One of the most exciting developments in recent years has been the increasingly widespread application of artificial intelligence (AI). According to IBM, this sophisticated technology “leverages computers and machines to mimic the problem-solving and decision-making capabilities of the human mind.”
For the purposes of dentistry, the two most relevant branches of AI are machine learning and deep learning. Machine learning involves training a system to recognize common patterns within a large data set and ascribe a particular meaning to them. For example, you may input a large data set of images that are labeled, for instance, for stage, phase or place in a process. The system learns how these labels and images match and is then able to recognize them on its own when presented with an unlabelled image. Deep learning builds on this, using more complex, layered neural networks that are capable of self-learning with very little human intervention.
1. Caries detection
Early detection of dental caries can make a big difference to the patient’s prognosis, enabling early intervention by the dental professional that can reverse or arrest incipient caries lesions. AI has shown great potential as a tool to support this goal.
In one study, a dental AI system detected caries with an impressive 92.5% success rate. While this was initial research into the technology and the authors acknowledged system limitations, these initial findings are promising. Another recent study found that clinicians evaluating bitewing radiographs with the support of an AI system were able to detect caries lesions more accurately, especially enamel lesions, associated with an increase in the number of true positives (caries lesions that were identified as carious). However, it was noted that this had led to an increase in both non-invasive and invasive treatment decisions, suggesting the need for more guidance when using AI systems. A similar study, this time assessing the clinician’s ability to detect proximal enamel caries lesions, found that AI increased the ability to correctly identify enamel caries lesions by 71%, while also finding an 11% increase in healthy proximal surfaces incorrectly identified as carious. In this case, the increase in enamel lesions correctly identified would support decision-making for early non-invasive intervention of those lesions and although there was a small increase in false positives the end result would also support only non-invasive preventive measures.
In addition, AI systems are being researched for their potential in the evaluation of periapical radiographs for a diagnosis of deep caries lesions and pulpitis. In a study evaluating three AI systems for the ability to assess depth, one system in particular showed significantly greater precision and accuracy than experienced clinicians although, as with other studies, the authors noted limitations that must be overcome. Refinement of AI systems for caries detection will help to unlock their full diagnostic potential.
2. Periodontal disease
As with dental caries, early detection and treatment of periodontal disease can greatly influence patient outcomes. AI tools relating to periodontal disease are currently in development and show promising early results. In a 2022 systematic review published in the Journal of Prosthetic Dentistry, intraoral images and radiographs were assessed using AI, which was found to:
Detect biofilm with 73.6% to 99% accuracy.
Diagnose gingivitis with 74% to 78.20% accuracy.
Detect alveolar bone loss with 73.4% to 99% accuracy.
Diagnose periodontal disease with 47% to 81% accuracy.
At the recent EuroPerio10 congress, hosted by the European Federation of Periodontology (EFP), researchers presented the findings of a study (Artificial intelligence shows promise for interpreting dental X-rays, 16 Jun 2022) in which AI was used to examine bitewing radiographs of patients with periodontitis. The AI system demonstrated good precision and sensitivity in detecting total alveolar bone loss, horizontal bone loss, and dental calculus. Sensitivity and precision were relatively low for furcation defects and the system was unable to detect vertical bone loss. While more research and development is required, the system is described as a ‘glimpse into the future’.
AI is also used for disease risk assessment. In an observational study, researchers investigated the use of AI technology and an algorithm that would predict an individual’s systemic risk of periodontal disease based on risk factors such as age, smoking habits and diet. Using non-invasive and easily collected information, this represents a future opportunity for rapid risk assessment, early intervention and possible prevention of periodontal disease.
AI is increasingly being used in endodontics, with applications that include: determining the location and measurements of certain tooth structures; identifying specific endodontic pathologies; assessing the viability of the tooth; and predicting the success of treatment.
In a systematic review, researchers described how AI had been used to detect periapical lesions with an impressive 92% reliability. They also discussed how an AI model was used to locate the apical foramen and determine the working length of the tooth. In one study in the review, working length determination in dry skulls was 93% accurate. In another study with cadavers, the accuracy of working length determination reached 96% with AI, compared to 76% accuracy for experienced endodontists.
Orthodontic treatment requires a significant amount of planning, for which AI is proving useful. It is increasingly used in treatment planning, anatomical landmark detection, assessment of growth and development, and evaluation of treatment outcomes.
One example is clear aligner treatment, which uses intelligent algorithms to determine the optimal movement of each individual tooth based on intraoral scans of the patient’s dentition. The algorithm plots the progression of the treatment to simulate the end result, relying on the clinician to feed back the actual progression. With every piece of data fed back, the treatment plan is adjusted accordingly and the algorithm is refined.
Other potential applications include estimating the size of unerupted teeth, assessing the need for extraction, classifying patient growth patterns, automated tooth segmentation, and detecting activation patterns of tongue musculature.
5. Osteoarthritis classification
Temporomandibular joint osteoarthritis (TMJOA) involves the progressive degradation of the hard and soft tissues around the temporomandibular joint. AI has demonstrated the ability to classify 3D images of the mandibular condyle into five stages of structural degenerative changes, achieving 91% close agreement with clinician consensus and an established classification system.
6. Cancer detection
Oral and oropharyngeal cancers will affect an estimated 54,000 American adults in 2022. The overall five-year survival rate is estimated to be 67%, but this increases to 83% with early detection and treatment — something AI may well be able to support.
In a recent study, an AI model was able to achieve an F1 score (which includes precision and recall) of 87% for the identification of images containing lesions and 78% for those in need of referral. It was concluded that this initial research was promising for AI identification and classification of oral and oropharyngeal cancers.
In another study, AI models were able to detect head and neck cancers with specificity and accuracy of 78% to 81.8% and 80% to 83.3%, respectively, compared to data showing specificity and accuracy of 83.2% and 82.9% respectively for specialists. AI has also demonstrated the ability to distinguish between two types of similar maxillary tumors with specificity and accuracy of 81.8% and 83.3% respectively, compared to 81.1% and 83.2% from clinical specialists. Further, AI did so in 38 seconds, compared to the clinicians’ average of 23.1 minutes.
The use of AI models is somewhat limited by the size and quality of the data available. However, as researchers work on expanding the available data sets, we can expect accuracy to increase.
In the meantime, researchers stress that AI is not a replacement for the expertise of skilled clinicians. AI is best viewed as an intelligent assistant in diagnostic and therapeutic care, providing rapid, reliable data to inform clinical decision making. Ultimately, this could translate to improved protocols and health outcomes for patients.