Table 5.

CNN selected studies.

Authors Name and Year	Methods	Results	Authors Suggestions/Conclusions
Prajapati et al., (2017) [16]	Transfer learning with VGG16 pre-trained model	Accuracy = 88.46%	Transfer learning with the VGG16 pre-trained model achieved better accuracy.
Lee et al., (2018) [56]	Pre-trained GoogLeNet Inception v3 network	Accuracy of 89%, 88%, and 82% was observed in the premolar, molar, and both the premolar-molar regions.	In terms of diagnosing dental caries, Deep CNN algorithms are anticipated to be among the best and most productive technique.
Vinayahalingam et al., (2021) [57]	CNN MobileNet V2	Accuracy = 0.87, sensitivity = 0.86, specificity = 0.88, AUC = 0.90	This method forms a promising foundation for the further development of automatic third molar removal assessment.
Choi et al., (2018) [63]	Customized CNN	F1max = 0.74, FPs = 0.88	This system can be used to detect proximal dental caries on several periapical images.
Lee et al., (2021) [65]	Deep CNN (U-Net)	Precision = 63.29%, recall = 65.02%, F1-score = 64.14%	Clinicians should not wholly rely on AI-based dental caries detection results, but should instead use them only for reference.
Yang et al., (2018) [67]	Customized CNN	F1 score = 0.749	The method doesn’t always work on images of molars.
Lee et al., (2018) [68]	Pre-trained deep CNN (VGG-19) and self-trained network	Premolars (accuracy = 82.8%), molars (accuracy = 73.4%)	Using a low-resolution dataset can reduced the accuracy of the diagnosis and prediction of PCT.
Al Kheraif et al., (2019) [69]	Hybrid graph-cut technique and CNN	Accuracy = 97.07%	The DL with convolution neural network system effectively recognizes the dental disease.
Murata et al., (2019) [70]	Customized AlexNet CNN	Accuracy = 87.5%, sensitivity = 86.7%, specificity = 88.3%, AUC = 0.875	The AI model can be a supporting tool for inexperienced dentists.
Krois et al., (2019)  [72]	Custom-made CNN	Accuracy = 0.81, sensitivity = 0.81, Specificity = 0.81	ML-based models could minimize the efforts.
Zhao et al., (2020) [77]	Customized Two-staged attention segmentation network	Accuracy = 96.94%, dice = 92.72%, recall = 93.77%	Failure to properly divide the foreground image into teeth areas due to inaccurate pixel segmentation.
Fariza et al., (2020) [78]	U-Net convolution network	Accuracy = 97.61%	Segmentation with the proposed U-Net convolution network results in fast segmentation and smooth image edges.
Lakshmi and Chitra, (2020) [79]	Sobel edge detection with deep CNN	Accuracy = 96.08%	Sobel edge detection with deep CNN is efficient for cavities prediction compared to other methods.
Khan et al., (2021) [80]	U-Net + Densenet121	mIoU = 0.501, Dice coefficient = 0.569	DL can be a viable option for segmentation of caries, ABR, and IRR in dental radiographs.
Moran et al., (2020) [81]	Pre-trained ResNet and an Inception model	Accuracy = 0.817, precision = 0.762, recall = 0.923, specificity = 0.711, negative predictive = 0.902	Clinically, the examined CNN model can aid in the diagnosis of periodontal bone deterioration during periapical examinations.
Chen et al., (2021)  [82]	Customized Faster R-CNN	Precision = 0.5, recall = 0.6	Disease lesions with too small sizes may not be indications for faster R-CNN.
Lin and Chang, (2021)  [84]	ResNet	Accuracy = 93.33%	In the second stage, endodontic therapy is the most vulnerable to incorrect labeling.
Zhang et al., (2022)  [85]	Customized multi-task CNN	Precision = 0.951, recall = 0.955, F-score = 0.953	The method can provide reliable and comprehensive diagnostic support for dentists.
Yu et al., (2020)  [91]	Customized ResNet50-FPN	Accuracy = 95.25%, sensitivity = 89.83%, specificity = 96.10%	Only implement caries detection for First Permanent Molar not all teeth.
Rana et al., (2017)  [92]	Customized CNN	AUC = 0.746, precision = 0.347, recall = 0.621	Dental professionals and patients can benefit from automated point-of-care early diagnosis of periodontal diseases provided.
Tanriver et al., (2021)  [94]	Multiple pre-trained NNs; EfcientNet-b4 architecture	sensitivity = 89.3, precision = 86.2, F1 = 85.7	The suggested model shows significant promise as a low-cost, noninvasive tool to aid in screening procedures and enhance OPMD identification.
Schlickenrieder et al., (2021) [95]	pre-trained ResNeXt-101–32x8d	accuracy = 98.7%, AUC = 0.996	More training is needed in AI-based detection, classification of common and uncommon dental disorders, and all types of restorations.
Takahashi et al., (2021)  [96]	YOLO v3 and SSD	mAP = 0.80, mIoU = 0.76	This method was limited accuracy in identifying tooth-colored prosthese.