Table 2.
Details of the studies that have reported on the application of AI-based models in endodontics.
| Serial No. | Authors | Year of Publication | Study Design | Algorithm Architecture | Objective of the Study | No. of Patients/Images/Photographs for Testing | Study Factor | Modality | Comparison If Any | Evaluation Accuracy/Average Accuracy/Statistical Significance | Results: (+) Effective, (−) Non-Effective, (N) Neutral |
Outcomes | Authors’ Suggestions/Conclusions |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Saghiri et al. [25] | 2011 | Comparative study | ANNs | AI-based model for locating the minor apical foramen | 50 samples | Apical foramen | Intraoral radiographs | Two experienced endodontists | For 93% of the samples, the model determined the location of the apical foramen correctly | (+) Effective | ANN-based model demonstrated good accuracy in detecting the apical foramen | The AI model can be useful for secondary opinion in order to achieve better clinical decision-making |
| 2 | Saghiri et al. [26] | 2012 | Comparative study | ANNs | AI-based model for determining the working length | 50 samples | Working length | Intraoral radiographs | Dentist | AI model demonstrated 96% accuracy in comparison with the experienced endodontists whose accuracy was 76% | (+) Effective | AI model demonstrated more accuracy in determining the working length in comparison with experienced endodontists | This model was efficient in determining the working length |
| 3 | Kositbowornchai et al. [27] | 2013 | Comparative study | ANNs | AI-based model for determining vertical root fracture (VRFs) | 200 samples (80 for training, 120 for testing) |
Vertical root fracture | Digital radiographs | Between groups | Sensitivity (98%), specificity (90.5%), and accuracy (95.7%) | (+) Effective | This AI model displayed sufficient sensitivity, specificity, and accuracy | This model make can be useful for making correct interpretations of root fractures |
| 4 | Tumbelaka et al. [28] | 2014 | Observational study | PNNs | AI-based model for identifying pulpitis | 20 samples | Pulpitis | Periapical radiographs | None | Mean square error around 0.0003 | (+) Effective | This model precisely diagnosed reversible and irreversible pulpitis | In order to obtain a better diagnosis, radiographs have to be digitalized |
| 5 | Johari et al. [29] | 2017 | Comparative study | Probabilistic neural networks (PNNs) | AI-based model for diagnosing VRFs in intact and endodontically treated teeth | 240 samples | Vertical root fracture | CBCT images | Other state-of-the-art approaches | Accuracy of 96.6%, sensitivity of 93.3%, and specificity of 100% | (+) Effective | This model is efficient in diagnosing VRFs using CBCT images | Additional training of AI-based models is required before clinical use |
| 6 | Shah et al. [30] | 2018 | Comparative study | CNNs | AI model for automatically detecting cracks in teeth | 6 samples | Cracked teeth | CBCT images | Frangi’s vessel enhancement algorithm | Mean ROC was 0.97 | (+) Effective | This model was efficient in detecting cracked teeth | The model can detect cracked teeth in earlier stages and prevent pain and suffering associated with them |
| 7 | Ekert et al. [31] | 2019 | Comparative study | CNNs | AI model for detecting apical lesions | 85 samples | Apical lesions | Panoramic radiographs | 6 independent examiners | AUC was 0.85 (0.04). Sensitivity was 0.65 (0.12) and specificity was 0.87 (0.04) | (+) Effective | This model showed a satisfying ability to detect apical lesions | Sensitivity of the model needs to be improved before application in clinics |
| 8 | Fukuda et al. [32] | 2019 | Comparative study | CNNs | AI model for detecting vertical root fractures (VRFs) |
300 samples (240 for training and 60 for testing) | Vertical root fracture | Panoramic radiographs | 2 radiologists and 1 endodontist | Recall was 0.75, precision was 0.93, and F measure was 0.83 | (+) Effective | This model showed promising results in detecting VRFs | This model has to be trained and applied on datasets from other hospitals |
| 9 | Hiraiwa et al. [33] | 2019 | Comparative study | ANNs | AI model for assessing the root morphology of the mandibular first molar | 760 samples | Root morphology | Panoramic and CBCT images | Radiologist | Accuracy of 86.9% | (+) Effective | The model displayed high accuracy in diagnosing a single or extra root in the distal roots of mandibular first molars | This model displayed a high level of diagnostic ability |
| 10 | Mallishery et al. [34] | 2019 | Comparative study | ML | AI-based ML model for predicting the difficulty level of the case | 500 samples | Case difficulty | Datasets | 2 endodontists | Sensitivity of 94.96% | (+) Effective | This model displayed an excellent prediction of case difficulty | This model displayed excellent prediction ability which can increase the speed of decision-making and referrals |
| 11 | Setzer et al. [35] | 2020 | Comparative study | CNNs | A deep learning model for automated segmentation of CBCT images and detecting periapical lesions | 20 CBCT images (16 CBCT images for training, 4 CBCT images for validation) |
Apical lesions | CBCT images | 1 radiologist, 1 endodontist, and 1 senior graduate | Accuracy of 0.93 and specificity of 0.88 | (+) Effective | With a limited CBCT training, this model displayed excellent results in detecting the lesion | This model can aid clinicians with automated lesion detection |
| 12 | Orhan et al. [36] | 2020 | Comparative study | ANNs | AI model for detecting periapical pathosis | 153 samples | Periapical lesions | CBCT images | 1 maxillofacial radiologist | Reliability of 92.8% in correctly detecting periapical lesions | (+) Effective | There was no difference in the accuracy of humans and AI model in detecting apical lesions | This model will be useful for detecting periapical pathosis in clinical scenarios |
| 13 | Endres et al. [37] | 2020 | Comparative study | CNNs | Deep learning model for detecting periapical disease |
197 samples (95 images for training and 102 images for testing) | Periapical disease | Panoramic radiographs | 24 oral and maxillofacial (OMF) surgeons |
Average precision of 0.60 and an F1 score of 0.58 | (+) Effective | This deep learning algorithm achieved a better performance than 14 of 24 OMF surgeons | The deep learning model has potential to assist OMF surgeons in detecting periapical lucencies |
| 14 | Qiao et al. [38] | 2021 | Comparative study | CNNs | Deep learning models for root canal length measurement |
21 samples | Root canal length | Tooth | Dual-frequency impedance ratio method | Accuracy of 95% | (+) Effective | This model demonstrated better accuracy in comparison with other models | The performance of this model can be enhanced by increasing the number of samples |
| 15 | Sherwood et al. [39] | 2021 | Comparative study | CNNs | Deep learning model for classifying C-shaped canal anatomy in mandibular second molars |
135 samples (100 images for training and 35 images for testing) | Canal shapes | CBCT images | U-Net, residual U-Net, and Xception U-Net architectures | The mean Dice coefficients were 0.768 ± 0.0349 for Xception U-Net, 0.736 ± 0.0297 for residual U-Net, and 0.660 ± 0.0354 for U-Net on the test dataset | (+) Effective | Both Xception U-Net and residual U-Net performed significantly better than U-Net | Deep learning models can aid clinicians in detecting and classifying C-shaped canal anatomy |
| 16 | Li et al. [40] | 2021 | Comparative study | CNNs | Deep learning model for detecting apical lesions | 460 samples (322 images for training and 138 images for testing) | Apical lesions | Periapical radiographs | 3 experienced dentists | Diagnostic accuracy of the model was 92.5% | (+) Effective | Deep neural models demonstrated excellent accuracy in detecting the periapical lesions | This automated model allows dentists to make the diagnosis process shorter and more efficient |
| 17 | Vicory et al. [41] | 2021 | Comparative study | ML | AI model for detecting tooth microfractures | 36 samples | Tooth microfractures | High-resolution (hr) CBCT and micro-CT scans |
Direction–projection–permutation | Significant separation result | (+) Effective | The data suggest that this approach can be applied to hr-CBCT (clinically) when the images are not over-processed | Early detection of microfractures can help in planning appropriate treatment |
| 18 | Zheng et al. [42] | 2021 | Comparative study | CNNs | Deep learning model for detecting deep caries and pulpitis | 844 samples (717 images for training and 127 images for testing) | Deep caries and pulpitis | Periapical radiographs | VGG19, Inception V3, ResNet18, 5 experienced dentists | Accuracy of 0.86, precision of 0.85, sensitivity of 0.86, specificity of 0.86, and AUC of 0.94 | (+) Effective | ResNet18 demonstrated the best performance also in comparison with experienced dentists | The promising potential of this model can be applied for clinical diagnosis |
| 19 | Moidu et al. [42] | 2021 | Comparative study | CNNs | Deep learning model for categorization of endodontic lesions | 1950 samples |
Periapical lesions | Periapical radiographs | 3 endodontists | Sensitivity/recall of 92.1%, 76% specificity, 86.4% positive predictive value, and 86.1% negative predictive value | (+) Effective | The model exhibited excellent sensitivity, positive predictive value, and negative predictive value | This AI model can be beneficial for clinicians and researchers |
| 20 | Pauwels et al. [44] | 2021 | Comparative study | CNNs | Deep learning model for detecting periapical lesions | 10 samples |
Periapical lesions | Periapical radiographs | 3 oral radiologists | Mean sensitivity of 0.87, specificity of 0.98, and ROC-AUC of 0.93 | (+) Effective | This CNN model displayed perfect accuracy for the validation data | This model showed promising results in detecting periapical lesions |
| 21 | Jeon et al. [45] | 2021 | Comparative study | CNNs | Deep learning model for predicting C-shaped canals in mandibular second molars | 2040 samples (1632 images for training and 408 images for testing) |
C-shaped canals | Panoramic radiographs | 1 experienced radiologist and 1 experienced endodontist | Accuracy of 95.1, sensitivity of 92.7, specificity of 97.0, and precision of 95.9% | (+) Effective | This CNN model displayed significant accuracy in predicting C-shaped canals |
This model can assist clinicians with dental image interpretation |
| 22 | Guo et al. [46] | 2021 | Comparative study | ANNs | Radial basis function neural network (RBFN)-based AI model for predicting thrust force and torque for root canal preparation | 2 samples |
Thrust force and torque | CT scans | Comparative ANN model | Prediction error less than 14% | (+) Effective | This model displayed an excellent prediction of thrust force and torque in canal preparation | Can be useful for instructing dentists during root canal preparations and also for improving the geometrical design of nickel titanium files |
| 23 | Lin et al. [47] | 2021 | Comparative study | ANNs | AI model for automatic and accurate segmentation of the pulp cavity and tooth | 30 samples (25 sets for training and 5 sets for testing) | Segmentation of the pulp cavity | CBCT images | 1 experienced endodontist | Dice similarity coefficient of 96.20% ± 0.58%, precision rate of 97.31% ± 0.38%, recall rate of 95.11% ± 0.97%, average symmetric surface distance of 0.09 ± 0.01 mm, and Hausdorff distance of 1.54 ± 0.51 mm in the tooth and Dice similarity coefficient of 86.75% ± 2.42%, precision rate of 84.45% ± 7.77%, recall rate of 89.94% ± 4.56%, average symmetric surface distance of 0.08 ± 0.02 mm, and Hausdorff distance 1.99 ± 0.67 mm in the pulp cavity |
(+) Effective | The analysis performed by the model was better than that of the experienced endodontist | This model demonstrated excellent accuracy and hence can be applied in research and clinical tasks in order to achieve better endodontic diagnosis and therapy |
| 24 | Gao et al. [48] | 2021 | Observational study | ANNs | Backpropagation (BP) AI model for predicting postoperative pain following root canal treatment | 300 samples (210 for training, 45 for validating, and 45 for testing) | Postoperative pain | Datasets | None | Accuracy of prediction was 95.60% | (+) Effective | This model displayed an excellent prediction of postoperative pain following RCT | The results displayed by this model have shown clinical feasibility and clinical application value |
| 25 | Ngoc et al. [49] | 2021 | Comparative study | CNNs | AI-based model for diagnosis of periapical lesions | 130 samples | Periapical lesions | Bitewing images | Endodontists | Sensitivity of 89.5, specificity of 97.9, and accuracy of 95.6% | (+) Effective | This model displayed excellent performance and can be used as a support tool in the diagnosis of periapical lesions | This model can be used in teledentistry for the diagnosis of periapical diseases where there is a lack of dentists |
| 26 | Kirnbauer et al. [50] | 2022 | Observational study | CNNs | AI model for the automated detection of periapical lesions | 144 samples | Periapical lesions | CBCT images | None | Sensitivity of 97.1% and specificity of 88.0% for lesion detection | (+) Effective | This AI model displayed excellent results compared with related literature | This model can be applied for testing under clinical conditions |
| 27 | Herbst et al. [51] | 2022 | Comparative study | ML | AI-based ML model for predicting failure of root canal treatment | 591 samples | Root canal failure | Datasets | Random forest, gradient boosting machine, extreme gradient boosting, predictive modeling |
logR 0.63, gradient boosting machine (GBM) 0.59, random forest (RF) 0.59, extreme gradient boosting (XGB) 0.60 | (N) Neutral | This study found tooth-level factors to be associated with failure | With this AI model, predicting failure was only limitedly possible |
| 28 | Bayrakdar et al. [52] | 2022 | Observational study | CNNs | AI-based deep convolutional neural network (D-CNN) model for the segmentation of apical lesions | 470 samples | Apical lesions | Panoramic radiographs | None | Sensitivity of 0.92, precision of 0.84, and F1-score of 0.88 | (+) Effective | This AI model was efficient in evaluating periapical pathology |
This AI model may facilitate clinicians in the assessment of periapical pathology |
| 29 | Zhao et al. [53] | 2022 | Comparative study | CNNs | AI model for evaluating the curative effect after treatment of dental pulp disease (DPD) | 120 samples | Dental pulp disease | Radiographs and CBCT images | Control group with healthy teeth | Segmentation accuracy was 85.5%; diagnostic rate of X-ray was 43.7% and diagnostic rate of CBCT was 100% | (+) Effective | CBCT evaluation using an AI model can be an effective method for evaluating the curative effect of dental pulp disease treatment during and after the surgery | This model has a higher application prospect in the diagnosis and treatment of DPD |
| 30 | Hamdan et al. [54] | 2022 | Comparative study | CNNs | AI model for detecting apical radiolucencies | 68 samples | Apical radiolucencies | Periapical radiographs | Eight experienced specialists | Alternative free-response receiver operating characteristic (AFROC) of 0.892, specificity of 0.931, and sensitivity of 0.733 | (+) Effective | This model has the potential to improve the diagnostic efficacy of clinicians | This AI model enhances clinicians’ abilities to detect apical radiolucencies |
| 31 | Calazans et al. [55] | 2022 | Comparative study | CNNs | AI models for classifying periapical lesions | 1000 samples (training 60%, validation 20%, testing 20%) | Periapical lesions | CBCT scans | Experienced oral and maxillofacial radiologist | Accuracy of 70%, specificity of 92.39% |
(+) Effective | DenseNet-121 network was superior to VGG-16 and human experts | The proposed models displayed a satisfactory classification performance |
| 32 | Yang et al. [56] | 2022 | Comparative study | CNNs | AI-based deep learning model for classifying C-shaped canals in mandibular second molars | 1000 samples | C-shaped canals | Periapical and panoramic radiographs | Specialist and general clinician | AUC of 0.98 on periapical and AUC of 0.95 on panoramic | (+) Effective | This model displayed high accuracy in predicting the C-shaped canal in both periapical and panoramic images and was similar to the performance of a specialist and better than a general dentist | This model was effective in diagnosing C-shaped canals and therefore can be a valuable aid for clinicians and also in dental education |
| 33 | Xu et al. [57] | 2022 | Comparative study | ML | AI-based models for identifying the history of root canal therapy | 920 samples (736 for training and 184 for testing) | Root canal therapy | Datasets | VGG16, VGG19, and ResNet50 | Accuracies were above 95% and AUC area was 0.99 | (+) Effective | This model displayed excellent accuracy and can aid in clinical auxiliary diagnosis based on image display | This AI-assisted diagnosis of oral medical images can be effectively promoted for clinical practice |
| 34 | Qu et al. [58] | 2022 | Comparative study | ML | AI-based machine learning models for predicting prognosis of endodontic microsurgery | 234 samples (80% for the training set and 20% for the test set) | Predicting prognosis | Datasets | Gradient boosting machine (GBM) and random forest (RF) models | Accuracy of 0.80, sensitivity of 0.92, specificity of 0.71, positive predictive value (PPV) of 0.71, negative predictive value (NPV) of 0.92, F1 score of 0.80, and area under the curve (AUC) of 0.88 | (+) Effective | The GBM model outperformed the RF model slightly on the dataset | The models can improve efficiency and assist clinicians in decision-making |
| 35 | Li et al. [59] | 2022 | Comparative study | ANNs | AI-based anatomy-guided multibranch transformer (AGMB-Transformer) network for assessing the result of root canal therapy |
245 samples | Root canal therapy evaluation | Datasets | 2 experienced specialists and other models (ResNet50, ResNeXt50, GCNet50, BoTNet50) | Accuracy ranged from 57.96% to 90.20%, AUC of 95.63%, sensitivity of 91.39%, specificity of 95.09%, F1 score of 90.48% | (+) Effective | This model achieved a highly accurate evaluation | The performance of this model has important clinical value in reducing the workload of endodontists |
| 36 | Hu et al. [60] | 2022 | Comparative study | CNNs | AI-based deep learning models for diagnosing vertical root fracture | 276 samples | Vertical root fracture | CBCT images | 2 experienced radiologists, ResNet50, VGG19, and DenseNet169 |
The accuracy, sensitivity, specificity, and AUC were 97.8%, 97.0%, 98.5%, and 0.99 | (+) Effective | ResNet50 presented the highest accuracy and sensitivity for diagnosing VRF teeth | ResNet50 presented the highest diagnostic efficiency in comparison with other models. Hence, this model can be used as an auxiliary diagnostic technique to screen for VRF teeth |
| 37 | Vasdev et al. [61] | 2022 | Comparative study | CNNs | AI-based deep learning model for detecting healthy and non-healthy periapical images | 16,000 samples | Periapical lesions | Periapical radiographs | ResNet-18, ResNet-34, and AlexNet | Accuracy of 0.852, precision and F1 score of 0.850 | (+) Effective | This AlexNet model outperformed the other models | This model generalizes effectively to previously unseen data and can aid clinicians in diagnosing a variety of dental diseases |
Footnotes: ML = machine learning, ANNs = artificial neural networks, CNNs = convolutional neural networks, DCNNs = deep neural networks, c-index = concordance index, CT = computed tomography, CBCT = cone-beam computed tomography, OCT = optical coherence tomography.