Table 3.

Conclusions and limitations of the included studies.

Author (Year)	Conclusion	Limitations (Risk of Bias *)
Okada [16] (2015)	The proposed model may assist clinicians to accurately differentiate periapical lesions.	A small training dataset *; Lacking data heterogeneity *; Dataset only consisted of scans from subjects with the condition of interest *; Lacking independent unseen testing data *; Manual ROI selection; Long execution time.
Abdolali [17] (2017)	The proposed model can improve the accuracy of the diagnosis of dentigerous cysts, radicular cysts, and keratocysts, and may have a significant impact on future AI diagnostic systems.	A small training dataset *; Lacking data heterogeneity *; Dataset only consisted of scans from subjects with the condition of interest *; Lacking independent unseen testing data *.
Yilmaz [18] (2017)	Periapical cysts and keratocysts can be classified with high accuracy with the proposed model. It can also contribute to the field of automated diagnosis of periapical lesions.	A small training dataset *; Lacking data heterogeneity *; Dataset only consisted of scans from subjects with the condition of interest *; Manual detection and segmentation of lesions.
Lee [19] (2020)	Periapical cysts, dentigerous cysts, and keratocysts can be effectively detected and diagnosed with the proposed deep CNN algorithm, but the diagnosis of these lesions using radiological data alone, without histological examination, is still challenging.	A relatively small training dataset *; Dataset only consisted of scans from subjects with the condition of interest *; Manual ROI selection; Potential overfitting problem in the training procedure *.
Orhan [28] (2020)	The proposed deep learning systems can be useful for detection and volumetric measurement of periapical lesions. The diagnostic performance was comparable to that of an oral and maxillofacial radiologist.	Relatively inaccurate segmentation of lesions in close contact with neighboring soft tissue
Abdolali [29] (2019)	The proposed system is effective and can automatically diagnose various maxillofacial lesions/conditions. It can facilitate the introduction of content-based image retrieval in clinical CBCT applications.	Relatively inaccurate detection of symmetric lesions
Johari [30] (2017)	The proposed deep learning model can be used for the diagnosis of vertical root fractures on CBCT images of endodontically treated and also vital teeth. With the aid of the model, the use of CBCT images is more effective than periapical radiographs.	A small training dataset *; Ex-vivo data only containing sound extracted premolars *; Lacking data heterogeneity *; Unknown diagnostic performance on multirooted teeth and teeth with caries or filling materials *.
Kise [32] (2019)	The deep learning model showed high diagnostic accuracy for SjS, which is comparable to that of experienced radiologists. It is suggested that the model could be used to assist the diagnosis of SjS, especially for inexperienced radiologists.	A small training dataset *; Lacking data heterogeneity *; Lacking subjects with other pathological changes of the parotid gland in the control subjects *; Manual ROI segmentation.
Kann [31] (2018)	The proposed deep learning model has the potential for use as a clinical decision-making tool to help guide head and neck cancer patient management.	The process of individual lymph node CT labeling in correlation with pathology reports is subject to some degree of uncertainty and subjectivity *; Only lymph nodes for which a definitive correlation could be made were included in the labeled dataset, potentially biasing the dataset to those nodes that could be definitively correlated with pathologic report *.
Ariji [20] (2019)	The proposed deep learning model yielded diagnostic results comparable to that of radiologists, which suggests that the model may be valuable for diagnostic support.	A small training dataset *; Lacking data heterogeneity *; Lacking independent unseen testing data *; Manual ROI segmentation;
Cheng [33] (2011)	The proposed model can efficiently assist clinicians in locating the odontoid process of the second vertebra.	A small training dataset *; Lacking data heterogeneity *; Inaccurate localization performance.
Shahidi [34] (2014)	The localization performance of the proposed model was acceptable with a mean deviation of 3.40 mm for all automatically identified landmarks.	A small training dataset *; Lacking data heterogeneity *; Inaccurate localization performance.
Montufar [21] (2018)	The proposed algorithm for automatically locating landmarks on CBCT volumes seems to be useful for 3D cephalometric analysis.	A small training dataset *; Lacking data heterogeneity *; Lacking independent unseen testing data *; Inaccurate localization performance.
Montufar [22] (2018)	The proposed hybrid algorithm for automatic landmarking on CBCT volumes seems to be potentially useful for 3D cephalometric analysis.	A small training dataset *; Lacking data heterogeneity *; Lacking independent unseen testing data *; Relatively inaccurate localization performance.
Torosdagli [35] (2019)	The proposed deep learning algorithm allows for orthodontic analysis in patients with craniofacial deformities exhibiting excellent performance.	A small training dataset *; Lacking data heterogeneity *; Analysis of pseudo-3D images instead of fully 3D images *;
Park [36] (2018)	The proposed deep learning algorithm is useful for super-resolution and de-noising.	A small training dataset *; Small anatomical structures may be easily buried and invisible in low-resolution images.
Minnema [23] (2019)	The proposed deep learning algorithm allows us to accurately classify metal artifacts as background noise, and to segment teeth and bony structures.	A small training dataset *; Lacking independent unseen testing data *; Potential bias in the overall accuracy of the gold standard segmentations *.
Miki [38] (2017)	The proposed deep learning algorithm to classify tooth types on CBCTs yielded a high performance. This can be effectively used for automated preparation of dental charts and might be useful in forensic identification.	A small training dataset *; Unstable classification performance due to the analyzed levels of the cross-sectional tooth images and metal artifacts;
Ghazvinian Zanjani [24] (2019)	The proposed end-to-end deep learning framework for the segmentation of individual teeth and the gingiva from intraoral scans outperforms state-of-the-art networks.	A small training dataset *; Ex-vivo data *; Lacking independent unseen testing data *;
Kim [45] (2020)	The proposed automated segmentation method for full arch intraoral scan data is as accurate as a manual segmentation method. This tool could efficiently facilitate the digital setup process in orthodontic treatment.	Ex-vivo data *; Unable to automatically detect the occlusion area.
Lian [25] (2020)	The proposed end-to-end deep neural network to automatically label individual teeth on raw dental surfaces acquired by 3D intraoral scanners outperforms the state-of-the-art methods for 3D shape segmentation.	A small training dataset *; Scans only containing the maxillary dental surfaces with the complete 14 teeth *; Failed to properly handle missing teeth and additional braces in challenging cases; Lacking independent unseen testing data *;
Liu [27] (2016)	The proposed machine learning algorithm based on face scanning patterns could support current clinical practice of the screening and diagnosis of ASD	A small training dataset *; Lacking independent unseen testing data *; Several influencing factors, such as age-/culture-adapted face scanning patterns and the characteristics of the ASD patients should be considered when applying the model to classify children with ASD *.
Knoops [26] (2019)	The proposed model can automatically analyze facial shape features and provide patient-specific treatment plans from a 3D facial scan. This may benefit the clinical decision-making process and improve clinical understanding of face shape as a marker for plastic and reconstructive surgery.	Lacking independent unseen testing data *

3D, three-dimensional; AI, artificial intelligence; ASD, autism spectrum disorder; CBCT, cone beam computed tomography; CT, computed tomography; CNN, convolutional neural network; ROI, region of interest; SjS, Sjögren’s syndrome; * risk of bias.