Table 3.
Summary of related studies for AI application in periodontology.
| 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 segmentation |
Li et al., 2022 [29] | CNN | Oral endoscope images | 607 images Train: 320 images Test: 287 images 320/287 | Plaque segmentation | Dentist | NA | Acc: 0.864 IoU: 0.859 | The model is helpful in plaque segmentation on small dataset |
| Li et al., 2019 [30] | A method based on contrast limited adaptive histogram (CLAHE), gray-level co-occurrence matrix (GLCM), and extreme machine learning | Digital photographs | 93 images Train: 73 images Test: 20 images 73/20 | Gingivitis identification | NA | NA | Accuracy: 0.74, Sensitivity: 0.75, Specificity: 0.73, Precision: 0.74 | The method is helpful for gingivitis identification | |
| Disease localization | Lin et al., 2015 [6] | Level segmentation based on Bayesian or KNN or SVM classifier | Periapical Radiographs | 31 images | Alveolar bone loss | NA | Leave one-out | Mean SD True Positive Fraction (TPF): 0.925, True Positive Fraction (FPF): 0.14 | The model localizes bone loss areas with high classification effectiveness |
| Disease detection | Lee et al., 2022 [41] | VGG+Individual CNN | Periapical Radiographs | 1740 images Train: 1218 images, Valid: 417 images, Test: 105 images 1218/417/105 | Bone loss | Three dentists | NA | Acc: 0.99, AUC: 0.98 | The proposed algorithm is helpful in diagnosing periodontal bone loss |
| Krois et al., 2019 [5] | Seven layered deep CNN | Panoramic radiographs | 1750 images, Train: 1400 images, Valid: 350 images, 1400/350 | Bone loss | Three examiners | Ten-fold | Acc: 0.81 | The model shows discrimination ability similar to that of dentists | |
| Kim et al., 2019 [63] | Deep CNN + Transfer learning | Panoramic radiographs | 12,179 images, Train: 11,189 images, Valid: 190 images, Test: 800 images 1189/190/800 | Bone loss | Dental clinicians | NA | AUROC: 0.95, F1-score: 0.75, Sensitivity: 0.77, Specificity: 0.95, PPV: 0.73, NPV: 0.96 | The model is useful in tooth numbering and achieved detection performance superior to that of dental clinicians | |
| Lee et al., 2019 [43] | GoogleNet InceptionV3 + Transfer learning | Panoramic radiographs and CBCT images | 2126 images including 1140 panoramic and 986 CBCT images Train: 1700 images, Test: 426 images, 1700/426 | Odontogenic cyst lesion | NA | NA | Panoramic images: AUC—0.847, Sensitivity—0.882, Specificity—0.77, CBCT images AUC—0. 914, Sensitivity—0.961, Specificity—0.771 | The model provides higher diagnostic performance on CBCT images in effectively detecting and diagnosing cystic lesions | |
| Disease classification | Moran et al., 2020 [7] | ResNet Inception | Periapical radiographs | 467 images, Train: 415 images, Test: 52 images, 415/52 | Periodontal bone destruction | NA | NA | Acc: 0.81, Precision: 0.76, Recall: 0.92, Specificity: 0.71, NPV: 0.90 | The inception model classifies regions based on the presence of periodontal bone destruction with encouraging performance |
| Disease segmentation | Khan et al., 2021 [44] | UNet + DenseNet121 | Periapical radiographs | 200 images, Train: 160 images, Test: 40 images | Bone recession and inter-radicular radioulency | Three experts | NA | mIoU: 0.501, Dice score: 0.569 | Automates the process of detecting the presence and shape of caries |
| Zheng et al., 2021 [8] | Automatically constrained dense U-Net | CBCT images | 100 images | bone lesion identification | Three reviewers | Four-fold | Dice score for different categories: Background: 0.961, Lesion: 0.709, Material: 0.822, Bone: 0.877, Teeth: 0.801 | The model is helpful in detecting the correct shape of the lesion and the bone | |
| Duong et al., 2019 [45] | UNet | High frequency ultrasound images | 35 images, Train: 30 images, Test: 5 images, 30/5 | Alveolar bone assessment | Three experts | NA | Dice Coefficient: 0.75, Sensitivity: 0.77, Specificity: 0.99 | The method yields a higher performance in delineating alveolar bone as compared to experts | |
| Nguyen et al., 2020 [46] | U-Net with ResNet34 encoder | Intraoral ultrasound images | 1100 images, Train: 700 images, Valid: 200 images, Test: 200 images 700/200/200 | Alveolar bone assessment | Three examiners | NA | Dice Coefficient: 0.853, Sensitivity: 0.885, Specificity: 0.998 | The model has the potential to detect and segment alveolar bone automatically | |
| Disease diagnosis | Li et al., 2020 [64] | Mask RCNN + novel caliberation method | Panoramic radiographs | 298 images, Train: 270 images, Test: 28 images 270/28 | Periodontitis prediction | Junior dentist | NA | mAP: 0.826, Dice score: 0.868, F1-score: 0.454, Accuracy: 0.817 | The model is useful for diagnosing the severity degrees of periodontitis |
| Papantonopoulos et al., 2014 [23] | Multilayer Perceptron ANN | Textual | 29 subjects | Aggressive periodontitis | NA | Ten-fold | Accuracy: 0.981 | The model provides effective periodontitis classification | |
| Geetha et al., 2020 [19] | Back propagation Neural Network | Intraoral digital radiographs | 105 images | Dental caries detection | NA | Ten-fold | Accuracy: 0.971, FPR: 2.8%, ROC: 0.987 | The model is helpful for the detection of tooth decay and is independent of visual errors | |
| Risk assessment | Shankarapillai et al. 2012 [9] | Multilayer Feedforward Propagation | Textual | 230 subjects | Periodontitis risk assessment | NA | NA | MSE: 0.132 | The model can be used for effective periodontitis risk prediction |