With increasing digitalization in orthodontics, certain orthodontic manufacturing processes such as the fabrication of indirect bonding trays, aligner production, or wire bending can be automated. However, orthodontic treatment planning and evaluation remains a specialist’s task and responsibility. As the prediction of growth in orthodontic patients and response to orthodontic treatment is inherently complex and individual, orthodontists make use of features gathered from longitudinal, multimodal, and standardized orthodontic data sets. Currently, these data sets are used by the orthodontist to make informed, rule-based treatment decisions. In research, artificial intelligence (AI) has been successfully applied to assist orthodontists with the extraction of relevant data from such data sets. Here, AI has been applied for the analysis of clinical imagery, such as automated landmark detection in lateral cephalograms but also for evaluation of intraoral scans or photographic data. Furthermore, AI is applied to help orthodontists with decision support for treatment decisions such as the need for orthognathic surgery or for orthodontic tooth extractions. One major challenge in current AI research in orthodontics is the limited generalizability, as most studies use unicentric data with high risks of bias. Moreover, comparing AI across different studies and tasks is virtually impossible as both outcomes and outcome metrics vary widely, and underlying data sets are not standardized. Notably, only few AI applications in orthodontics have reached full clinical maturity and regulatory approval, and researchers in the field are tasked with tackling real-world evaluation and implementation of AI into the orthodontic workflow.

Natural language processing (NLP) is an intersection between Computer Science and Linguistic which aims to enable machines to process and understand human language. We here summarized applications and limitations of NLP in dentistry. Narrative review. NLP has evolved increasingly fast. For the dental domain, relevant NLP applications are text classification (e.g., symptom classification) and natural language generation and understanding (e.g., clinical chatbots assisting professionals in office work and patient communication). Analyzing large quantities of text will allow understanding diseases and their trajectories and support a more precise and personalized care. Speech recognition systems may serve as virtual assistants and facilitate automated documentation. However, to date, NLP has rarely been applied in dentistry. Existing research focuses mainly on rule-based solutions for narrow tasks. Technologies such as Recurrent Neural Networks and Transformers have been shown to surpass the language processing capabilities of such rule-based solutions in many fields, but are data-hungry (i.e., rely on large amounts of training data), which limits their application in the dental domain at present. Technologies such as federated or transfer learning or data sharing concepts may allow to overcome this limitation, while challenges in terms of explainability, reproducibility, generalizability and evaluation of NLP in dentistry remain to be resolved for enabling approval of such technologies in medical devices and services. NLP will become a cornerstone of a number of applications in dentistry. The community is called to action to improve the current limitations and foster reliable, high-quality dental NLP. NLP for text classification (e.g., dental symptom classification) and language generation and understanding (e.g., clinical chatbots, speech recognition) will support administrative tasks in dentistry, provide deeper insights for clinicians and support research and education.

The development of artificial intelligence (AI) in dentistry requires large and well-annotated datasets. However, the availability of public dental imaging datasets remains unclear. This study aimed to provide a comprehensive overview of all publicly available dental imaging datasets to address this gap and support AI development. This observational study searched all publicly available dataset resources (academic databases, preprints, and AI challenges), focusing on datasets/articles from 2020 to 2023, with PubMed searches extending back to 2011. We comprehensively searched for dental AI datasets containing images (intraoral photos, scans, radiographs, etc.) using relevant keywords. We included datasets of >50 images obtained from publicly available sources. We extracted dataset characteristics, patient demographics, country of origin, dataset size, ethical clearance, image details, FAIRness metrics, and metadata completeness. We screened 131,028 records and extracted 16 unique dental imaging datasets. The datasets were obtained from Kaggle (18.8%), GitHub, Google, Mendeley, PubMed, Zenodo (each 12.5%), Grand-Challenge, OSF, and arXiv (each 6.25%). The primary focus was tooth segmentation (62.5%) and labeling (56.2%). Panoramic radiography was the most common imaging modality (58.8%). Of the 13 countries, China contributed the most images (2,413). Of the datasets, 75% contained annotations, whereas the methods used to establish labels were often unclear and inconsistent. Only 31.2% of the datasets reported ethical approval, and 56.25% did not specify a license. Most data were obtained from dental clinics (50%). Intraoral radiographs had the highest findability score in the FAIR assessment, whereas cone-beam computed tomography datasets scored the lowest in all categories. These findings revealed a scarcity of publicly available imaging dental data and inconsistent metadata reporting. To promote the development of robust, equitable, and generalizable AI tools for dental diagnostics, treatment, and research, efforts are needed to address data scarcity, increase diversity, mandate metadata completeness, and ensure FAIRness in AI dental imaging research.

Convolutional Neural Networks (CNNs) have long dominated image analysis in dentistry, reaching remarkable results in a range of different tasks. However, Transformer-based architectures, originally proposed for Natural Language Processing, are also promising for dental image analysis. The present study aimed to compare CNNs with Transformers for different image analysis tasks in dentistry. Two CNNs (U-Net, DeepLabV3+), two Hybrids (SwinUNETR, UNETR) and two Transformer-based architectures (TransDeepLab, SwinUnet) were compared on three dental segmentation tasks on different image modalities. Datasets consisted of (1) 1881 panoramic radiographs used for tooth segmentation, (2) 1625 bitewings used for tooth structure segmentation, and (3) 2689 bitewings for caries lesions segmentation. All models were trained and evaluated using 5-fold cross-validation. CNNs were found to be significantly superior over Hybrids and Transformer-based architectures for all three tasks. (1) Tooth segmentation showed mean±SD F1-Score of 0.89±0.009 for CNNs, 0.86±0.015 for Hybrids and 0.83±0.22 for Transformer-based architectures. (2) In tooth structure segmentation CNNs also outperformed with 0.85±0.008 compared to Hybrids 0.84±0.005 and Transformers 0.83±0.011. (3) Even more pronounced results were found for caries lesions segmentation; 0.49±0.031 for CNNs, 0.39±0.072 for Hybrids and 0.32±0.039 for Transformer-based architectures. CNNs significantly outperformed Transformer-based architectures and their Hybrids on three segmentation tasks (teeth, tooth structures, caries lesions) on varying dental data modalities (panoramic and bitewing radiographs). As deep-learning-based image analysis is part of modern dentistry, practitioners and dental researchers should be aware of strength and limitations of modern model architectures for dental-image analysis. Models that demonstrate optimal performance in other domains do not necessarily constitute the optimal selection for the purpose of dental imaging.

The imbalanced nature of real-world datasets is an ongoing challenge in the field of machine and deep learning. In medicine and in dentistry, most data samples represent patients not affected by pathologies, and on imagery, pathologic image areas are often smaller than healthy ones. Selecting suitable loss functions during deep learning is essential and may help to overcome the resulting imbalance. We assessed six different loss functions for one exemplary task, tooth structure segmentation on bitewing radiographs, for their performance. Six different loss functions (Focal Loss, Dice Loss, Tversky Loss and hybrid losses of Cross-Entropy and Dice Loss, Focal and Dice Loss, Focal and Generalized Dice Loss) were compared on a tooth structure segmentation task of 1,625 bitewing radiographs. Training was performed using three different model architectures (U-Net, Linknet, DeepLavbV3+) over a 5-fold cross-validation. Tooth structures consisted of the classes (occurrence in% of samples/captures areas measured on pixel level) enamel (100 %/25 %), dentin (100 %/50 %), root canal (100 %/10 %), filling (81 %/8 %) and crown (28 %/5 %). Hybrid loss functions significantly outperformed standalone ones and provided robust results over the different architectures for the classes enamel, dentin, root canal and filling. Specifically, the Dice Focal loss reached high performance to conquer both image level and pixel level class imbalance, respectively. In dental use cases it is often important to predict minority classes such as pathologies accurately. Using specific loss function may be an effective strategy to overcome data imbalance when training deep learning models.

Supervised deep learning requires labelled data. On medical images, data is often labelled inconsistently (e.g., too large) with varying accuracies. We aimed to assess the impact of such label noise on dental calculus detection on bitewing radiographs. On 2584 bitewings calculus was accurately labeled using bounding boxes (BBs) and artificially increased and decreased stepwise, resulting in 30 consistently and 9 inconsistently noisy datasets. An object detection network (YOLOv5) was trained on each dataset and evaluated on noisy and accurate test data. Training on accurately labeled data yielded an mAP50: 0.77 (SD: 0.01). When trained on consistently too small BBs model performance significantly decreased on accurate and noisy test data. Model performance trained on consistently too large BBs decreased immediately on accurate test data (e.g., 200% BBs: mAP50: 0.24; SD: 0.05; p < 0.05), but only after drastically increasing BBs on noisy test data (e.g., 70,000%: mAP50: 0.75; SD: 0.01; p < 0.05). Models trained on inconsistent BB sizes showed a significant decrease of performance when deviating 20% or more from the original when tested on noisy data (mAP50: 0.74; SD: 0.02; p < 0.05), or 30% or more when tested on accurate data (mAP50: 0.76; SD: 0.01; p < 0.05). In conclusion, accurate predictions need accurate labeled data in the training process. Testing on noisy data may disguise the effects of noisy training data. Researchers should be aware of the relevance of accurately annotated data, especially when testing model performances.