This study reviews radiographic data segmentation as a cornerstone of machine learning (ML) and deep learning (DL) in dentistry. After outlining artificial intelligence (AI), ML, and DL concepts, it highlights convolutional neural networks-driven tasks—classification, detection, and pixel/voxel segmentation—across panoramic, periapical, bitewing, and cone beam computed tomography imaging. Automated tooth numbering, restoration and implant labeling, caries delineation, endodontic morphology and fractures, periapical and periodontal lesions, and peri-implant bone loss show strong performance metrics, often matching or surpassing clinicians while markedly accelerating workflows. The study underscores AI’s potential to improve accuracy and efficiency while maintaining essential human oversight.
Key points
-
•
Deep-learning radiographic segmentation across panoramic, periapical, bitewing, and CBCT enables classification, detection, pixel-level delineation, automated tooth numbering and restoration/implant labeling, enhancing diagnostic accuracy.
-
•
AI models accurately detect caries, endodontic morphology, root fractures, periapical lesions, and periodontal bone loss, with F1, Dice, sensitivity-specificity often surpassing experienced clinicians.
-
•
In implantology, ML/DL segment alveolar bone, sinuses, mandibular canal; identify implant systems; predict osseointegration, fractures, peri-implant bone loss, improving surgical planning, clinical efficiency, and outcomes.
-
•
While performance is strong across modalities, rare diseases remain underrepresented; larger datasets, standardized validation, and workflow integration are essential for generalization, safety, and clinical adoption.
-
•
Automated segmentation replaces time-consuming manual methods, handling image variability and complex anatomy; models deliver rapid processing, sometimes thousands-fold faster than experts, with accuracies around 80% to 99%.
Abbreviations
| AI | artificial intelligence |
| ANNs | Artificial Neural Networks |
| CBCT | cone beam computed tomography |
| CNNs | Convolutional Neural Networks |
| DL | deep learning |
| DT | Decision Trees |
| ML | machine learning |
| PCA | Principal Component Analysis |
| QLF | quantitative light-induced fluorescence |
| RF | Random Forests |
| SVMs | Support Vector Machines |
| VRF | vertical root fracture |
Artificial intelligence
The term artificial intelligence (AI) was first introduced in 1956 during a workshop led by John McCarthy, Nathaniel Rochester, Marvin Minsky, and Claude Shannon. AI is defined as machines capable of performing human-like tasks—or assisting humans in such tasks—through cognitive abilities such as information processing, communication, learning, and inductive reasoning. AI-based technologies incorporate various cognitive models aimed at analyzing existing data and predicting future trends. For many years, AI remained largely a theoretic concept with limited real-world impact. However, over the past 15 years, it has begun to demonstrate human-level performance across numerous domains ( Fig. 1 ).
The relationship and interplay among artificial intelligence, machine learning, and deep learning.
Machine learning
A subfield of AI, machine learning (ML) refers to the process of developing computer systems that utilize statistical models and algorithms to identify patterns and structures within data. These algorithms are capable of improving over time by learning from new data without human intervention, thereby adapting and refining their performance. Core ML methodologies include supervised learning, unsupervised learning, and reinforcement learning. In supervised learning, models are trained on labeled datasets to produce known outcomes. , Commonly employed ML subalgorithms within supervised learning include Decision Trees (DT), Random Forests (RF), Support Vector Machines (SVMs), and Artificial Neural Networks (ANNs).
DT algorithm resembles a tree structure with branching nodes and leaves. Each node contains a specific threshold value, and the path progresses through branches until a leaf node is reached. Upon reaching a leaf, the instance is assigned to the corresponding class. RF is constructed by aggregating multiple DTs that differ slightly from one another. SVMs operate by mapping data into a hypothetical high-dimensional space, positioning instances based on their features, and identifying the optimal hyperplane that separates them most effectively. ANNs, inspired by the biological neural networks of the human brain, embody mathematically nonlinear models composed of artificial neurons. ,
Unsupervised learning aims to extract new insights from unlabeled data. Common algorithms in this category include Principal Component Analysis (PCA) and k-means clustering. PCA transforms the data into a new set of uncorrelated variables—known as principal components—while retaining as much information as possible and reducing dimensionality. k-means clustering groups similar data points into clusters based on patterns identified within the input data.
Reinforcement learning, on the other hand, leverages reward-based mechanisms to optimize decision-making processes, and holds potential for applications such as treatment planning in dentistry.
Deep learning
Deep learning (DL) is composed of multilayered ANNs, where each layer is responsible for analyzing specific aspects of the input data. As a result, this architecture is capable of generating predictions based on unlabeled and unstructured data as well. DL has the potential to achieve high levels of accuracy, at times even surpassing human performance. The foundation of this success lies in the ability of deep, multilayered Convolutional Neural Networks (CNNs) to learn from large datasets. Unlike traditional approaches, DL models possess an inherent capacity for feature extraction, thereby significantly reducing the need for manual engineering ( Fig. 2 ).
A conventional DL workflow applied to real-world scenarios.
The use of AI in health care was first demonstrated in 1976, when Gunn explored the diagnosis of abdominal pain through computer-based analysis. Today, AI facilitates digital workflows and offers significant improvements in diagnosis, prognostic assessments, image analysis, treatment planning, and patient care. In dentistry, AI-powered solutions promote proactive treatment approaches by providing personalized recommendations to help patients maintain their oral health. Furthermore, virtual dental assistants that offer advice on routine check-ups and cleaning procedures assist individuals in managing their oral hygiene more effectively.
The Fédération Dentaire Internationale recognizes DL as a foundational technology in the field of dentistry, emphasizing its potential to improve global oral health outcomes. One of the key areas where AI excels is the analysis of dental radiographs. In dental radiology, CNNs are frequently employed to perform tasks such as classification, detection, and segmentation ( Fig. 3 A–C).
-
•
Classification: This task encompasses a broad range of applications, such as identifying the presence of disease or distinguishing between different types of conditions. With advances in computational power, increasingly deeper and more complex DL architectures are being employed.
-
•
Detection: Aims to identify and localize pathologic regions or specific anatomic structures within radiographic images. While similar to classification models, these DLs also incorporate additional functionalities such as region proposals and regression.
-
•
Segmentation: Utilized to delineate anatomic structures or lesions in various radiographic images. It involves classifying each pixel or voxel within an image or volumetric dataset into predefined categories.
( A ) Detection. ( B ) Segmentation. ( C ) Classification.
Tooth segmentation and reporting of dental conditions
Automated tooth segmentation holds critical importance across all specialties of dentistry. Replacing manual segmentation—which is often time-consuming and prone to subjective errors—with automated methods can improve accuracy due to their capacity to handle variations in image quality and structural complexity. Advances in this field have the potential to significantly enhance both clinical applications such as orthodontics, prosthodontics, and dental implant planning, as well as forensic dentistry and bone age estimation.
Tooth segmentation and numbering using AI algorithms have primarily been performed on panoramic radiographs. ,,,, A recent review reported that the accuracy of AI-based models for tooth numbering and segmentation ranged between 81.8% and 99%, while sensitivity values ranged from 84.5% to 99.94%. In studies utilizing cone beam computed tomography (CBCT) for dental segmentation, sensitivity values between 93.8% and 99% have been reported. ,, Shaheen and colleagues stated that their DL-based model for automatic tooth segmentation and classification from CBCT images was 1800 times faster than an expert. Alqaftani and colleagues employed CBCT images for the segmentation and classification of teeth with orthodontic brackets, achieving a model accuracy of 99%.
Additionally, several studies have utilized intraoral radiographs such as periapical and bitewing images for tooth segmentation and classification. ,,, Tekin and colleagues used a Mask R-Convolutional Neural Network (CNN) architecture on bitewing radiographs and reported an F1 Score value of 97.36% for tooth segmentation. Ayyıldız and colleagues successfully numbered teeth using a polygonal segmentation method on periapical radiographs via the DL-based YOLOv5 algorithm. Table 1 summarizes studies using AI for segmentation of teeth.
Table 1
Studies using artificial intelligence in the segmentation of teeth
| Study | Objectives | AI Algorithm | Data Modality | Result |
|---|---|---|---|---|
| Lee et al, 2020 | Tooth segmentation | R-CNN | Panoramic Radiography |
F1 score: 0.875
Precision: 0.858 Recall: 0.893 |
| Koch et al, 2019 | Tooth segmentation | U-Net | Panoramic Radiography | Dice Score: 0.936 |
| Leite et al, 2021 | Tooth detection and segmentation | Resnet-101 |
Panoramic
Radiography |
Precision of Detection: 99.6%
Precision of Segmentation: 99.3% |
| Sheng et al, 2023 | Tooth segmentation | SWin-Unet |
Panoramic
Radiography |
F1 Score: 0.6372–0.8203 |
| Adnan et al, 2023 | Tooth segmentation and numbering | U-Net (Segmentation) and Faster R-CNN (Numbering) |
Panoramic
Radiography |
Accuracy of segmentation: 88.2%
Accuracy of numbering: 93.8% |
| Ayidh et al, 2024 | Tooth segmentation and classification | CNN | CBCT |
Accuracy of segmentation: 0.99
Accuracy of classification: 1 |
| Shaheen et al, 2021 | Tooth segmentation and classification | 3D U-Net | CBCT |
Precision and recall of segmentation: 0.98% and 0.83%
Precision and recall of classification: 97.9% and 97.5% |
| Tekin et al, 2017 | Tooth segmentation and numbering | Mask R-CNN | Bitewing Radiography |
F1 Score of segmentation: 97.36%
F1 Score of numbering: 93.33% |
| Ayyıldız et al, 2024 | Tooth numbering with polygonal segmentation | YOLOv5 | Periapical Radiography | Recall: 0.9882, precision: 0.9548, F1 score: 0.9712 |
Various AI algorithms have also been developed for the automatic segmentation of not only teeth but also restorative materials, crowns, root canal fillings, and dental implants. ,,,, The R-CNN model proposed by Vinayahalingam and colleagues achieved F1-scores of 0.993, 0.952, and 0.97, respectively, for the automatic detection, segmentation, and labeling of teeth, crowns, fillings, root canal fillings, implants, and root remnants ( Fig. 4 ). Karataş and colleagues reported satisfactory results with their model designed for the segmentation of amalgam, composite resin, and metal-ceramic restorations in periapical and bitewing radiographs.
The workflow of segmentation process.
Cariology
Dental caries is a chronic, infectious disease resulting from the demineralization of the tooth’s hard tissues. Traditionally, radiographic imaging and visual-tactile examination have been widely used for caries detection. In addition to these methods, more advanced diagnostic tools have been developed, including fiber-optic transillumination, quantitative light-induced fluorescence (QLF), laser fluorescence, electrical conductivity measurements, and optical coherence tomography. However, the relatively low specificity and high setup costs associated with these techniques have limited their widespread adoption in clinical practice.
The advancement of AI algorithms has played a significant role in improving diagnostic accuracy and enhancing the standard of care ( Fig. 5 ).
Mask removal of the carious area from the original image of the tooth by DL segmentation.
Various radiographic modalities, including panoramic, periapical, bitewing, and CBCT images, have been utilized for caries segmentation with AI algorithms. ,,,,,, While studies using non-CBCT radiographic methods contribute approximately 21% to 29% to caries segmentation research, CBCT-based studies account for only around 2%.
Lian and colleagues reported that their nnUNet model for caries segmentation on panoramic radiographs demonstrated diagnostic accuracy comparable to that of dental professionals. Liu and colleagues, employing a stepwise learning approach for periapical radiographs, achieved an accuracy of 82.5% in caries segmentation. Similarly, Cantu and colleagues’ DL-based model attained an accuracy of 0.80 using bitewing radiographs, outperforming dental practitioners in segmentation tasks.
In CBCT imaging, successful results have been achieved using a model based on ML and image processing techniques to segment and classify carious lesions.
Endodontics
Endodontics is a branch of dentistry focused on the diagnosis, prevention, and treatment of diseases affecting the dental pulp and periapical tissues. Traditionally, it relies heavily on clinical experience and radiographic interpretation, which are inherently subject to human error and subjective judgment. The advancement of AI offers the potential to overcome these limitations and transform endodontic practice. In particular, CNN-based models within the field of DL have significantly enhanced diagnostic accuracy and treatment planning.
CNNs have demonstrated high performance in tasks such as identifying root canal morphology, determining working length, detecting vertical root fractures (VRFs), and recognizing pathologic signs in radiographic images. Accurate determination of working length is critical for successful root canal treatment. In one study where CBCT was considered the gold standard, a DL model successfully segmented the root morphology of mandibular first molars using panoramic radiographs.
Variations in root canal morphology are also major factors influencing endodontic success. Jeon and colleagues utilized a CNN-based model to segment C-shaped canals of mandibular second molars on panoramic images and reported that the model showed notable accuracy in predicting these anatomic variations.
In another study, Duman and colleagues evaluated the diagnostic performance of a DL-based YOLOv5 model for the automatic segmentation of the mesiobuccal second root—a variation commonly found in maxillary molars and known to greatly impact treatment success. The model achieved a sensitivity of 92%, demonstrating promising diagnostic capability.
Vertical root fractures and periapical lesions
VRFs are diagnostically challenging conditions that may necessitate root resection or tooth extraction, accounting for approximately 2% to 5% of crown/root fractures. Fukuda and colleagues reported that CNN-based models can serve as useful tools for detecting VRFs in panoramic radiographs. In a study where 2 AI models were developed to predict VRFs in endodontically treated and intact teeth using both periapical and CBCT images, the model based on CBCT images demonstrated higher diagnostic accuracy compared to the one using periapical images.
Accurate diagnosis and treatment planning for teeth presenting with periapical lesions and associated symptoms can be difficult. Apical periodontitis is a common condition, representing approximately 75% of radiolucent jaw lesions. AI models have the potential to effectively analyze characteristic features of radiolucency and alveolar bone resorption. Li and colleagues demonstrated that their CNN model facilitated automatic diagnosis of apical lesions on periapical radiographs with an accuracy of 92.5%.
In another study focused on the segmentation of periapical lesions in panoramic radiographs, a DL-based model was found to reliably detect these lesions ( Fig. 6 ). Zheng and colleagues successfully segmented apical lesions on CBCT images using DL algorithms that integrated anatomic region information.
The semantic segmentation task for apical lesion.
Periodontal diseases
Periodontitis is a common chronic inflammatory disease that affects the gums and the surrounding supporting tissues. , Early diagnosis of this condition can improve treatment outcomes and prevent complications. Detection of periodontal diseases typically relies on clinical examination, periodontal probing, radiographic evaluation, QLF, and oral bacterial DNA testing. However, these methods often require specialized expertise and may be expensive or have limited accessibility.
AI holds significant potential in periodontology by enhancing the accuracy of risk assessments, supporting radiographic evaluations, and enabling early diagnosis of periodontitis through the identification of periodontal changes—ultimately contributing to personalized treatment planning for patients. An AI system developed by Chifor and colleagues demonstrated the ability to automatically segment normal periodontal tissues using ultrasound imaging. In another study, Nguyen and colleagues aimed to automatically segment alveolar bone and detect the alveolar crest from intraoral ultrasound images using an ML approach, achieving a sensitivity of 88.5%. Such systems can assist dental professionals and periodontists in diagnosis and treatment planning.
Numerous studies have applied AI models to segment alveolar bone loss and diagnose periodontal disease using various radiographic modalities. ,,,, A recent review reported that the diagnostic accuracy of AI models in detecting alveolar bone loss and periodontal disease using panoramic, periapical, and bitewing radiographs ranged from 73.4% to 99%. Kurt Bayrakdar and colleagues developed models using CBCT images that successfully segmented teeth, total alveolar bone loss, suprabony defects, infrabony defects, perio-endo lesions, buccal defects, and furcation defects ( Fig. 7 ).
( A , B ) The automatic segmentation of the periodontal problems utilizing the DL model in axial cone-beam CT slices.
Lee and colleagues trained and tested an AI model using intraoral radiographs to identify periodontally compromised teeth and predict tooth extraction. Their model demonstrated diagnostic accuracy rates of 81% to 76.7% for identifying compromised teeth and 82.8% to 73.4% for predicting extractions. Similarly, Thanathornwong and colleagues reported satisfactory performance from a DL model trained with input images and corresponding segmentation maps to detect periodontally endangered teeth on panoramic radiographs.
Implantology
Dental implants have become one of the routine treatment modalities for oral rehabilitation. With the integration of AI technologies, implantology has undergone a significant paradigm shift. AI algorithms enhance the precision of implant placement, contributing to safer treatment processes while offering the potential to optimize esthetic outcomes.
Prior to implant placement, radiographic techniques are commonly used to assess bone structure and dimensions. The CNN model developed by Kurt Bayrakdar and colleagues demonstrated substantial potential in analyzing the morphology of edentulous alveolar bone and segmenting surrounding anatomic structures such as the nasal fossa, maxillary sinus, and mandibular canal using CBCT images. Similarly, Moufti and colleagues reported that ML achieved higher accuracy than manual segmentation in delineating edentulous alveolar bone on CBCT scans prior to implant placement.
The wide variety of dental implant systems presents a significant challenge for clinicians in both diagnosis and treatment planning. To address this, various AI models have been developed to identify implant types using periapical and panoramic radiographs. ,,,, Revilla-León and colleagues reported that AI models designed to determine implant types from radiographs achieved accuracy rates ranging between 93.8% and 98%. Similarly, Chaurasia and colleagues demonstrated that DL models used to identify and classify different implant systems from periapical and panoramic radiographs achieved accuracies ranging from 70.75% to 98.19%.
Accurately predicting therapeutic outcomes in implantology is critical for ensuring treatment success. AI is expected to play an increasingly significant role in this domain, as AI-based analyses allow for the evaluation of patient-specific data to forecast treatment success and minimize potential risks. AI models developed to predict osseointegration or implant success using various input data have demonstrated accuracy levels ranging from 62.4% to 80.5%. Additionally, CNN models have shown strong performance in predicting implant fractures and segmenting peri-implant bone loss.
Oral and maxillofacial pathology
Oral and maxillofacial pathology encompasses a wide range of diseases and conditions affecting the oral cavity, jaws, and facial structures. However, the complexity and variability of these pathologies often make accurate diagnosis challenging. Additionally, interobserver and intraobserver variability among pathologists can lead to inconsistencies in the diagnostic process. In this context, AI-based analyses have significant potential to standardize the diagnostic workflow and enhance accuracy.
Numerous studies have demonstrated the ability of AI systems to segment various maxillofacial pathologies—including ameloblastomas, dentigerous cysts, odontogenic keratocysts, radicular cysts, follicular cysts, and a variety of tumors—with high accuracy on both 2 dimensional and 3 dimensional imaging data ,,, ( Fig. 8 ).
( A ) Object detection: The model detects cystic lesions and draws a box around them. ( B ) Segmentation: Different models segment the upper jaw, lower jaw, sinuses, teeth, and mandibular canal. ( C ) Overlap analysis: Overlaps between lesion boxes and segmented structures are calculated.
The model developed by Endres and colleagues outperformed 14 oral and maxillofacial surgeons in segmenting infections, granulomas, cysts, and tumors in panoramic radiographs.
In a review of ML-based studies focused on the identification of maxillofacial cysts and tumors, the average sensitivity and specificity of studies using panoramic radiographs were reported as 0.83 and 0.82, respectively. In contrast, studies utilizing CBCT images achieved superior performance, with reported sensitivity and specificity values of 0.88 each.
The maxillary sinuses play a critical role in many dental procedures due to their anatomic proximity to the posterior maxillary teeth. Ohashi and colleagues reported that diagnosing maxillary sinus diseases can be particularly challenging for inexperienced dentists or dental assistants. To assist in the automatic identification of maxillary sinus pathologies on panoramic and CBCT images, several AI-based models have been developed. ,,,,
In one study, a CNN model demonstrated diagnostic performance for maxillary sinusitis on panoramic radiographs that was comparable to 2 radiologists and superior to that of 2 dental assistants. The diagnostic accuracy of DL algorithms in predicting maxillary sinus diseases has been reported to range between 75.7% and 99.7%.
Treatment planning
Complex procedures involved in orthodontic and maxillofacial treatments require a comprehensive understanding of craniofacial anatomy. With the application of AI-based systems, anatomic and dental relationships can be analyzed to assess growth patterns and developmental components.
For orthodontic treatments to be effective in the long term, accurate evaluation of growth and development is essential. AI systems have demonstrated promising results in predicting growth rate and development, as well as understanding how these factors impact orthodontic outcomes. These applications include the analysis of cervical vertebral maturation, segmentation of the pharyngeal airway, and the creation of 3 dimensional models prior to orthodontic interventions ,, ( Fig. 9 ).
