Table 1.
Description of included studies.
| Study | Country | Year | Data Type | Subject Total | ML Architecture |
Annotators | Performance Comparison | CNN Performance Comment | Brief Description | |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Papantonopoulos [45] | Greece | 2014 | Patient data | 29 | MLP ANN | n.a. | Not comparative | ANN’s gave 90–98% accuracy in classifying patients into AgP or CP. | ANNs used to classify periodontitis by immune response profile to aggressive periodontitis (AgP) or chronic periodontitis (CP) class. |
| 2 | Bezruk [47] | Ukraine | 2017 | Saliva | 141 | CNN (no description) |
n.a. | Not comparative | Precision of CNN 0.8 in predicting gingivitis based upon crevicular fluid markers. | CNN was used for the learning task to build an information model of salivary lipid peroxidation and periodontal status and to evaluate the correlation between antioxidant levels in unstimulated saliva and inflammation in periodontal tissues. |
| 3 | Rana [48] | USA | 2017 | Photographs | 405 | CNN Autoencoder | Dentist | Not comparative | AU ROC curve of 0.746 for classifier to distinguish between inflamed and healthy gingiva. | Machine learning classifier used to provide pixel-wise inflammation segmentations from photographs of colour-augmented intraoral images. |
| 4 | Feres [41] | Brazil | 2018 | Plaque | 435 | SVM | n.a. | Not comparative | AUC > 0.95 for SVM to distinguish between disease and health. AUC for ability to distinguish between CP and AgP was 0.83. | SVM was used to assess whether 40 bacterial species could be used to classify patients into CP, AgP, or periodontal health. |
| 5 | Lee [49] | South Korea | 2018 | Periapical | 1740 | CNN encoder + 3 dense layers |
Periodontist | Periodontist | CNN showed AU ROC curve of 73.4–82.6 (95% CI 60.9–91.1) in predicting hopeless teeth. | The accuracy of predicting extraction was evaluated and compared between the CNN and blinded board-certified periodontists using 64 premolars and 64 molars diagnosed as severe n the test dataset. For premolars, the deep CNN had an accuracy of 82.8% |
| 6 | Yoon [50] | USA | 2018 | Patient data | 4623 | Deep neural network-BigML | n.a. | Not comparative | DNN used as multi-regressional tool found correlation between ageing and mobility. | 78 variables assessed by DNN were used to find a correlation that can predict tooth mobility. |
| 7 | Aberin [51] | Philippines | 2019 | Plaque | 1000 | AlexNet | Pathologists | Not comparative | Accuracy in predicting health or periodontitis from plaque slides reported at 75%. | CNN was used to classify which microscopic dental plaque images were associated with gingival health. |
| 8 | Askarian [52] | USA | 2019 | Photographs | 30 | SVM | n.a. | Not comparative | 94.3% accuracy of SVM in detection of periodontal infection. | Smartphone-based standardised photograph detection using CNN to classify gingival disease presence. |
| 9 | Duong [31] | Canada | 2019 | Ultrasound | 35 | U-Net | n.a. | Orthodontist | CNN yielded 75% average dicemetric for ultrasound segmentation. | The proposed method was evaluated over 15 ultrasound images of teeth acquired from porcine specimens. |
| 10 | Hegde [39] | USA | 2019 | Patient data | 41,543 | SVM | n.a. | Not comparative | Comparison of ML vs. MLP vs. RF vs. SVM for data analysis. Similar accuracy found between all methods. | The objective was to develop a predictive model using medical-dental data from an integrated electronic health record (iEHR) to identify individuals with undiagnosed diabetes mellitus (DM) in dental settings. |
| 11 | Joo [53] | South Korea | 2019 | Photographs | 451 | CNN encoder + 1 dense layer |
n.a. | Not comparative | Reported CNN accuracy of 70–81% for validation data. | Descriptive analysis of preliminary data for concepts of imaging analysis. |
| 12 | Kim [54] | South Korea | 2019 | Panoramic | 12,179 | DeNTNet | Hygienists | Hygienists | Superior F1 score (0.75 vs. 0.69), PPV (0.73 vs. 0.62), and AUC (0.95 vs. 0.85) for balanced setting DeNTNet vs. clinicians for assessing periodontal bone loss. | CNN used to develop an automated diagnostic support system assessing periodontal bone loss in panoramic dental radiographs. |
| 13 | Krois [55] | Germany | 2019 | Panoramic | 85 | CNN encoder + 3 dense layers |
Dentist | Dentists | CNN performed less accurately than the original examiner segmentation and independent dentists’ observers. | CNNs used to detect periodontal bone loss (PBL) on panoramic dental radiographs. |
| 14 | Moriyama [56] | Japan | 2019 | Photographs | 820 | AlexNet | Dentist | Not comparative | Changes in ROC curves can have a significant effect on outcomes—looking at predicted pocket depth photographs and distorting images to improve accuracy. | CNN was used to establish if there is a correlation between pocket depth probing and images of the diseased area. |
| 15 | Yauney [57] | USA | 2019 | Patient data | 1215 | EED-net (custom net) |
Dentist | Not comparative | AUC of 0.677 for prediction of periodontal disease based on intraoral fluorescent porphyrin biomarker imaging. | CNN was used to establish a link between intraoral fluorescent porphyrin biomarker imaging, clinical examinations, and systemic health conditions with periodontal disease. |
| 16 | Alalharith [58] | Saudi Arabia | 2020 | Photographs | 134 | Faster R-CNN | Dentist | Previously published outcomes | Faster R-CNN had tooth detection accuracy of 100% to determine region of interest and 77.12% accuracy to detect inflammation. | An evaluation of the effectiveness of deep learning based CNNs for the pre-emptive detection and diagnosis of periodontal disease and gingivitis by using intraoral images. |
| 17 | Bayrakdar [59] | Turkey | 2020 | Panoramic | 2276 | GoogLeNet Inception v3 | Radiologist and periodontist | Radiologist and periodontist | CNN showed 0.9 accuracy to detect alveolar bone loss. | CNN used to detect alveolar bone loss from dental panoramic radiographic images. |
| 18 | Chang [60] | South Korea | 2020 | Panoramic | 340 | ResNet | Radiologist | Radiologists | 0.8–0.9 agreement between radiologists and CNN performance. | Automatic method for staging periodontitis on dental panoramic radiographs using the deep learning hybrid method. |
| 19 | Chen [61] | China (and the UK) | 2020 | Photographs | 180 | ANN (no description) |
n.a. | Not comparative | ANN accuracy of 71–75.44% for presence of gingivitis from photographs. | Visual recognition of gingivitis testing a novel ANN for binary classification exercise—gingivitis or healthy. |
| 20 | Farhadian [38] | Iran | 2020 | Patient data | 320 | SVM | n.a. | Not comparative | The SVM model gave an 88.4% accuracy to diagnose periodontal disease. | The study aimed to design a support vector machine (SVM)-based decision-making support system to diagnose various periodontal diseases. |
| 21 | Huang [40] | China | 2020 | Gingival crevicular fluid | 25 | SVM | n.a. | n.a. | Classification models achieved greater than or equal to 91% in classifying SP patients, with LDA being the highest at 97.5% accuracy. | This study highlights the potential of antibody arrays to diagnose severe periodontal disease by testing five models (SVM, RF, kNN, LDA, CART). |
| 22 | Kim [42] | South Korea | 2020 | Saliva | 692 | SVM | n.a. | n.a. | Accuracy ranged from 0.78 to 0.93 comparing neural network, random forest, and support vector machines with linear kernel, and regularised logistic regression in the R caret package. | CNN was used to assess whether biomarkers can differentiate between healthy controls and those with differing severities of periodontitis. |
| 23 | Kong [46] | China | 2020 | Panoramic | 2602 | EED-net (custom net) |
Expert? | Expert? | The custom CNN performed better than U-Net or FCN-8, all with accuracies above 98% for anatomical segmentation. | CNN was used to complete maxillofacial segmentation of images, including periodontal bone loss recognition. |
| 24 | Lee [62] | South Korea | 2020 | Periapicals and panoramic | 10,770 | GoogLeNet Inception v3 | Periodontist | Periodontist | The CNN (0.95) performed better than human (0.90) for OPGs, but the same for PAs (0.97). | CNN used for identification of implants systems and their associated health. |
| 25 | Li [63] | Saudi Arabia | 2020 | Panoramic | 302 | R-CNN | Dentist | Other CNNs and dentist | Proposed architecture gave accuracies of 93% for detecting no periodontitis, 89% for mild, 95% for moderate, and 99% for severe. | This study compared different CNN models for bone loss recognition. |
| 26 | Moran [64] | Brazil | 2020 | Periapicals | 467 | ResNet, Inception | Radiologist and dentist | Compares two CNN approaches for accuracy | AUC ROC curve for ResNet and Inception was 0.86 for identification of regions of periodontal bone destruction. | Assessment of whether a CNN can recognise of periodontal bone loss improve post-image enhancement? |
| 27 | Romm [65] | USA | 2020 | Metabolites | N/A | CNN (No description), PCA |
n.a. | n.a. | Oral cancer identified rather than a periodontal disease with 81.28% accuracy. | CNN to analyse metabolite sets for different oral diseases to distinguish between different forms of oral disease. |
| 28 | Shimpi [43] | USA | 2020 | Patient data | N/A | SVM, ANN | n.a. | n.a. | ANN presented more reliable outcomes than NB, LR, and SVM. | The study reviewed classic and CNN regression to assess accuracy in prediction for periodontal risk assessment based on EHIR. |
| 29 | Thanathornwong [66] | Thailand | 2020 | Panoramic | 100 | Faster R-CNN | Periodontist | Periodontist | 0.8 precision for identifying periodontally compromised teeth using radiographs. | CNN used to assess periodontally compromised teeth on OPG. |
| 30 | You [67] | China | 2020 | Photographs | 886 | DeepLabv3+ | Orthodontist | Orthodontist | No statistically significant difference in the ability to discern plaque on photographs compared to clinician. | CNN used to assess plaque presence in paediatric teeth. |
| 31 | Cetiner [68] | Turkey | 2021 | Patient data | 216 | MLP ANN | n.a. | n.a. | The DT was most accurate, with accuracy of 0.871 compared to LR (0.832) and LP (0.852). | Assessment of three models of data mining to provide a predictive decision model for peri-implant health. |
| 32 | Chen [69] | China | 2021 | Periapicals | 2900 | r-CNN | Dentist | n.a. | CNN used to locate periodontitis, caries, and PA pathology on PAs. | |
| 33 | Danks [70] | UK | 2021 | Periapicals | 340 | ResNet | Dentist | Dentist | Predicting periodontitis stage accuracy of 68.3%. | CNN used to find bone loss landmarks using different tools to provide staging of disease. |
| 34 | Kabir [30] | USA | 2021 | Periapicals | 700 | Custom CNN combining Res-Net and U-Net | Periodontitis | Periodontitis | Agreement between professors and HYNETS of 0.69. | CNN calibrated with bone loss on PAs applied to OPGs for staging and grading of whole-mouth periodontal status. |
| 35 | Khaleel [71] | Iraq | 2021 | Photographs | 120 | BAT algorithm, PCA, SOM | Dentist | n.a. | BAT method provided 95% accuracy against ground truth | Assessment of different algorithms’ efficacy in recognising gingival disease. |
| 36 | Kouznetsova [72] | USA | 2021 | Salivary metabolites | N/A | DNN | n.a. | n.a. | Model performance assessment only of different CNNs. | CNN predicts which molecules should be assessed for metabolic diagnosis of periodontitis or oral cancers. |
| 37 | Lee [35] | South Korea | 2021 | Panoramic | 530 | U-Net, Dense U-Net, ResNet, SegNet | Radiologists | Radiologists | The accuracy of the resulting model was 79.54%. | Assessment of a variety of CNN architectures for detecting and quantifying the missing teeth, bone loss, and staging on panoramic radiographs. |
| 38 | Li [73] | China | 2021 | Photographs | 3932 | Fnet, Lnet, cnet | Dentist | Dentists | Low agreement between three dentists and CNN in heatmap analysis. | CNN used for gingivitis detection photographs. |
| 39 | Li [29] | China | 2021 | Photographs | 110 | DeepLabv3+ | Dentist | n.a. | MobileNetV2 performed in a similar manner to Xception65; however, Mob, was 20× quicker. | Different CNNs trialled for RGB assessment of gingival tissues to assess inflamed gum detection on photographs. |
| 40 | Ma [74] | Taiwan | 2021 | Panoramic | 432 | ConvNet, U-Net | Unknown | n.a. | ConvNet analysis post U-Net segmentation—the model showed moderate levels of agreement (F2 score of between 0.523 and 0.903) and the ability to predict periodontitis and ASCVD. | CNNs used to assess for atherosclerotic cardiovascular disease and periodontitis on OPGs. |
| 41 | Moran [75] | Brazil | 2021 | Periapicals | 5 | Inception and for super-resolution SRCNN | Dentist | n.a. | Minimal enhancement of CNN performance was noted from super resolution, which may introduce additional artefacts. | The study compared the effects of super resolution methods on the ability of CNNs to perform segmentation and bone loss identification. |
| 42 | Ning [76] | Germany | 2021 | Saliva | N/A | DisGeNet, HisgAtlas | n.a. | n.a. | DL-based model able to predict immunosuppression genes in periodontitis with an accuracy of 92.78%. | CNN to identify immune subtypes of periodontitis and pivotal immunosuppression genes that discriminated periodontitis from the healthy. |
| 43 | Shang [32] | China | 2021 | Photographs | 7220 | U-Net | Dentist | Dentist | U-Net to have a 10% increased recognition of calculus, wear facets, gingivitis, and decay | Comparison of U-Net vs. comparison between U-Net and DeepLabV3/PSPNet architecture for image recognition on oral pictures for wear, decay, calculus, and gingivitis. |
| 44 | Wang [77] | USA | 2021 | Metabolites | N/A | FARDEEP | n.a. | n.a. | ML successfully used in logistic regression of plaque samples. | CNN is used as a processing tool for clinical, immune, and microbial profiling of peri-implantitis patients against health. |
| 45 | Jiang [37] | China | 2022 | Panoramic | 640 | U-Net, YOLO-v4 | Periodontist | Periodontist | Compared to the ground truth, accuracy of 0.77 was achieved by the proposed architecture. | CNN used to provide % bone loss and resorption/furcation lesion and staging of periodontal disease from OPGs. |
| 46 | Lee [36] | USA | 2022 | Periapicals | 693 | U-Net, ResNet | Dentist | Dentist | The accuracy of the diagnosis based upon staging and grading was 0.85 | Full mouth PA films were used to review bone loss—staging and grading were then performed. |
| 47 | Li [73] | China | 2022 | Photographs | 2884 | OCNet, Anet | Dentist | Dentist | CNN provided AUC prediction of 87.11% for gingivitis and 80.11% for calculus. | Research trialling different methods of segmentation to assess plaque on photographs of tooth surfaces (inc ‘dye labelling’). |
| 48 | Liu [78] | China | 2022 | Periapicals | 1670 | Faster R-CNN | Dentist | Dentist | The results confirm the advantage of utilising multiple CNN architectures for joint optimisation to increase UTC ROC boosts of up to 8%. | CNN used to assess implant marginal bone loss with dichotomous outcomes. |
| 49 | Pan [33] | USA | 2022 | Ultrasound | 627 | U-Net | Dentist | Dentist | Showed a significant difference between CNN outcome and dental experts’ labelling. | CNN was used to provide an estimation of gingival height in porcine models. |
| 50 | Zadrozny [34] | Poland | 2022 | Panoramic | 30 | U-Net | Radiologists | Dentists | Tested CNN showed unacceptable reliability for assessment of caries (ICC = 0.681) and periapical lesions (ICC = 0.619), but acceptable for fillings (ICC = 0.920), endodontically treated teeth (ICC = 0.948), and periodontal bone loss (ICC = 0.764). | Testing of commercially available product Diagnocat in the evaluation of panoramic radiographs. |