Artificial intelligence in orthodontic diagnosis

INTRODUCTION

Day by day, the world is becoming increasingly digitalized, from the entrance of our homes, where artificial intelligence (AI) is used to ring the doorbell, to the various automobiles we use in our daily lives. Right now, the gadgets we are reading about are also a creation of AI. AI, a subfield of computer science, refers to the capability of machines to mimic human cognitive functions.^[1] Over the past decade, AI has demonstrated significant potential, offering solutions to a wide range of tasks across various domains.^[2] Earlier use of digitalization across the globe has made the clinical work of a dentist quite complicated, as the patient uses AI, but is unaware of the use of certain methods, when, and how to use them as a clinical approach.^[3] However, certain software has been a boon for dentists in today’s era, where taking an impression is not required, and planning for orthodontic treatment for a certain number of cases has been easy via aligners.^[4]

Clinical orthodontic practice often demands considerable time for performing multiple analyses that rely heavily on the orthodontist’s clinical expertise.^[5] These extensive workloads can reduce the overall efficiency of orthodontic care and limit accessibility for non-specialists due to the experience required. Recent studies have demonstrated that AI can markedly improve the efficiency of orthodontic workflows. Several commercially available AI-based software systems, such as 3Shape Dental System 2.22.0.0, Uceph 4.2.1, and Mastro 3D V6.0, have already been integrated into orthodontic practice. With continuous advancements in AI algorithms, computational power, and the growing availability of large-scale datasets, the scope of AI applications in orthodontics continues to expand, leading to ongoing performance enhancements. Regular reviews and updates on these developments enable researchers and clinicians to gain timely and comprehensive insights into this rapidly evolving field. Despite the promising outcomes achieved thus far, there remains substantial potential for further progress in AI integration within orthodontics.^[6]

AI has been extensively explored across dental and medical disciplines through numerous studies. Its applications in dentistry have already been implemented in, or are under development for, various domains such as diagnosis, clinical decision-making, treatment planning, outcome prediction, and disease prognosis.^[7] Several reviews have discussed AI in dentistry; however, the present review seeks to trace the evolution of AI from its early conceptualization to its current state, classify its various types, summarize recent advancements in dental AI research, and elucidate its relationship with evidence-based dentistry.^[8]

NEURAL NETWORKING SYSTEMS IN AI – THE TECHNICAL ASPECTS

A neural network is an AI approach that enables computers to process data in a manner inspired by the structure and functioning of the human brain. A neural network is a type of machine learning (ML) process known as deep learning. It utilizes interconnected nodes, or “neurons,” arranged in layers that mimic the structure of the human brain. In most neural networks, these nodes are connected from one layer to the next, with each connection assigned a weight that determines the influence of one neuron on another. As data move through the network, from the input layer to the hidden layers and finally to the output layer, the system progressively learns patterns and relationships within the data. This learning process allows the network to generate accurate outputs based on the input it receives. Artificial neural networks, often called neural nets, are modeled after the structure and function of biological neurons. The input layer is the first layer and receives data from external sources, passing it to the hidden layer(s). Each neuron in the hidden layer takes the weighted inputs from the previous layer, computes a weighted sum, and transfers the result to the next layer. The weights assigned to connections between neurons determine the impact of each input. These weights are adjusted during training to improve the model’s accuracy and performance. This optimization process allows the network to learn complex relationships in the data and make more accurate predictions over time. The advantages and disadvantages of neural networking system has been given in Supplementary Table 1.

AI IN ORTHODONTICS

A vast contribution can be seen of AI in orthodontics among all segments or phases during the treatment plan. Accurate orthodontic diagnosis depends on a comprehensive set of evaluations, including cephalometric, dental, and facial analyses, assessment of skeletal maturation, and examination of upper airway obstruction.^[9] These assessments collectively provide a detailed understanding of the patient’s facial profile, dental and skeletal relationships, growth stage, and airway status. Cephalometric analysis serves as a cornerstone in orthodontic diagnosis, treatment planning, and post-treatment evaluation. Traditional manual landmark identification, however, is often labor-intensive, highly dependent on clinician expertise, and prone to intra- and inter-operator variability, which can compromise both efficiency and accuracy in clinical workflows. Efforts to automate cephalometric landmark detection began as early as the 1980s, but early models exhibited significant error rates, limiting their clinical utility. In recent years, advancements in AI have led to the development of more precise and reliable automated cephalometric analysis systems, greatly improving consistency, speed, and diagnostic accuracy.^[10-13]

RADIOGRAPHIC INTERPRETATION

Radiographic assessment plays a vital role in contemporary dental diagnosis and treatment planning. Among extraoral imaging modalities, the panoramic radiograph (PR) is routinely employed to visualize the dentition and adjacent maxillofacial structures. Although its resolution is lower than that of intraoral radiographs for detailed evaluation of individual teeth, PRs can reveal important changes in mineralization, bone architecture, and surrounding anatomical structures, helping clinicians identify conditions such as dental caries, periodontal bone loss, and various osseous pathologies.

With the increasing availability of CBCT, three-dimensional imaging has become more common; however, PRs remain widely used because they are quicker to obtain and easier to interpret. Despite this, comprehensive and accurate analysis of PRs still demands significant clinical expertise and time. To support clinicians, computer-assisted diagnostic technologies have been introduced in dental imaging.

One such AI–driven platform utilizing CNNs is Diagnocat (LLC Diagnocat, Moscow, Russia). This cloud-based system allows clinicians to upload various dental radiographs, including panoramic images, which are then processed by the algorithm. The PR analysis typically takes around two minutes, after which the software generates a detailed report. This output can help direct the clinician’s attention to potential abnormalities and also serve as an effective communication tool when explaining findings and proposed treatments to patients.^[17]

COMPUTED TOMOGRAPHY (CT) AND MAGNETIC RESONANCE IMAGING (MRI) IMAGING

The Dual-Branch Generative Adversarial Network (DBGAN) is an advanced AI method designed to enable clearer and more clinically meaningful interpretation of CT and MRI scans. This innovative approach integrates the high-density bone details captured by CT imaging with the soft-tissue contrast characteristic of MRI, producing a single, comprehensive image. The fusion of these two modalities offers clinicians a more complete and informative representation of patient anatomy, potentially improving diagnostic accuracy and treatment planning across a wide spectrum of medical conditions where both imaging types are routinely used but have individual limitations.

CT imaging employs X-ray technology to produce detailed cross-sectional images of anatomical structures, particularly bones, which appear opaque to X-rays. In contrast, MRI utilizes strong magnetic fields and radiofrequency waves to generate detailed images of soft tissues, including organs, muscles, and pathological changes. By combining these complementary imaging modalities, clinicians can visualize both bone and soft-tissue structures simultaneously— enhancing diagnostic precision and providing insights that would not be apparent from either modality alone.

The DBGAN model is built upon deep learning principles and adopts a dual-branch architecture comprising multiple generators and discriminators. The generators synthesize fused images that capture the essential and complementary features from CT and MRI datasets, while the discriminators evaluate these generated images, comparing them to real ones to ensure authenticity and high visual quality. This adversarial training process continues iteratively until the system produces realistic, high-quality fused images with minimal artifacts and maximal diagnostic information.

A key innovation within DBGAN is the integration of two specialized modules: the Multiscale Extraction Module (MEM) and the Self-Attention Module (SAM). The MEM facilitates the extraction of critical multiscale features from both CT and MRI images, ensuring that fine structural details are preserved. The SAM, on the other hand, emphasizes the most relevant and distinguishing features in the fusion process, enhancing image clarity and interpretability.

Extensive testing, using both subjective visual evaluation and objective quantitative assessment, has demonstrated that DBGAN outperforms conventional fusion algorithms in terms of image quality and diagnostic value. By leveraging AI-driven fusion, radiologists can utilize the strengths of both CT and MRI while mitigating their respective limitations. This technology promises to advance diagnostic imaging by maintaining the detailed bone information typical of CT alongside the soft-tissue contrast of MRI.

A broader implication of such AI applications in radiology is their capacity to detect subtle imaging patterns that may elude human interpretation. For instance, ML analysis of brain MRI data has shown potential in recognizing early tissue alterations indicative of ischemic stroke within a narrow time frame after onset—achieving higher sensitivity than traditional human-based assessments.

Replication of AI-guided algorithms in other cancer types would be conducive to generating an unbiased, low-variance machinery for patient-focused imaging interpretation.^[18,19]

OBSTRUCTIVE SLEEP APNOEA (OSA)

OSA is a clinical disorder marked by repeated episodes of airflow cessation resulting from upper airway obstruction despite ongoing respiratory effort. If left untreated, OSA can lead to serious systemic complications, including cardiovascular and metabolic diseases, cerebrovascular incidents, cognitive decline, and an increased risk of motor vehicle accidents. Although accurate diagnosis remains a challenge, treatment selection and adherence are equally critical in managing the condition effectively. The current gold standard treatment is nasal continuous positive airway pressure (CPAP), which functions as a pneumatic splint to maintain upper airway patency. When used consistently, CPAP is highly effective. However, alternative management strategies such as behavioral interventions, oral appliance (OA) therapy, myofunctional therapy, and surgical options are also utilized depending on patient-specific needs.

ML techniques have increasingly been applied to improve diagnostic accuracy, risk stratification, and treatment prediction in OSA. Supervised classification models rely on labeled datasets for training, whereas CNNs, a subtype of ML, require less preprocessing and are particularly effective in analyzing complex physiological data such as that obtained from polysomnography (PSG). Zhu et al. (2022) employed ML algorithms and genomic data to identify genes associated with OSA and its response to CPAP therapy.^[20] Their predictive analysis revealed that individuals at elevated risk for OSA exhibited heightened immune cell activation and increased gene expression levels, both of which diminished following CPAP treatment. These findings highlight the potential of ML-based genetic profiling to improve risk assessment and predict therapeutic outcomes.

Other investigations have similarly utilized ML to identify biomarkers indicative of OSA presence and severity, including plasminogen activator inhibitor-1, tissue plasminogen activator, and soluble E-selectin. These biomarkers were shown to decrease after CPAP therapy, suggesting their utility in monitoring treatment response and guiding cardiovascular risk management. Collectively, these studies demonstrate the promise of ML in leveraging biomarker analysis to facilitate individualized and precision-based OSA management strategies.

AI USED IN ORTHODONTIC TREATMENT PLANNING

In recent decades, orthodontic practice has experienced a notable transition from traditional fixed appliances to clear aligner therapy. This trend is primarily driven by growing patient demand for treatment options that offer greater comfort, convenience, and esthetic appeal. Because clear aligners are removable, they do not hinder routine oral hygiene practices such as brushing and flossing, nor do they impose dietary restrictions. From the clinician’s perspective, aligners provide an effective alternative for managing many mild to moderate malocclusions. Reported treatment outcomes indicate a success rate of approximately 80–90%, underscoring their reliability when used in appropriately selected cases.

Clear aligners are associated with reduced pain and may lower the risk and severity of root resorption when compared to traditional orthodontic treatments. As clear aligner technology has progressed, parallel advancements have occurred in the field of AI. AI is now being used in various aspects of orthodontics, including decision-making support for permanent tooth extractions or orthognathic surgery, teledentistry applications, estimation of cervical vertebral maturation stages, and assessment of remaining patient growth. One significant application of AI is in tooth segmentation—a critical step in creating individualized treatment plans. Conventionally, this process relied on mathematical algorithms with fixed thresholds. However, recent research has demonstrated that AI-based methods can perform tooth segmentation more efficiently and, in many cases, more accurately than conventional techniques. When combined with three-dimensional (3D) dental data, AI-driven segmentation has eliminated the need for physical dental models.

The various AI-based aligner software have been featured below

ETHICAL AND LEGAL CHALLENGES

CONCLUSION

AI has experienced rapid growth across medicine and dentistry, and its influence within orthodontics continues to expand. It is increasingly recognized as a valuable problem-solving tool capable of supporting clinicians in diagnosis, treatment planning, and automated cephalometric landmark prediction. The field has seen the emergence of AI-driven workflows that streamline processes such as image anonymization, segmentation, landmark detection, orientation standardization, registration, and various quantitative or classificatory assessments.

Orthodontists continue to seek technologies that enhance diagnostic precision, improve reliability, and reduce the time required for clinical decision-making. However, much of AI training relies on expert-derived annotations, which introduces subjective variability and may affect the consistency of model outputs. Therefore, AI should be viewed as an adjunctive decision-support system that augments, rather than replaces, the clinician’s expertise and judgment. Approaches such as multimodal learning hold promise for improving diagnostic accuracy and advancing the principles of precision orthodontics. Nevertheless, challenges related to data security, patient privacy, and ethical management of clinical records must be given high priority during algorithm development. Ensuring diverse and representative datasets is essential for reducing model bias and improving generalizability. As deep learning approaches, including large-scale foundation models, become more complex and less inherently interpretable, comprehensive validation and performance testing become crucial. Furthermore, well-designed prospective clinical trials are needed to assess the real-world applicability, safety, and effectiveness of AI tools within orthodontic practice.