Table 1.
Characteristics of studies describing machine learning-based artificial intelligence (AI) models applied in dentomaxillofacial radiology (DMFR).
| Author (Year) | Application | Imaging Modality | AI Technique | Image Data Set Used to Develop the AI Model | Independent Testing Image Data Set / Validation Technique | Performance |
|---|---|---|---|---|---|---|
| Diagnosis of Dental and Maxillofacial Diseases | ||||||
| Okada [16] (2015) | Diagnosis of periapical cysts and granuloma | CBCT | LDA | 28 scans from patients with periapical cysts or granuloma | 7-fold CV | 94.1% (accuracy) |
| Abdolali [17] (2017) | Diagnosis of radicular cysts, dentigerous cysts, and keratocysts | CBCT | SVM; SDA | 96 scans from patients with radicular cysts, dentigerous cysts, or keratocysts | 3-fold CV | 94.29–96.48% (accuracy) |
| Yilmaz [18] (2017) | Diagnosis of periapical cysts and keratocysts | CBCT | k-NN; Naïve Bayes; Decision tree; Random forest; NN; SVM | 50 scans from patients with cysts or tumors | 10-fold CV/LOOCV | 94–100% (accuracy) |
| 25 scans from patients with cysts or tumors | 25 scans from patients with cysts or tumors | |||||
| Lee [19] (2020) | Diagnosis of periapical cysts, dentigerous cysts, and keratocysts | Panoramic radiography and CBCT | CNN | 912 panoramic images and 789 CBCT scans | 228 panoramic images and 197 CBCT scans | Panoramic radiography 0.847 (AUC); 88.2% (sensitivity); 77.0% (specificity) CBCT 0.914 (AUC); 96.1% (sensitivity); 77.1% (specificity) |
| Orhan [28] (2020) | Diagnosis of periapical pathology | CBCT | CNN | 3900 scans acquired using multiple FOVs from 2800 patients with periapical lesions and 1100 subjects without periapical lesions | 109 scans acquired using multiple FOVs from 153 patients with periapical lesions | 92.8% (accuracy) |
| Abdolali [29] (2019) | Diagnosis of radiolucent lesion, maxillary sinus perforation, unerupted tooth, and root fracture | CBCT | Symmetry-based analysis model | 686 scans acquired using a large FOV (12 × 15 × 15 cm3), collected from several dental imaging centers in Iran | 459 scans acquired using a large FOV (12 × 15 × 15 cm3), collected from several dental imaging centers in Iran | 0.85–0.92 (DSC) |
| Johari [30] (2017) | Detection of vertical root fractures | Periapical radiography and CBCT | CNN | 180 periapical radiographs and 180 CBCT scans of the extracted teeth | 60 periapical radiographs and 60 CBCT scans of the extracted teeth | Periapical radiography 70.0% (accuracy); 97.8% (sensitivity); 67.6% (specificity) CBCT 96.6% (accuracy); 93.3% (sensitivity); 100% (specificity) |
| Kise [32] (2019) | Diagnosis of Sjögren’s syndrome | CT | CNN | 400 scans (200 from 20 SjS patients and 200 from 20 control subjects) acquired using a large FOV | 100 scans (50 from 5 SjS patients and 50 from 5 control subjects) acquired using a large FOV | 96.0% (accuracy); 100% (sensitivity); 92.0% (specificity) |
| Kann [31] (2018) | Detection of lymph node metastasis and extranodal extension in patients with head and neck cancer | Contrast-enhanced CT | CNN | Images of 2875 CT-segmented lymph node samples with correlating pathology labels | Images of 131 lymph nodes (76 negative and 55 positive) | 0.91 (AUC) |
| Ariji [20] (2019) | Detection of lymph node metastasis in patients with oral cancer | Contrast-enhanced CT | CNN | Images of 441 lymph nodes (314 negative and 127 positive) from 45 patients | 5-fold CV | 78.2% (accuracy); 75.4% (sensitivity); 81.0% (specificity), 0.80 (AUC) |
| Localization of Anatomical Landmarks for Orthodontic and Orthognathic Treatment Planning | ||||||
| Cheng [33] (2011) | Localization of the odontoid process of the second vertebra | CBCT | Random forest | 50 scans | 23 scans | 3.15 mm (mean deviation) |
| Shahidi [34] (2014) | Localization of 14 anatomical landmarks | CBCT | Feature-based and voxel similarity-based algorithms | 8 scans acquired using a large FOV from subjects aged 10–45 years | 20 scans acquired using a large FOV from subjects aged 10–45 years | 3.40 mm (mean deviation) |
| Montufar [21] (2018) | Localization of 18 anatomical landmarks | CBCT | Active shape model | 24 scans acquired using a large FOV | LOOCV | 3.64 mm (mean deviation) |
| Montufar [22] (2018) | Localization of 18 anatomical landmarks | CBCT | Active shape model | 24 scans acquired using a large FOV | LOOCV | 2.51 mm (mean deviation) |
| Torosdagli [35] (2019) | Localization of 9 anatomical landmarks | CBCT | CNN | 50 scans | 48 scans | 0.9382 (DSC); 93.42% (sensitivity); 99.97% (specificity), |
| Improvement of Image Quality | ||||||
| Park [36] (2018) | Improvement of image resolution | CT | CNN | 52 scans | 13 scans | The CNN network can yield high-resolution images based on low-resolution images |
| Minnema [23] (2019) | Segmentation of CBCT scans affected by metal artifacts | CBCT | CNN | 20 scans | Leave-2-out CV | The CNN network can accurately segment bony structures in CBCT scans affected by metal artifacts |
| Other | ||||||
| Miki [38] (2017) | Tooth classification | CBCT | CNN | 42 scans with the diameter of the FOV ranged from 5.1 to 20 cm | 10 scans with the diameter of the FOV ranged from 5.1 to 20 cm | 88.8% (accuracy) |
AI, artificial intelligence; AUC, area under the receiver operating characteristic curve; CBCT, cone beam computed tomography; CNN, convolutional neural network; CT, computed tomography; CV, cross validation; DSC, dice similarity coefficient; FOV, field of view; k-NN, k-nearest neighbors; LDA, linear discriminant analysis; LOOCV, leave-one-out cross-validation; NN, neural network; SDA, sparse discriminant analysis; SjS, Sjögren’s syndrome; SVM, support vector machine.