Table 4.

Performance of the developed AI models in comparison to specialists/general practitioners

Author (Year)	Application	Imaging modality	AI software/ deep learning model	Test dataset	Performance of the developed software/model versus human		Main findings
					AI	Human Mean (range)
Dental caries
Srivastava et al. (2017) 19	Detection of dental caries	Bitewing radiography	CNN	500 images from nearly 100 clinics across USA	SEN = 0.81 PPV = 0.62 F1 = 0.7	three dentists SEN = 0.42 (0.34–0.48) PPV = 0.78 (0.63–0.89) F1 = 0.53 (0.5–0.56)	The model achieved significantly higher F1-score and sensitivity for detecting caries than three dentists.
Cantu et al. (2020) 20	Detection of initial (enamel) and advanced (dentin) proximal caries	Bitewing radiography	CNN (U-Net)	141 bitewings from the dental clinic at Charité - Universitätsmedizin Berlin	All caries ACC = 0.80 SEN = 0.75 SPE = 0.83 PPV = 0.70 NPV = 0.86 F1 = 0.73 MCC = 0.57 Initial caries SEN > 0.7 Advanced caries SEN > 0.7	seven experienced dentists All caries ACC = 0.71 SEN = 0.36 (0.19–0.65) SPE = 0.91 (0.69–0.98) PPV = 0.75 (0.41–0.88) NPV = 0.72 (0.68–0.82) F1 = 0.41 (0.26–0.63) MCC = 0.35 (0.14–0.51) Initial caries SEN <0.25 Advanced caries SEN = 0.40–0.75	The model achieved higher overall accuracy than seven dentists. The seven dentists were far less sensitive, but slightly more specific than the model. For initial caries, the risk of under detection by dentists was very high while the model showed robust sensitivity regardless of the lesion depth.
Mertens et al. (2021) 99	Detection of proximal enamel, early dentin, and advanced dentin caries	Bitewing radiography	dentalXr.ai Pro software	20 bitewings from the dental clinic at Charité - Universitätsmedizin Berlin	ten images evaluated by 22 dentists with the aid of dentalXr.ai Pro AUC = 0.89 ACC = 0.94 SEN = 0.81 SPE = 0.97 PPV = 0.82 NPV = 0.97 F1 = 0.81	ten images evaluated by 22 dentists without the aid of dentalXr.ai Pro AUC = 0.85 ACC = 0.93 SEN = 0.72 SPE = 0.97 PPV = 0.80 NPV = 0.95 F1 = 0.76	dentalXr.ai Pro software can significantly increase dentists’ sensitivity for detecting enamel caries.
Devlin et al. (2021) 21	Detection of proximal enamel caries	Bitewing radiography	AssistDent software	24 images from one dental hospital and nine general dental practice sites in UK	24 images evaluated by 12 dentists with the aid of AssistDent SEN = 0.76 SPE = 0.85	24 images evaluated by 11 dentists without the aid of AssistDent SEN = 0.44 SPE = 0.96	AssistDent software can significantly increase dentists’ sensitivity for detecting proximal enamel caries in enamel
Endodontic evaluation
Hamdan et al. (2022) 101	Detection of apical radiolucencies	Periapical radiography	Denti.AI software	68 images from one dental center	six operative dentistry residents, one general dentist and one endodontist with the aid of Denti.AI AUC = 0.89 SEN = 0.93 SPE = 0.73	six operative dentistry residents, one general dentist and one endodontist without the aid of Denti.AI AUC = 0.82 SEN = 0.94 SPE = 0.60	Denti.AI software can enhance dental practitioner’s ability to detect apical radiolucencies on periapical images.
Jeon et al. (2021) 34	Detection of C-shaped canals in mandibular second molars	Panoramic radiography	CNN (Xception)	408 cropped images of mandibular second molars	AUC = 0.98 ACC = 0.95 SEN = 0.93 SPE = 0.97 PPV = 0.96	OMF radiologist/endodontist AUC = 0.87/0.89 ACC = 0.87/0.89 SEN = 0.93/0.92 SPE = 0.82/0.86 PPV = 0.84/0.86	The model outperformed the OMF radiologist and endodontist
Sherwood et al. (2021) 35	Segmentation and classification of C-Shaped canals in mandibular second molars	CBCT	CNN (U-Net, Residual U-Net, or XceptionU-Net)	35 scans	SEN = 0.72–0.79	one endodontist and 1 OMF radiologist SEN = 0.97	The model performed less well than the OMF radiologist and endodontist while it may aid clinicians with the detection and classification of C-shaped canal anatomy.
Yang et al. (2022) 36	Classification of C-shaped canals in mandibular second molars	Periapical and panoramic radiography	CNN	100 cropped images consisting of 56 mandibular second molars without C-shaped canals and 44 molars with C-shaped canals	Periapical images (PA) AUC = 0.95 ACC = 0.90 SEN = 0.93 SPE = 0.87 PPV = 0.90 NPV = 0.91 F1 = 0.91 Panoramic images (Pano) AUC = 0.93 ACC = 0.85 SEN = 0.72 SPE = 0.93 PPV = 0.87 NPV = 0.84 F1 = 0.79	one specialist AUC = 0.95 (PA); 0.96 (Pano) ACC = 0.95 (PA); 0.96 (Pano) SEN = 0.95 (PA); 0.97 (Pano) SPE = 0.94 (PA); 0.95 (Pano) PPV = 0.94 (PA); 0.95 (Pano) NPV = 0.95 (PA); 0.97 (Pano) F1 = 0.94 (PA); 0.96 (Pano) one general dentist AUC = 0.89 (PA); 0.91 (Pano) ACC = 0.89 (PA); 0.91 (Pano) SEN = 0.91 (PA); 0.93 (Pano) SPE = 0.87 (PA); 0.89 (Pano) PPV = 0.86 (PA); 0.89 (Pano) NPV = 0.92 (PA); 0.93 (Pano) F1 = 0.89 (PA); 0.91 (Pano)	The model’s diagnostic performance using only the root portion of the tooth was similar to the specialist and superior to the general dentist. Both the specialist and general dentist showed better diagnostic performance when reading panoramic radiographs compared with periapical images.
Periodontal evaluation
Kim et al. (2019) 22	Detection of periodontal bone loss	Panoramic radiography	Deep neural transfer network	800 images from Korea University of Anam Hospital	AUC = 0.95 SEN = 0.77 SPE = 0.95 PPV = 0.73 NPV = 0.96 F1 = 0.75	five dentists AUC = 0.85 (0.84–0.87) SEN = 0.78 (0.74–0.80) SPE = 0.92 (0.91–0.93) PPV = 0.62 (0.59–0.65) NPV = 0.96 (0.95–0.97) F1 = 0.69 (0.68–0.70)	The model outperformed five dentists in detecting periodontal bone loss.
Krois et al. (2019) 23	Detection of periodontal bone loss	Panoramic radiography	CNN	353 cropped images of individual tooth	AUC = 0.89 ACC = 0.81 SEN = 0.81 SPE = 0.81 PPV = 0.76 NPV = 0.85 F1 = 0.78	six dentists AUC = 0.77 ACC = 0.76 SEN = 0.92 SPE = 0.63 PPV = 0.68 NPV = 0.90 F1 = 0.78	The model outperformed six dentists in detecting periodontal bone loss.
Dental implants
Liu et al. (2022) 37	Detection of peri-implant bone loss	Periapical radiography	Faster R-CNN	150 images of bone level dental implants placed in patients	SEN = 0.67 SPE = 0.87 PPV = 0.81	two dentists SEN = 0.62–0.93 SPE = 0.64–0.77 PPV = 0.69–0.70	The model performed similarly to two dentists, but inferior to one experienced dentist (ground truth)
Lee et al. (2020) 41	Classification of six dental implant systems	Periapical and panoramic radiography	18-layer deep CNN	2,396 cropped images of individual dental implant placed in patients from three centers including Daejeon Dental Hospital, Wonkwang University; Ilsan Hospital, National Health Insurance Service; and Mokdong Hospital, Ewha Womans University	AUC = 0.90–0.98 SEN = 0.83–0.97 SPE = 0.83–0.98	six board-certified periodontists AUC = 0.50–0.97 SEN = 0.78–0.97 SPE = 0.39–0.99 eight periodontology residents AUC = 0.50–0.92 SEN = 0.10–0.95 SPE = 0.38–0.99 eleven residents not specialized in periodontology AUC = 0.54–0.92 SEN = 0.49–0.89 SPE = 0.39–0.96	The model outperformed most of the participating periodontists, periodontal residents, and residents not specialized in periodontology.
Cystic, nodal, and tumor lesions
Poedjiastoeti et al. (2018) 50	Detection of ameloblastomas and keratocysts	Panoramic radiography	CNN (VGG-16)	100 images from 50 patients with ameloblastomas and 50 patients with keratocysts	ACC = 0.83 SEN = 0.82 SPE = 0.83 Diagnostic time: 38 s	5 OMF surgeons ACC = 0.83 SEN = 0.81 SPE = 0.83 Diagnostic time: 23 mins	The model’s performance was on par with five OMF surgeons.
Endres et al. (2020) 49	Detection and segmentation of infection, granuloma, cysts, and tumors in the jaws	Panoramic radiography	CNN (U-Net)	102 images from the Department of Oral and Maxillofacial Surgery, Charité, Berlin	SEN = 0.51 PPV = 0.67	24OMF surgeons SEN = 0.51 (0.26–0.76) PPV = 0.69 (0.42–0.93)	The model outperformed 14 of 24 OMF surgeons
Ariji et al. (2022) 77	Identification of metastatic cervical lymph nodes	Contrast-enhanced CT	CNN (U-Net)	72 image slices of 24 metastatic and 68 non-metastatic lymph nodes from 59 OSCC patients	AUC = 0.95 ACC = 0.96 SEN = 0.98 SPE = 0.95	two radiologists AUC = 0.90 ACC = 0.89 SEN = 0.94 SPE = 0.86	The model outperformed two radiologists in identifying metastasis with a short time period of 7 sec.
Others
Kunz et al. (2020) 60	Localization of cephalometric landmarks	Cephalometric radiography	CNN	50 images from a private orthodontic dental practice	Mean absolute differences between AI and gold standard ranging 0.46–2.18° for angular and 0.44–0.64 mm for linear analyses	Mean absolute differences between 12 orthodontists and gold standard ranging 0.55–1.80° for angular and 0.35–0.88 mm for linear analyses	The model’s performance reached the level equivalent to that of experienced orthodontists.
Ezhov et al. (2021) 78	Segmentation of teeth and jaws, numbering of teeth, detection of caries, periapical lesions, and periodontitis	CBCT	Diagnocat software	30 scans selected from 1,135 scans acquired from 17 scanners	Cross-condition SEN = 0.92 SPE = 0.99	4 OMF radiologists Cross-condition SEN = 0.93–0.94 SPE = 0.99–1.00	Diagnocat‘s performance was on par with four radiologists
Choi et al. (2022) 46	Determination and classification of positional relationships between lower third molars and the mandibular canal	Panoramic radiography	CNN (ResNet-50)	Cropped images of lower third molars with their roots overlapping the mandibular canal from 25% of 571 panoramic images	Determination of the true contact position ACC = 0.72 SEN = 0.86 SPE = 0.55 Classification of the bucco-lingual position ACC = 0.81 SEN = 0.87 SPE = 0.75	6 OMF surgeons Determination of the true contact position ACC = 0.53–0.70 SEN = 0.25–0.88 SPE = 0.17–0.92 Classification of the bucco-lingual position ACC = 0.32–0.52 SEN = 0.40–1.0 SPE = 0–0.56	The model outperformed six OMFS specialists with much higher accuracy for determining the true contact position and classifying the bucco-lingual position between lower third molars and the mandibular canal.
Vollmer et al. (2022) 58	Prediction of oroantral communication after tooth extraction	Panoramic radiography	CNN (VGG16, InceptionV3, MobileNetV2, EfficientNet, or ResNet50)	60 images from patients with or without postoperative OAC	The highest performance (MobileNetV2) AUC = 0.67 ACC = 0.74 SEN = 0.43 PPV = 0.75 F1 = 0.55	4 OMF experts AUC = 0.55–0.71 SEN = 0.14–0.60	Although the MobileNetV2 model and one expert reached AUCs of nearly 0.7, the overall accuracy for predicting oroantral communication after tooth extraction from panoramic images was not sufficiently reliable.
Murata et al. (2019) 55	Diagnosis of maxillary sinusitis	Panoramic radiography	CNN (AlexNet)	120 images consisting of 60 healthy and 60 inflamed sinuses	ACC = 0.88 SEN = 0.87 SPE = 0.88 PPV = 0.88 NPV = 0.87	Radiologists/dental residents ACC = 0.90/0.77 SEN = 0.90/0.78 SPE = 0.89/0.75 PPV = 0.89/0.76 NPV = 0.90/0.78	The model performed similarly to two OMF radiologists and outperformed two dental residents.
Jung et al. (2021) 69	Diagnosis of temporomandibular joint osteoarthritis	Panoramic radiography	CNNs (ResNet-152 or EfficientNet-B7)	20% of 858 images from 395 patients with normal TMJs and 463 with TMJ osteoarthritis	ResNet/EfficientNet AUC = 0.94/0.95 ACC = 0.88/0.88 SEN = 0.95/0.86 SPE = 0.80/0.91	Specialists/general dentists ACC = 0.88/0.67 SEN = 0.86/0.69 SPE = 0.91/0.65	The model outperformed three general dentists and three specialists in the diagnosis of TMJ osteoarthritis
Kise et al. (2019) 76	Diagnosis of Sjögren’s syndrome	CT	CNN (AlexNet)	100 CT slices from 5 patients diagnosed with Sjögren’s syndrome and five individuals without any parotid gland abnormalities	ACC = 0.96 SEN = 1.0 SPE = 0.92	Experienced/inexperienced radiologists ACC = 0.98/0.84 SEN = 0.99/0.78 SPE = 0.97/0.89	The model performed similarly to three experienced OMF radiologists and outperformed three inexperienced OMF radiologists.

ACC, accuracy; AI, artificial intelligence; AUC, area under the ROC curve; CBCT, cone-beam computed tomography; CT, computed tomography; CNN, convolutional neural network; DSC, Dice similarity coefficient; F1, F1-score; MCC, Matthew’s correlation coefficient; NPV, negative predictive value; OAC, Oroantral communication; OMF, oral and maxillofacial; PA, periapical images; Pano, panoramic images; PPV, positive predictive value (Precision); SEN, sensitivity (Recall); SPE, specificity; TMJ, temporomandibular joint;