Table 4.
Summary of related studies for AI application in endodontics.
| 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 detection and classification |
Ghaedi et al., 2014 [31] | Circular Hough Transform Based Segmentation | Intraoral optical occlusal tooth surface images | 88 images with blue background | Detect and score caries lesions | International Caries Detection and Assessment System (ICDAS) experts | Ten fold | Accuracy: 0.863, Specificity: 0.983, Sensitivity: 0.83 | The automated system was helpful in detecting and score caries lesions. |
| Berdouses et al., 2015 [32] | Random Forest | Photographic colored images | 103 digital images | Occlusal caries lesions | Two pediatric dentists | Ten-fold | Accuracy: 0.80, Precision: 0.86, Recall: 0.86, F1-score: 0.85, ROC: 0.98 | The model was able to provide detection performance similar to that of trained dentists | |
| Disease Detection | Pauwels et al., 2021 [47] | CNN + Transfer Learning | Intraoral radiographs | 280 images, Train: 168 images, Test 112 images 168/112 | Periapical lesions | 3 oral radiologists | Five-fold | Sensitivity: 0.79, Specificity: 0.88, ROC-AUC: 0.86 | The study explored the potential of CNN-based assessment of periapical lesions and achieved superior performance compared to human observers |
| Fukuda et al., 2020 [49] | DetectNet (CNN) | Panoramic radiographs | 300 images, Train: 240 images, Test: 60 images 240/60 | Vertical root fracture | Three observers | Five-fold | Recall: 0.75, Precision: 0.93, F1-score: 0.83 | The model was useful in identifying teeth with vertical root fracture | |
| Orhan et al., 2020 [48] | Deep CNN (U-Net) | 3D CBCT images | 153 images | Periapical lesions | Two Oral and maxillofacial radiologist | NA | Recall: 0.89, Precision: 0.95, F1-score: 0.93 | The model was able to detect periapical pathosis with 92.8% reliability | |
| Ekert et al. 2019 [50] | 7 layer feed forward CNN | Panoramic radiographs | 85 images, Train: 56 images, Valid: 29 images 56/29 | Apical lesions | Six independent and experienced dentists | Ten-fold | Avg AUC: 0.85, Confidence Interval (CI): 95% | A moderately deep CNN was helpful in detecting apical lesions and can be helpful in reducing dentists’ diagnostic efforts | |
| Disease diagnosis | Bayraktar & Ayan, 2022 [51] | Deep CNN (YOLO) pretrained using DarkNet-53 | Bitewing radiographs | 1000 images, Train: 800 images, Test: 200 images 800/200 | Interproximal caries lesions | Two experienced dentists | Hold-out | Accuracy: 0.945, Sensitivity: 0.722, Specificity: 0.981, PPV: 0.865, NPV: 0.954, AUC: 0.871 | The study shows promising outcomes for detecting caries in bitewing images achieving an accuracy above 90% |
| Disease classification | Hiraiwa et al., 2019 [52] | AlexNet GoogleNet | CBCT images and panoramic images | Training image patches: Single root group: 11,472, Extra root group: 11,004, Testing image patches: Single root group:32, Extra root group—32 | Assessing number of distal roots of mandibular first molars | Two radiologists | Five-fold | Accuracy: 87.4, Sensitivity: 77.3, Specificity: 97.1, PPV: 96.3, NPV: 81.8, AUC: 0.87 | The model achieved detection performance superior to that of dental radiologists in differentiating whether distal root was single or with an extra root |
| Disease segmentation | Casalegno et al., 2019 [53] | Symmetric Autoencoder with skip connections similar to U-Net and encoding path similar to VGG16 | Near Infrared transillumination (TI) images | 217 grayscale images of upper and lower molars and premolars, Train: 185 images, Test: 32 images 185/32 | Proximal and occlusal caries lesion | Two dentists | Monte Carlo | IoU score: Occlusal—0.49, Proximal—0.49, AUROC: Occlusal—0.83, Proximal—0.85 | The proposed system has the potential to support dentists by providing higher throughput in detecting occlusal and proximal lesions |
| Disease diagnosis | Kositbowornchai et al., 2013 [20] | Probabilistic Neural Network | Intraoral radiographs | 200 images (50 sound and 150 vertical root fractures), Train: 120, Test: 80 120/80 | Vertical root fracture detection | N/A | Three-fold | Sensitivity: 0.98, Specificity: 0.905, Accuracy: 0.957 | The model was helpful in diagnosing vertical root fractures using intraoral digital radiographs |