The specialty of endodontics primarily addresses diseases of the pulp and periradicular tissues. The clinical aspects encompass a broad treatment spectrum, including primary root canal treatment, nonsurgical retreatment, endodontic surgery, dental trauma, vital pulp therapy, and regenerative endodontic procedures. The larger scope of the field encompasses research in pulp biology and the surrounding tissues, as well as tissue engineering for pulp and tooth regeneration.

Artificial intelligence (AI) has found numerous applications in the medical field. However, in dentistry, its use is still severely limited. Endodontic procedures are carried out by specialists and general practitioners. Endodontics can be complicated for clinicians due to varying symptoms, changing clinical presentations, and challenges with radiographic image interpretation. AI-supported applications may have the potential to aid generalists and specialists with diagnosis, decision-making, and treatment planning for endodontic disease. AI, in various forms, is likely to be implemented in endodontic education. Furthermore, AI applications such as chatbots may be employed for everyday practice functions such as scheduling, note-taking, or dealing with insurance companies, and educating patients on endodontic topics.

The application of AI in the field of endodontics aims at improving diagnosis, clinical decision-making, and treatment planning. Utilizing complex mathematical operations, AI allows for the objective evaluation of patient data that augments clinical expertise, reduces diagnostic variability, and may improve procedural outcomes. AI was originally defined as “the science and engineering of making intelligent machines.”

Today, powerful computational systems emulate tasks traditionally requiring human intelligence. Machine learning (ML), a subset of AI, allows algorithms to learn from data and improve predictions or classifications over time. Deep learning (DL), itself a subset of ML, is built on multilayered artificial neural networks and is a powerful tool used for pattern detection within complex datasets such as biomedical images, for example, cone-beam computed tomography (CBCT) scans. Convolutional Neural Networks (CNNs) and advanced DL architectures like U-Net are commonly applied to capture spatial and contextual features in 2 dimensional (2D) or 3 dimensional (3D) images with high precision.

The biological goal of endodontics is the prevention and treatment of periradicular disease. Therefore, the accurate detection of periapical lesions (PLs) is a key component of endodontic decision-making, as their presence or absence significantly influences treatment choices, including to watch and observe, or to intervene by nonsurgical or surgical retreatment. Historically, intraoral radiography has been the standard diagnostic tool for PL detection. The introduction of 3D radiography in the form of cone-beam computed tomography (CBCT) has significant potential to uproot this traditional role of periapical radiography.

CBCT images are encoded in the digital imaging and communications in medicine (DICOM) format, which offers general advantages, including interoperability, standardization, and metadata attachment, allowing for easy transfer between machines and applications from different manufacturers. Interpreted by human clinicians, CBCT allows for an excellent lesion detection accuracy (sensitivity 0.95 and specificity 0.88) for periapical disease when compared to 2D radiographs. Currently, it is considered the most effective tool for odontogenic lesion detection.

However, while CBCT allows for improved PL detection compared to 2D radiography, limitations still exist for its practical use. A low interobserver and intraobserver agreement in CBCT interpretation has been observed. Low sensitivity and specificity for PL detection in CBCT images have been detected if a lesion is associated with endodontically treated teeth. Thus, there is an unmet need to develop and clinically implement AI applications to detect PL in CBCT images. Moreover, the interpretation of CBCT volumes is time-consuming and demands substantial experience and expertise. ^,

For AI-assisted lesion detection, Sadr and colleagues provided a systematic review and meta-analysis. The authors included 12 studies on diagnostic test accuracy and reported a sensitivity range for radiographic detection of PL to be 0.65 to 0.96. However, there was significant variation in the methods of these selected publications. For example, the study by Ekert and colleagues applied CNNs for PL detection in panoramic radiographs, with a sensitivity of 0.65 and a specificity of 0.87. While the latter study presented a 2D approach, Orhan and colleagues utilized CBCT and reported a lesion detection accuracy of 0.95 sensitivity. Setzer and colleagues developed a DL algorithm for automated PL detection that also allows for multilabel segmentation. Segmentation refers to the partition of an image into multiple categories. It is a common approach to delineate and classify different structures in biomedical images. It is primarily a classification task. Their study labeled PL, tooth structure, bone, restorative materials, and background and achieved a high PL detection accuracy (sensitivity of 0.93 and specificity of 0.88). Kirnbauer and colleagues utilized a periradicular region of interest approach, first detecting PL in whole arch CBCTs, followed by segmenting and classifying the PL, their model achieved 0.93 sensitivity and 0.88 specificity. Recently, transformer-based architectures have emerged as strong contenders in biomedical imaging. Chen and colleagues compared a transformer-based algorithm, pretrained on diverse CBCT scans of medical origin to a classic U-Net architecture. The transformer model achieved perfect sensitivity (1.00) and 0.94 specificity, indicating that these architectures may outperform traditional CNNs for volumetric imaging tasks ( Fig. 1 ).

Full-3D multilabel segmentation with periapical lesion detection of dental limited field-of-view CBCT of mandibular anterior sextant. Periapical lesion on mandibular right central incisor. Comparison of original CBCT slices with fully automated segmentation of the identical areas with the AI platform.

( Courtesy o f Rui Qi Chen, Center for Machine Learning, Georgia Institute of Technology, Atlanta, GA.)

AI-based applications have also been used for attempts to obtain differential diagnosis of the subtype of odontogenic PLs. PLs may present as an abscess, cyst, or granuloma. While abscesses can be diagnosed by, for example, pain or swelling, and so forth, if acute, or by the presence of a sinus tract if chronic, cysts and granulomas traditionally could only be distinguished histologically. A traditional view is that periradicular cysts may persist or even continue to proliferate after nonsurgical endodontic intervention. While there is no definite evidence of this hypothesis, knowledge of the subtype of PL may influence treatment decisions, for example, opting for a surgical approach after failed root canal treatment if the lesion has definitely been identified as a cyst. An earlier approach to employ CBCT-based technology for differential diagnosis was undertaken by Simon and colleagues, who achieved a 76.5% accuracy in differentiating granulomas from cysts by grayscale analysis. AI-based applications include ML approaches such as the one by Okada and colleagues, who used semiautomated segmentation of CBCT volumes in conjunction with an AI algorithm and reached a 94.1% agreement with Simon and colleagues’ earlier result from their dataset. The DL model presented by Ver Berne and colleagues reached an area under the curve of 0.97 for cystic lesions and 0.88 for granulomas, evaluating 2D panoramic radiographs.

DL algorithms are particularly known for their excellent performance regarding image analysis. This is owed to their ability to objectively extract features from biomedical images, which has surpassed human performance as early as 2016. Thus, AI-based applications hold the promise further to distinguish lesions of endodontic origin from other jaw lesions. For example, for lesions greater than 10 mm in diameter, several studies have shown excellent results in differentiating odontogenic keratocystic tumors from ameloblastomas. ^,,

A recent scoping review by Aminoshariae and colleagues investigated the potential for the use of AI-based applications for endodontic education. The authors sought out AI applications applied in endodontics and further reviewed publications from the medical field to draw comparisons to dentistry, with a specific focus on endodontics. Of the 185 studies initially identified, 35 were included in a more detailed review to isolate 10 areas where AI can enhance endodontic education. These areas included (1) radiographic interpretation, (2) differential diagnosis, (3) treatment planning and decision-making, (4) case difficulty assessment, (5) preclinical training, (6) advanced clinical simulation, (7) real-time clinical guidance, (8) robotics and autonomous, systems, (9) student progress evaluation, and (10) calibration and standardization of teaching curricula. Notably, the selected studies demonstrated applications of core AI, such as ML or DL, but also the use of virtual reality (VR) and augmented reality (AR) for training and guidance purposes. While preclinical training and advanced clinical simulation can be used for predoctoral and postgraduate endodontic training, these techniques may also become useful for continuing education for clinicians and calibration of educators. From an administrative perspective, AI can help to assess student performance, and adapt individual learners’ curricula to address weaknesses, create study aids and examination questions, and provide interactive feedback.

Clinical areas involve radiographic interpretation, differential diagnosis, treatment planning and decision-making, case difficulty assessment, and robotics and autonomous systems that will aid not only students during their education but also clinicians in everyday endodontic practice. AI-powered chatbots, driven by Large Language Models, will play a significant role in the educational aspects of endodontics. Learners and educators are provided easier access and improved interaction with the medical knowledge by interpreting clinical scenarios, simulating patient interactions, and providing personalized learning experiences with 24/7, on-demand support. This applies not only to the aforementioned diagnostic aid and clinical guidance but also to academic aspects of the education. Kavadella and colleagues demonstrated that students who used ChatGPT for literature searches performed better during assignments, as the iterative questioning process encouraged critical thinking. Chatbots may also be integrated with VR and AR platforms to provide realistic clinical simulations.

AI can provide objective, consistent analysis. In contrast to human practitioners, it can help to reduce interobserver variability and improve alignment with evidence-based guidelines. However, the responsibility of biomedical image interpretation lies with the prescribing provider. Especially when considering the significant data contained in a CBCT volume, reviewing complete datasets is not only time-intensive but also bears the risk of overlooking anomalies. AI-driven decision support systems can offer “second opinions” by identifying suspicious regions. By reducing the variability inherent to manual interpretations, AI helps ensure that patients receive uniform, high-quality care regardless of the treating clinician’s experience level. AI can provide risk scores based on large-scale datasets and existing clinical outcome data. In endodontics, prognostic tools have been developed for nonsurgical root canal treatment. Lee and colleagues predicted the outcome of endodontic treatment after 3 years based on a database of periapical radiographs of 598 single-rooted premolars. The authors defined success as a periapical index (PAI) score of 1, and failure as a PAI score of 4 or 5, or radiographic evidence of extraction. The PRESSAN-17 DL model achieved a significantly higher accuracy of predicting the correct prognosis compared to residual network (RESNET-18) (67.0% vs 63.4%; Fig. 2 ). While this was yet a modest outcome, no additional data beyond the radiographic information were available to the AI algorithms. Future applications may include patient factors such as medical history or tooth-related clinical factors, for example, the quality of an existing restoration, likely leading to improved accuracy.

( A–C ) Visualization of 3 “test set” case examples. Each example showing the gray-scale preoperative periapical radiographs and the corresponding gradient-weighted class activation map (Grad-CAM) heat map of each feature or endodontic prediction. A red region represents a larger weight. Four clinical features and the endodontic outcome prediction Grad-CAM heat map were superimposed on a preoperative preprocessed image. COD, coronal defect; PAR, periapical radiolucency; PRF, previous root filling; ∗incorrect prediction.

(Junghoon Lee et al., An Endodontic Forecasting Model Based on the Analysis of Preoperative Dental Radiographs: A Pilot Study on an Endodontic Predictive Deep Neural Network, Journal of Endodontics, 49 (6), 2023, 710-719, https://doi.org/10.1016/j.joen.2023.03.015 .)

AI tools can automate much of this decision-making process by prescreening volumetric datasets for anomalies, pathologies, or structural variations. This aids in streamlining the diagnostic workflow and complex decision-making processes, reduces unnecessary referrals, and ultimately saves health care costs. AI-supported triage tools, already employed during the coronavirus disease 2019 pandemic in the medical field to prioritize cases that require urgent attention, can help practitioners not only to improve workflow efficiency in high-volume clinical settings but also to render appropriate treatment, essentially working as a “digital assistant,” seamlessly integrated with clinical software to provide real-time guidance.

A recent example is the development of a chatbot-based application for treatment guidance for dental trauma, as predicted by Aminoshariae and colleagues. The management of dental trauma remains a challenge for clinicians, as patients often present on an emergency basis, requiring time-sensitive and appropriate treatment decisions, as errors may have significant long-term consequences. The Dental Trauma Evo Chatbot was designed to provide immediate, standardized recommendations based on the guidelines of the International Association of Dental Traumatology. Based on ChatGPT-4 application programming interface (API), the application works in over 50 languages and was validated based on 32 varying trauma scenarios, with an almost perfect score of correct treatment recommendations.