Machine learning in dental photography

Introduction

The use of photography in dentistry has undergone significant advancements, becoming an essential tool in clinical practice, documentation, and communication between dentists, patients, and dental technicians. Origi­nally utilized for basic record-keeping, dental photo­graphy now plays a pivotal role in patient care, enhancing diagnosis, treatment planning, and documentation. Digital photography, in particular, has revolutionized the field of dentistry, providing instant access to high resolution images, which can be stored, analyzed, and shared effortlessly. Photography in dentistry serves various purposes, including documenting clinical cases, facilitating medico-legal processes, and educating both patients and practitioners. Additionally, dental photography has become fundamental in marketing dental practices by showcasing treatment outcomes through images before and after treatment [1]. Despite its growing importance, the adoption of dental photography is still limited due to various factors, such as high cost and the need for specialized training. In one study, only a third of surveyed dentists used photography regularly, citing expensive equipment and lack of knowledge as the barriers. However, the advantages of digital photo­graphy, such as easy storage, rapid sharing, and the ability to correct images instantly, suggest significant potential for broader adoption in the future [2].
As technology continues to evolve, the field of dental photography is likely to expand further. The integration of smartphones and advanced digital cameras with specia­lized lenses, enables clinicians to capture detailed intraoral and extraoral images efficiently. Additionally, innovations, such as the use of mobile applications for photography and the rise of teledentistry, underscore the potential for photography to further enhance dental care by bridging the gaps in communication and diagnosis [1, 2].
Machine learning (ML) represents a rapidly advanc­ing field at the intersection of statistics and computer science, enabling computers to identify patterns and make decisions from large sets of data. In the realm of medicine, ML has the potential to revolutionize dia­gnostic accuracy and treatment planning by analyzing provided detailed images. In dental photography, ML offers significant advantages over traditional methods by automating image enhancement, abnormality detection, and disease classification. Conventional approaches rely on manual interpretation, which can be time-consuming and subjective, whereas ML-powered analysis ensures faster, more consistent, and objective assessment of dental images. These algorithms can refine image quality, adjust lighting and contrast, and highlight areas of concern, aiding clinicians in more accurate diagnoses. Super­vised learning, a key category of ML, can be applied to classify images based on known outputs, such as identifying dental diseases or abnormalities. For example, ML algorithms can help in detecting cavities, gingivitis, or even early stages of oral cancer by analyzing photographic data, ultimately augmenting diagnostic capabilities of dental professionals. Unsupervised learning, on the other hand, offers the potential to uncover new patterns in dental images that are not immediately visible, paving the way for personalized treatment plans based on unique patient features [3]. ML is transforming medicine, especially in medical imaging where it enhances diagnostic accuracy and clinical decision-making. ML techniques, i.e., radiomics and deep learning, are applied to various tasks, such as risk assessment, detection, diagnosis, prognosis, and response to therapy. These methods convert medical images into quantitative, mineable data, creating image-based phenotypes that support precision medicine. Deep learning, a subset of ML, allows computers to learn complex patterns from large datasets, improving disease detection and characterization, notably in cancer. However, the successful implementation of ML in medical imaging requires large, well-annotated datasets and robust computational infrastructure. The potential of ML in clinical decision-making is already being explored, combining radiologists’ expertise with cutting-edge technology to deliver more personalized and efficient healthcare [4, 5]. ML enhances medical imaging by improving diagnosis, prognosis, and treatment planning. In stroke care, ML analyzes imaging data to assess severity and predict outcomes [6]. It also advances disease detection in conditions, such as cancer and diabetic retinopathy through automated image analysis [7]. In oncology, ML-driven histopathology improves oral cancer diagnosis and prognosis, identifying key biomarkers for personalized treatment [8].
ML is making significant advancements in dentistry, especially in clinical decision-making and treatment planning. Recent developments have led to clinical decision support systems (CDSS), which assist dentists across specialties, including orthodontics, periodontics, and prosthodontics. These systems utilize algorithms, such as neural networks and genetic algorithms to analyze complex clinical data and provide treatment recommendations, improving diagnostic accuracy and patient outcomes. For instance, ML models predict the need for tooth extractions in orthodontics, classify periodontal diseases, and forecast the longevity of dental restorations [9].
ML, as a subset of AI, has also demonstrated beneficial in improving dental diagnostics. Using algorithms to detect patterns in complex datasets, ML aids clinicians in making more accurate diagnoses. Deep learning, a type of ML utilizing multi-layered neural networks, has been especially effective in analyzing dental radiographs to detect caries and periodontal bone loss. This approach reduces human error and improves predictive accuracy by learning from new data continuously. Artificial intelligence (AI) has the potential to revolutionize dentistry by streamlining tasks, reducing costs, and enabling more personalized patient care [10]. In orthodontics, ML has emerged as a powerful tool, advancing diagnosis, decision-making, and treatment planning. Neural networks and deep learning are increasingly used to enhance the accuracy of malocclusion diagnosis and treatment decisions. ML models help predict bone age, identify anatomical landmarks, such as skeletal structures in cone-beam computed tomography (CBCT) scans, and reduce manual work, leading to greater diagnostic precision. These models also support decision-making, e.g., predicting whether tooth extraction is necessary or determining the optimal timing for intervention [11]. Moreover, ML and AI are transforming orthodontics by integrating AI into routine tasks, such as cephalometric analyses and 3D image segmentation. These tools minimize manual work and enhance diagnostic accuracy and efficiency. AI-powered CDSS predict treatment outcomes and assist with key decisions, such as whether extractions are necessary in orthodontic treatment. As data storage and analysis capabilities continue to expand, AI’s potential to process large datasets will further improve both clinical outcomes and orthodontic research [12].
The aim of this paper was to assess the available lite­rature on the use of ML in dentistry, focusing on dental photography and its application in various fields of dentistry.

Material and methods

Based on the PubMed database, 51 articles were obtained using the following key words: “machine learning dental photography”, “artificial intelligence”, “digital dentistry”. Studies types included systematic reviews, reviews, meta-analyses, and original research. Inclusion criteria were as follows: publications from 2014 to 2024, English-written articles, and studies related to dental photography application. Papers concerning other fields of medicine and lack of access to full-text articles were excluded. After applying exclusion criteria, 23 articles were selected and grouped according to their respective fields of dentistry and applications.

Results

The utilization of ML in dental photography can be very wide, encompassing many fields, such as odonto­logy, diagnostic process in conservative dentistry, resto­rative dentistry, orthodontics, oral medicine, and esthetic medicine.

Odontology

Dataset of non-carious teeth presented in a study by Chaudhary et al. [13] played a pivotal role in advancing ML applications in dental imaging. Images captured from various maxillary and mandibular views allowed for the training of AI models to identify and assess diffe­rent dental conditions. By focusing on healthy dentition across a wide age range, from children aged 1 to 14 years, this dataset offered significant value for developing baseline models capable of detecting dental anomalies in future phases. Such datasets enabled a thorough explo­ration of non-carious dental conditions, and could serve as a benchmark for evaluating diagnostic accuracy in automated systems. For instance, image segmentation techniques might benefit from the dataset’s diversity in terms of lighting conditions, angles, and occlusal surfaces, ensuring robust model development for practical, real-world applications in dentistry. This study showed the potential to improve early diagnostic capabilities in remote areas or regions with limited dental care access. However, it is important to acknowledge limitations, including the need for expanding the dataset to encompass broader geographical diversity that would enhance the generalizability of AI models in diverse populations. Moreover, a study by Ragodos et al. [14] explored the use of convolutional neural networks (CNNs) in detecting dental anomalies from intraoral photographs. The model employed in this study was based on ResNet-18, and achieved promising results in identifying different types of dental anomalies, such as hypoplasia and hypocalcification. The study’s use of transfer learning allowed the model to leverage pre-trained weights from larger datasets, significantly enhancing its predictive power. This approach presented a significant advantage in dental anomaly detection, especially in large-scale research settings where manual analysis of photographs is labor-intensive and prone to human error. Notably, the model’s speed and efficiency, processing 38,486 photos in just 16 minutes compared with a probable 1 year of human analysis, highlighted the potential of deep learning in revolutionizing dental diagnostics. However, challenges remain, particularly in addressing data imbalance and ensuring accuracy in complex cases involving orthodontic appliances.
In a study conducted by Li et al. [15], deep learning models were applied to classify isolated human teeth using photographs, emphasizing the potential of AI in dental education and forensic odontology. The top-3 accuracy of over 90% in all classification types and the average area under the curve (AUC) of 0.95 demonstrated the effectiveness of AI in automated tooth identification, mostly in educational settings where such technology could reduce the workload of dental professionals. However, the study also underscored some challenges, such as the anatomical similarities between teeth, which can lead to misclassifications, especially in non-standard anatomy cases. The ability of AI to perform at levels comparable to human experts despite difficulties, suggested that with further refinement and larger datasets, AI models could become reliable tools for tooth identification in various applications, including archaeology and forensics.
The use of AI in dental photography showed pro­mising advancements, particularly in the detection and classification of dental diseases, such as enamel hypo­mineralization (EH) and molar-incisor hypominera­lization (MIH). A recent study validated the external performance of an AI-based method that classified EH/MIH based on dental photographs. The AI system demonstrated an overall accuracy of 94.3% in detecting and classifying hypomineralized enamel defects, with specificity and sensitivity values between 81.7% and 98.7%, depending on the type of lesion. The summary of AI models used in dental diagnostics, along with their reported accuracy, is presented in Table 1. These results highlighted the AI potential to support early dia­gnosis and treatment planning, by providing reliable, image-based classification and segmentation of hypo­mineralized regions, especially in cases where visual diagnosis can be challenging. This approach could revolutionize diagnostic workflows, reducing clinician workload while maintaining high diagnostic accuracy [16].

Optimization of hygiene procedure

A study by Chen et al. [17] introduced DeepPlaq, a deep learning-based method for dental plaque indexing using intraoral images. This model employed a three-stage approach, integrating YOLOv8 for tooth detection, a segment anything model (SAM) for tooth segmentation, and a CNN-based classification system for plaque assessment based on a Quigley-Hein plaque index (QHI). DeepPlaq achieved an accuracy of 0.776, outperforming other models, including ResNeXt50 and ResNet50. The model’s ability to provide precise and consistent plaque scoring was shown as a promising step towards achieving automating repetitive tasks in dental care, alleviating the clinicians’ workload and improving the consistency of plaque assessments. However, the study emphasized the need for further improvements, including expanding the dataset and addressing class imbalance to enhance generalizability of the model in diverse populations.
In dental photography, the ability to visualize micro­structures, such as cavitation bubbles, plays a crucial role in understanding the dynamics of dental instruments, e.g., ultrasonic scalers. High-speed imaging, coupled with ML for image analysis, as demonstrated in a study on cavitation around scaler tips, provides a novel approach for enhancing dental procedures. The study successfully used high-speed cameras to capture cavitation bubble formation around different scaler tips, applying ML algorithms to quantify the cavitation area. This method not only enabled better visualization of dental processes, but also optimized the use of cavitation as a cleaning mechanism for dental plaque removal. The integration of image processing techniques, such as binary thresholding and ML, simplified the analysis process, making it accessible and replicable in various clinical settings. This advancement was shown significant, as it aids in optimizing instrument performance and minimizing the risk of tissue damage during dental procedures [18].

Caries diagnostics

Artificial intelligence can facilitate and accelerate the work of a dentist. In general dentistry, tools assisting and accelerating the process of detecting carious lesions, especially in intraoral photographs, are particularly useful. A study conducted by Duong et al. [19] presented a novel approach for automated caries detection using smartphone color photography combined with ML algo­rithms. The application of support vector machines (SVMs) for classifying caries lesions based on smartphone images, demonstrated a promising, cost-effective solution for clinical diagnostics, particularly in settings where the access to advanced imaging equipment may be limited. The model achieved a high accuracy of 92.37% in the classification of cavitated versus non-cavitated lesions, with sensitivity and specificity rates of 88.1% and 96.6%, respectively. This performance, while robust, highlighted the potential of smartphones as accessible diagnostic tools in teledentistry, especially in underserved communities. However, the model’s lower specificity for visually non-cavitated versus no-surface-change categories (66.7%) indicated that further refinement is required to improve early lesion detection. Despite these limitations, the integration of smartphone imaging and ML in caries detection, demonstrated a significant step towards more widespread use of AI-driven diagnostics in dental care.
To summarize the application of AI in caries detection in a standardized manner, a systematic review was conducted by Moharrami et al. [20], to assess the efficacy of AI models in detecting dental caries using oral photographs, highlighting the growing relevance of this technology in teledentistry. AI models showed a wide range of performance across different image acquisition devices, with F1-scores for caries detection ranging from 42.8% to 95.4%. Notably, professional cameras yielded the highest performance, followed by intraoral and smartphone cameras. This variability indicated the difficulty of achieving consistent accuracy, parti­cularly with smartphone images, where performance is more limited (F1-scores between 42.8% and 80%). Even though the authors suggested that the use of AI in caries detection can enhance clinical decision-making by providing objective assessments, further studies with robust designs and standardized metrics are needed. The potential of AI to reduce diagnostic variability, especially in resource-limited settings, underscores its importance, yet the need for higher-quality data from smartphones remains a significant challenge.
As described by Xiong et al. [21], the use of deep learning models, such as ToothNet, marks a significant step forward in the application of AI for diagnosing dental issues, including caries and fissure sealants in dental photography. ToothNet performed impressively, achieving an AUC of 0.925 for detecting caries, outperforming a dentist with one year of experience. This highlighted the strength of AI in dental diagnostics. However, while ToothNet showed superior accuracy and sensitivity in detecting caries compared with less-experienced dentists, its performance in detecting fissure sealants was slightly lower, with an AUC of 0.902. The authors suggested that while AI can greatly enhance diagnostic accuracy and efficiency, improvements are still needed, particularly for models handling multiple tasks. Therefore, to increase the versatility of such models in clinical settings, expanding of the dataset is critical to ensure they can accurately diagnose a wider range of dental conditions in diverse populations.
The application of deep learning in dental photography showed substantial potential in automating the detection of common dental anomalies, such as white spot lesions (WSLs). A pilot study explored the use of CNNs, i.e., SqueezeNet, to detect white spot lesions from clini­cal photographs. The results demonstrated promising accuracy, with models achieving a mean accuracy of up to 84%. This study emphasized the challenges of detecting subtle dental anomalies, such as WSLs, where human clinicians may struggle, particularly in distinguishing between caries lesions, fluorosis, and other hypomineralized enamel defects. The SqueezeNet model proved effective in classifying fluorotic white spot lesions, achieving a sensitivity of 0.66 and specificity of 0.86. These findings suggested that deep learning models can assist clinicians in the early detection and classification of white spot lesions, providing a valuable diagnostic tool in preventive dentistry. However, light reflections on teeth were identified as a primary source of false positives, indicating the need for improved image acquisition protocols and further model refinement to enhance diagnostic reliability in clinical practice [22].

Orthodontic diagnostics

A study by Noeldeke et al. [23] investigated the use of CNNs in detecting crossbites using 2D intraoral photographs, demonstrating the growing role of deep learning in orthodontic diagnostics. Among six models tested, the Xception architecture achieved the highest accuracy (98.57%) in distinguishing between crossbite and non-crossbite images, while DenseNet performed best in differentiating between lateral and frontal crossbites. These results highlighted CNNs’ ability to identify malocclusion types from photographic data, supporting their potential use in orthodontic diagnosis, particularly in remote or underserved areas, where access to specialists may be limited. Nevertheless, the study reported challenges in classifying more complex cases, such as distinguishing between different types of crossbites, due to the smaller sample size and overlapping clinical features. This limitation emphasized the need for larger datasets and further refinement of AI models for more nuanced orthodontic conditions.
On the other hand, CNNs can be a useful tool to assess soft tissue profiles from facial photographs and predict the necessity of orthognathic surgery, according to Jeong et al. [24]. The CNN model, based on the VGG19 architecture, achieved an accuracy of 89.3% in distinguishing between patients who required surgery and those who did not, based solely on their facial photographs. This underscored the potential of AI in augmenting the diagnostic process in maxillofacial surgery, particularly for early assessment based on non-invasive imaging. However, the study highlighted limitations, such as the model’s inability to account for functional aspects of dentofacial deformities, including occlusion or breathing issues, which are critical in determining the need for surgery. While CNNs provide a promising tool for initial screening, the authors noted that a comprehensive diagnosis should still include functional evaluations alongside esthetic considerations.
The application of ML in dental photography, particularly in orthodontics, demonstrated promising advancements. One significant area of focus was the use of deep learning models for cephalometric landmark detection. Automated detection through ML models was shown to improve both the accuracy and efficiency of orthodontic diagnosis rather than using traditional methods that often rely on manual tracing and subjective interpretation. These traditional methods are prone to intrapersonal and interpersonal variations, leading to inconsistencies in treatment planning. By incorporating CNNs, which are capable of learning from large datasets, the margin for error in landmark identification is significantly reduced. Studies reported high detection accuracy, with rates reaching over 86% within a 2 mm tolerance. This reduction in diagnostic time and the minimization of errors holds substantial potential for improving clinical outcomes in orthodontics and dental photography [25].
The integration of AI into orthodontic diagnostics has revolutionized the way clinicians approach the eva­luation of dental crowding and treatment planning. A recent study showcased the application of CNNs to assess crowding severity and the necessity for tooth extractions using intraoral photographs. Among the models tested, VGG19 demonstrated the highest accuracy for both landmark detection and categorization of dental crowding, with minimal mean errors of 0.84 mm in the maxillary region and a Cohen’s weighted k coefficient of 0.73, indicating substantial agreement in crowding categorization. In addition, this model achieved an accuracy of 92.2% for predicting the need for extractions, highlighting the potential of CNNs in aiding complex orthodontic decision-making that traditionally rely on subjective judgments of clinicians. This advancement not only streamlines the diagnostic process, but also enhances the precision of treatment planning, reducing potential biases related to clinical experience and interpretation [26].

Oral medicine

A study by Vinayahalingam et al. [27] highlighted the promising role of deep learning models, such as Mask R-CNN with Swin Transformers, particularly in detecting oral potentially malignant disorders (OPMDs). The ability of the model to accurately identify oral squamous cell carcinoma (OSCC), demonstrated significant progress in the non-invasive diagnosis of OSCC from photographic images. However, the model’s effectiveness in detecting both leukoplakia and oral lichen planus was moderately successful, indicating that although ML has great potential, there are challenges in achieving consistent accuracy in all types of OPMDs. These results suggested that while the technology can greatly assist clinicians in early cancer detection, mainly OSCC, further refinement is needed for other types of OPMDs, especially in differentiating between disorders presenting with similar clinical features.
A systematic review by Gomes et al. [28] showed the growing integration of deep neural networks and AI in various medical fields, including oral medicine. In dental photography, AI-based models exhibited promising outcomes in the automated classification of lesions, with high accuracy and sensitivity when compared with human experts. Such technological advancements hold potential for enhancing diagnostic precision, especially in detecting early-stage oral pathologies. However, challenges, including heterogeneity of oral lesions and uneven quality of images, affect the reproducibility of AI models in real-life clinical settings. The development of standardiz­ed imaging protocols and rigorous pre-processing, such as lighting adjustments and image scaling, was reported essential in overcoming these issues and improving the robustness of the models in oral health diagnostics. While AI holds the potential to complement human expertise, particularly in underserved areas, the technology is not yet ready to replace clinical judgment.
Moreover, AI-driven methods hold significant poten­tial for the automated detection of periodontal diseases, such as gingivitis, from dental photographs. In a comparative study of deep learning models, the ResNet and GoogLeNet models demonstrated superior performance, achieving accuracy levels of up to 97% in the identification of chronic gingivitis from intraoral images. The study further explored ensemble learning, combining predictions from multiple models, which improved classification performance. This technology could augment traditional diagnostic methods by enabling rapid, non-invasive detection of gingivitis based on photographs, reducing clinician burden and facilitating early intervention. Also, the application of these models in mobile devices or intraoral cameras could empower patients to perform self-assessments, further advancing preventive care in dentistry [29].
A study by Pedrinaci et al. [30] underscored the importance of integrating digital tools, e.g., multifunctional anatomical prototypes (MAPs), in dental photography and treatment planning for patients with excessive gingival display due to altered passive eruption. MAPs allow clinicians to use advanced diagnostic imaging technologies, including CBCT and intraoral scans, to enhance precision during diagnosis, treatment planning, and surgical execution. This digital workflow not only supports the identification of key anatomical landmarks, but also serves as an effective communication tool for patients, providing them with a clear visualization of expected outcomes before undergoing procedures, such as crown lengthening. While MAPs represent a step forward in the precision and patient communication, it is important to consider that their success relies heavily on the operator’s expertise and accuracy of digital workflows, highlighting the need for further developments in 3D printing and AI-based tooth segmentation techno­logies to fully realize their potential in clinical practice.

Extraoral analysis

In a study by Patcas et al. [31], the application of AI to evaluate the impact of orthognathic treatment on facial attractiveness and estimated age, demonstrated the growing potential of ML in dentistry. The CNNs used in this research showed a high level of accuracy in determining changes in both the appearance and attractiveness after treatment. A significant improvement in facial attractiveness was observed in 74.7% of patients, especially in those undergoing lower jaw surgery. Interestingly, AI models were able to estimate the apparent age of patients pre- and post-surgery, revealing a rejuvenating effect in patients who had undergone profile-altering surgery. This underscores AI’s potential to provide objective and data-driven assessments, which supplement the subjective nature of clinical evaluation, reducing bias in esthetic assessments, and improving treatment outcomes in facial surgeries. Nevertheless, while AI offers consistency in evaluation, the authors emphasized that it should complement rather than replace patient’s own perceptions and expectations regarding their esthetic results.
A review conducted by the same first author empha­sized the transformative potential of AI in medico-dental diagnostics, particularly when applied to facial images. The review outlined how AI, specifically through deep learning models, demonstrated superior accuracy in recognizing and diagnosing facial features, pathologies, and even esthetic outcomes. One of the key opportunities was the use of AI to evaluate facial attractiveness and predict age, with neural networks trained on vast datasets proving to be valuable tools for personalized treatment planning. For instance, AI can objectively assess the impact of orthodontic or maxillofacial interventions by quantifying changes in facial features, making clinical decisions less subjective. However, the review underscored significant challenges, such as ethical concerns regarding data privacy and the potential bias in AI models trained on non-medical datasets. This illustrates the need for careful curation of training data and rigorous regulatory oversight, to ensure AI applications in dentistry are both safe and reliable [32].

Machine learning limitations

Despite its potential, the implementation of ML in dental photography faces several challenges. Data privacy concerns arise due to the need for large, annotated datasets, often requiring sharing of sensitive patient information. Additionally, bias in ML models remains a significant issue, as algorithms trained on non-representative samples may produce inaccurate or inconsistent results across diverse patient populations. Addressing this bias requires various, high-quality training data and rigorous validation. Another key limitation is the generalizability of ML models in clinical practice. Many ML algorithms perform well in controlled research settings, but struggle when applied to real-world clinical environments with variable imaging conditions and patient factors.

Conclusions

The application of ML in dental photography pre­sents substantial opportunities across diverse fields, such as odontology, orthodontics, and oral medicine. ML models, particularly CNNs, have proven effective in automating the detection and classification of dental conditions, enhancing diagnostic precision and reducing clinician workload. With many ML-based diagnostic systems achieving accuracy rates of around 90%, these tools demonstrate significant reliability, making them valuable assets in clinical decision-making and everyday practice. From the early detection of caries using smartphone images to the assessment of orthodontic anomalies and periodontal diseases, AI-driven technologies are reshaping diagnostic workflows. Despite these advancements, challenges remain, including data imbalance, variability in image quality, and the need for more extensive datasets. As research continues to refine these models, the potential for ML to transform dental diagnostics, especially in remote and underserved areas, grows ever stronger.

Disclosures

1. Institutional review board statement: Not applicable.
2. Assistance with the article: None.
3. Financial support and sponsorship: None.
4. Conflicts of interest: The authors declare no potential conflicts of interest concerning the research, author­ship, and/or publication of this article.

References