Table 2.

Advanced AI applications in oral radiology.

Author	Summarized Abstract	Methods Used	Results	Conclusions
Zhu J. [27]	-AI framework for dental disease diagnosis on panoramic radiographs.	-Developed AI framework based on BDU-Net and nnU-Net. -Dentists with different levels of seniority independently diagnosed the evaluation dataset.	-AI framework showed high specificity in diagnosing dental diseases. -Performance comparable to dentists with 3–10 years of experience.	-Diagnostic time of the AI framework was significantly shorter.
Mima Y. [26]	-Tooth detection method using Faster R-CNN for dental X-ray images.	-Applied Faster R-CNN to extract a rectangular area including all teeth from each X-ray image. -Classified bounding boxes for each tooth into one of 32 tooth types using the trained Faster R-CNNs.	-Detection rate per tooth: 98.9%. -Mean intersection over union for each tooth: 0.748.	-Improved tooth detection with Faster R-CNNs in divided areas.
Başaran [25]	-Developed AI model for diagnostic charting in panoramic radiography.	-Developed an AI model (CranioCatch, Eskişehir, Turkey) using a deep CNN method. Employed the Faster R-CNN Inception v2 architecture from the TensorFlow library for model implementation.	-Highest precision for prosthesis, implant supported prosthesis, lowest for caries, dental calculus.	-AI model detected dental conditions in panoramic radiographs, aiding diagnosis and treatment. -Highest sensitivity for prosthesis, implant, impacted tooth, lowest for caries, dental calculus.
Lee [24]	-Deep learning neural network for cystic lesion diagnosis in imaging.	-Periapical radiographic images split into training/validation. -Utilized a pre-trained GoogLeNet Inception v3 CNN for preprocessing and transfer learning.	-Deep CNN achieved 84.6% accuracy with panoramic images, 91.4% with CBCT.	-Deep CNN architecture enhanced diagnosis of cystic lesions. -Further studies need to include ameloblastoma in the dataset.
Minnema [23]	-MS-D network developed for bone segmentation in CBCT scans.	-Bone segmentation performance was evaluated using a leave-2-out cross-validation method. -MS-D network’s performance was compared against a clinical snake evolution algorithm and two CNN architectures.	-MS-D network accurately segmented bony structures. ResNet introduced wave-like artifacts, U-Net mislabeled background voxels.	-The MS-D network can be utilized for segmentation of bony structures in CBCT scans.
Kuwana [22]	-AI system detects periapical pathosis on CBCT images accurately.	-Learning process conducted over 1000 epochs using DetectNet with training images and labels to create a learning model.	-Reliability of correctly detecting a periapical lesion was 92.8%. -Volume measurements by AI and humans were comparable.	-AI systems accurately detected periapical lesions with high reliability. -Deep learning AI useful for detecting periapical pathosis on CBCT images.
Mackie [20]	-Study focuses on TMJ OA diagnosis using bone imaging biomarkers.	-Investigated the articular fossa radiomic biomarkers and condyle-to-fossa distance, noting differences in the condyle-to-fossa distance between control and TMJ OA patients utilizing a LightGBM (light gradient boosting machine) model.	-No statistically significant difference in articular fossa radiomic biomarkers between TMJ OA and control patients.	-Articular fossa imaging features may have a larger contribution in diagnosis.
Tajima [18]	-AI system detects cyst-like lesions in jaws on panoramic radiographs.	-Deep learning algorithm with transfer learning used for training data. -Development of a deep convolutional neural network (DCNN) for automatic detection.	-Results included detection of cyst-like radiolucent lesions on panoramic radiographs. -Identified lesions: radicular cysts, dentigerous cysts, odontogenic keratocysts, simple bone cysts, and ameloblastomas.	-AI system detected cyst-like lesions with high accuracy. -AI may contribute to diagnostic support in future clinical practice.
Fukuda [17]	-Compares three CNNs for mandibular third molar and canal relationship. -Evaluated time, storage, diagnostic performance, and consistency of CNNs.	-A DCNN was constructed using transfer learning techniques to automatically detect the lesions.	-Good or very good consistency values for all CNNs. -No significant differences in diagnostic performance among CNNs with smaller patches.	-Time and storage requirements depended on CNN depth and parameters. -Image patch size crucial for high diagnostic performance and consistency.
Ariji [16]	-Deep learning detects mandibular radiolucent lesions with high sensitivity.	-Utilized 210 training images with corresponding labels for model training on a deep learning GPU training system (DIGITS) using the DetectNet neural network.	-Sensitivity was 0.88 for both testing 1 and 2. -False-positive rate per image was 0.00 for testing 1. -False-positive rate per image was 0.04 for testing 2.	-Deep learning can achieve high sensitivity in radiolucent lesion detection in the mandible. -Dentigerous cysts showed best detection and classification sensitivity.
Shaheen [15]	-Develops AI system for tooth segmentation and classification on CBCT.	-Developed an artificial intelligence framework using a three-step segmentation approach, each step employing a 3D U-Net architecture.	-AI model outperformed ground truth with 0.56 ± 0.38 mm Hausdorff distance -AI was 1800 times faster than an expert in teeth segmentation	-3D U-Net AI system was accurate and time-efficient for tooth segmentation. -AI system reduced clinical workload in dental diagnostics and treatment planning.
Cantu [14]	-Compares the performance of the neural network of the caries lesions of different radiographic extension on bitewings test dataset against seven independent evaluators.	-Utilized a convolutional neural network (U-Net) for analysis. -Stratified analysis based on lesion depth, categorizing into enamel lesions and dentin lesions.	-Neural network accuracy: 0.80, dentists’ mean accuracy: 0.71. -Neural network sensitivity: 0.75, dentists’ sensitivity: 0.36. -Dentists’ specificity: 0.91, neural network specificity: 0.83.	-Dentists under-detected lesions, while the network slightly over-detected.
Lee [13]	-Deep CNNs for dental caries detection and diagnosis on radiographs.	-Utilized a pre-trained GoogLeNet Inception v3 CNN for preprocessing and transfer learning	-Diagnostic accuracies for premolar, molar, and both models were provided. -Premolar model had the best area under the ROC curve.	-Deep CNN algorithm showed good performance in detecting dental caries.
Kuwada [12]	-Deep learning systems classify impacted supernumerary teeth in maxillary incisor region.	-Three different learning models were developed using AlexNet, VGG-16, and DetectNet.	-VGG-16 showed significantly lower values compared to DetectNet and AlexNet.	-DetectNet and AlexNet had potential for classifying impacted supernumerary teeth.
Yılmaz [11]	-Decision support system for classifying dental lesions using CBCT imaging.	-Utilized 50 CBCT images identified as periapical cysts and keratocystic odontogenic tumors, based on clinical, radiographic, and histopathologic features. Custom-developed software was utilized for segmentation.	-SVM classifier achieved 100% accuracy and F1 scores. -SVM showed 96.00% accuracy and 96.00% F1 scores.	-Periapical cyst and KCOT lesions can be classified with high accuracy. -The study contributed to computer-aided diagnosis of dental apical lesions.