolutional neural network; SVM, support vector machine; RF, random forest; ANN, artificial neural network; DT, decision tree.