Personalized dental medicine, artificial intelligence, and their relevance for dentomaxillofacial imaging

Personalized medicine refers to the tailoring of diagnostics and therapeutics to individuals based on one’s biological, social, and behavioral characteristics. While personalized dental medicine is still far from being a reality, advanced artificial intelligence (AI) technologies with improved data analytic approaches are expected to integrate diverse data from the individual, setting, and system levels, which may facilitate a deeper understanding of the interaction of these multilevel data and therefore bring us closer to more personalized, predictive, preventive, and participatory dentistry, also known as P4 dentistry. In the field of dentomaxillofacial imaging, a wide range of AI applications, including several commercially available software options, have been proposed to assist dentists in the diagnosis and treatment planning of various dentomaxillofacial diseases, with performance similar or even superior to that of specialists. Notably, the impact of these dental AI applications on treatment decision, clinical and patient-reported outcomes, and cost-effectiveness has so far been assessed sparsely. Such information should be further investigated in future studies to provide patients, providers, and healthcare organizers a clearer picture of the true usefulness of AI in daily dental practice.

Current concepts of managing dental diseases have by large been developed over the course of the last 50 years.While knowledge generated by continuous research efforts towards the biological foundation of the main dental diseases (caries and periodontitis) has been gradually integrated into contemporary therapy approaches, the backbone of treatments employed in dental practices has been established decades ago.For example, restorative treatments remain the cornerstone for carious lesions while deep scaling and root planing remain central for periodontal disease, both of which are increasingly accompanied by preventive efforts. 1,2iven the evolving understanding of dental diseases, their etiology and pathogenesis, and the resulting chance and need to adequately describe different disease stages and grades to deduct appropriate therapies, this simplification may not suffice any longer.Notably, it is grounded in a similarly simplified diagnostic approach; what is missing is a systematic and holistic evaluation of individual health and disease on patient, tooth, and site level, and the synthesis of the gathered data into adequately granular diagnoses.Such an approach would need to be built on a detailed multimodal data collection and would allow to assign individualized treatment pathways based on personalized diagnosis.
At present, however, such individualized diagnostics and treatment pathways are not at all available in dentistry.Instead, we are stuck in the era of stratification of individuals and lesions into risk groups, characterized mainly by simple shared phenotypic characteristics (e.g., caries experience for caries risk assessment, smoking or poor oral hygiene for periodontal risk assessment, etc.).Currently, the accuracy and generalizability of most of these risk assessment systems are insufficiently validated.Even if these risk assessment systems were valid, they would only describe groups of individuals and lesions sharing a similar "risk" and subsequently assign identical management strategies to all individuals in a certain risk group (i.e., the one-size-fits-all approach). 3hile being the next step beyond stratification, true personalized management is not possible at the moment.Personalized management is closely linked to "precision medicine", defined as "the tailoring of a therapy to individuals with one's biological (genomic, microbiomic, proteomic, etc.), social (economic, educational, etc.) and behavioral (lifestyle) characteristics". 3Personalized care should, ideally, allow to provide the safest, most efficacious and efficient diagnostics and therapies, which is jointly with precision medicine and closely related to another concept "P4 medicine". 4The four Ps stand for a more precise, personalized, preventive, and participatory healthcare approach (Figure 1).What is needed, however, to make personalized, precision, and P4 dentistry come true is a deep understanding of individuals and the option to predict what will happen to this individual, a specific organ or lesion.
To allow such understanding and prediction, the discussed concept of stratification and the employed few risk indicators or factors (Table 1) are obviously insufficient.What is needed is a shift towards a healthcare model centered around broad and deep data.As discussed elsewhere, 7 many recent academic breakthroughs in astronomy, 8 biology 9 and other disciplines are mainly driven by making use of large amounts of data.Dentistry should also make use of the wealth of available dental data and transform into something that was previously referred to as "data dentistry". 7The data needed could be generated from advanced sensor technologies, including wearables, ingestibles, and implantables as well as social media and electronic health records (eHR), to name a few. 10 Many of these data sources will not solely rely on being collected in clinical settings, but routinely, even by patients who may actively donate data from social media, food consumption, healthcare apps, behavioral diaries, or toothbrushes.In addition, prospectively collected omics data may become more available if costs for generating them decrease further and technologies are becoming available in routine settings. 10gure 1 The confluence of different data sources and technologies (e.g., AI, specifically deep learning, systems medicine involving genomic, metabolomic, or microbiomic data, as well as clinical data sources or those provided publicly or by the patient) will enable P4 medicine and dentistry. 4 of 22 Dentomaxillofac Radiol, 52, 20220335 Personalized dental medicine, AI and their relevance for DMFR Hung et al

Artificial intelligence and its use in dental medicine
The analysis of such diverse, multimodal, large, and complex data, including speech and imagery, requires advanced data analytic approaches. 11One major strategy adopted over recent years for this purpose is "artificial intelligence" (AI).The term was coined in the 1950's and refers to the idea of building machines that are capable of performing tasks that are normally performed by humans.Machine learning (ML) is a subfield of AI where algorithms are applied to learn the inherent statistical patterns and structures in data, which allows for predictions of unseen data.More complex machine learning algorithms frequently used for data like images are neural networks (NNs), which are constituted of artificial neurons (i.e., mathematical non-linear models that can be stacked and concatenated in layers using mathematical operations to form a network).The term "deep learning" is a reference to deep (multilayered) NNs, which are able to represent hierarchical features in complex data and frequently used for detecting edges, corners, shapes, and macroscopic patterns in images. 12L and NNs as a subtype of AI are "trained" to automatically perform specific tasks, and the most common type of training is supervised learning where data points and corresponding data information (e.g., labels, tasks, etc.) are repetitively passed through the network to detect the intrinsic statistical patterns in the data.During the training process, the connections between the neurons, also referred to as model weights, are optimized to minimize the so-called prediction error (difference of the true vs the predicted data information).A trained NN can predict the outcome of unseen data by passing the new data point through the network.AI, ML, and NNs are increasingly used in dentistry to work with the increasing amount of data available, as described above.A number of such forms of use are currently discussed or already clinically available:

Data analytics and precision dentistry
As discussed, there is an increasing strive towards more precise data-centered dentistry, making use of not only clinical and historical data, claims and treatment data, image and further test data, but also data provided by patients as outlined above.A big advantage in dentistry is that these multimodal datasets are usually available repeatedly as many patients visit dentists regularly.Using such longitudinal data will help to foster a deeper understanding of individual health and disease and to develop AI models to predict disease onset or progression individually.Currently, however, many of these data remain siloed or unavailable.Meanwhile, existing AI prediction models remain limited in their predictive power and generalizability as useful predictions need to be better than plainly guessing the so-called majority class (i.e., the more frequent event). 3Predicting this majority class is easy but models which focus on exclusively predicting it may not be clinically useful. 13

Evidence-based care
Gathering a more comprehensive picture of an individuals' health and objectifying diagnosis through imagery and AI-assisted analysis will support evidence-based care.Data-centric approaches will further allow to embed external evidence, for example from guidelines and standards of care, into decision making, and then fostering reliable high-quality and cost-effective care. 14n additional benefit of more data-driven care is the option to objectively assess treatment needs, actually provided treatments, and the yielded outcomes.Ultimately, this should foster value-based care (i.e., quantifying the "value" of a certain treatment to individuals and the society).

Beyond the dental chair
AI and data-driven approaches will facilitate better information and decision making on the dental public health level, including workforce planning.Automated

Risk factor Risk indicator
Definition "A characteristic that may make an individual more susceptible to a certain disease" 3 ; can be "environmental, behavioral, or biologic and "if present directly increases the probability of a disease occurring, and if absent or removed reduces the probability". 5 marker that is not necessarily causally linked, but can be used to predict risk, like past disease experience or social, educational or economic factors" 3 ; "may be a probable, or putative, risk factor, but […] a temporal association usually cannot be specified".

Current use of AI in dentomaxillofacial imaging
Radiographic examination is an integral component in most diagnostic and treatment planning processes in daily dental practice.With the growing use of digital dental radiography, images generated by dental radiographic examinations are commonly automatically stored as digital data in the archiving system and associated databases.These data can be analyzed using AI and specifically deep learning based on convolutional NNs. 17 Currently, a range of deep learning models have been trained and tested on dentomaxillofacial radiographic images to fulfill tasks of image classification (e.g., "is there a certain pathology detectable on this image?"),object detection (e.g., "in which image area is this certain pathology located?") and pixelwise segmentation (e.g., "which pixels of this image show a certain pathology"), and for image quality improvement (Table 2). 95,96agnosis

Dental caries
Intraoral radiographic examination is essential for the detection of dental carious lesions, particularly early non-cavitated ones.The sensitivity and specificity of intraoral radiography for detecting dental caries were reported to range from 27-66% and 76-97%, respectively. 97,98The relatively low sensitivity reported implies a high underdetection of dental caries, which may be related to clinicians' experience and caries lesion depth (i.e., enamel or dentin caries).Several deep learning models have been developed to assist clinicians in detecting and classifying dental caries. 96Lee et al developed three CNN models to automatically detect dental caries in posterior teeth on periapical images. 18The models showed higher detection accuracy for premolars than for molars, which could be related to differences in their anatomical characteristics.Srivastava et al developed a CNN model to detect dental caries on bitewings.The AI model achieved significantly higher sensitivity (81%) than three general dentists (34-48%). 19More recently, caries detection using AI additionally focused on the detection of early enamel caries.The CNN model developed by Cantu et al outperformed seven experienced dentists in detecting initial enamel and advanced dentin caries. 20The seven dentists showed greatly different sensitivities for detecting initial (<25%) and advanced (40-75%) caries while the model achieved robust sensitivities (>70%) for both initial and advanced caries.Currently, commercial AI software programs including AssistDent (Manchester, UK), Denti.AI (Toronto, Canada), Diagnocat (Tel Aviv, Israel), CranioCatch (Eskişehir, Turkey) and dentalXr.ai (Berlin, Germany) (Table 3) are available to assist clinicians in the diagnosis of dental caries on two-dimensional (2D) radiographic images.The use of AssistDent and dentalXr.ai significantly increased dentists' sensitivity especially for detecting enamel caries. 21,99Notably, automatic detection of buccal/lingual caries or secondary caries (i.e., caries next to restorations) remains challenging for AI models.This is, however, also the case for human observers and mainly grounded in the 2D nature of most intraoral images.While cone-beam computed tomography (CBCT) allows caries detection in threedimensions (3D), it is not recommended for caries diagnostics.

Periodontal bone loss
Deep learning models have also been developed for the detection and segmentation of periodontal bone loss and the associated classification of periodontitis stages on periapical and panoramic images.In 2018, Lee et al developed a CNN model on periapical images to automatically identify periodontally compromised posterior teeth and predict tooth loss in the future. 28The accuracy of the model was higher for premolars (>80%) than for molars.Thanathornwong et al developed a CNN model to identify periodontally compromised teeth on panoramic images. 29im et al 22 and Krois et al 23 trained their CNN models to automatically detect periodontal bone loss on panoramic radiographs.The diagnostic accuracies of their models (AUCs [area under the curves] of 0.89-0.95)were higher than that of several general dentists (AUCs of 0.77-0.85).
In addition, periodontitis stages can also be classified automatically using deep learning on periapical and panoramic images. 26,27Danks et al 24 and Lee et al 25 developed CNN models to measure the extent of periodontal bone loss on periapical images and subsequently classify the identified sites into three/four severity stages according to the bone loss extent measured.The model by Lee et al achieved high classification accuracy with an AUC value of 0.98.Future applications are expected to detect changes in the bone density and textures of the alveolar ridge for early detection of the onset of periodontitis.Detection of ameloblastomas and odontogenic keratocysts 50 Detection/classification of ameloblastomas, odontogenic keratocysts, dentigerous cysts, radicular cysts, and/or bone cysts in the maxilla/mandible 51,52 Differentiation of Stafne's bone cavity from mandibular radiolucent lesions 53

Maxillary sinus
Detection of maxillary sinus lesions 54,55 Detection and segmentation of maxillary sinus lesions 56,57 Prediction of oroantral communication after tooth extraction 58

Maxillofacial fracture
Detection and classification of mandibular fracture 72

Lymph node metastasis
Segmentation and identification of metastatic cervical lymph nodes 77

Reporting of the dental status
Segmentation of teeth and jaws, numbering of teeth, detection of caries, periapical lesions, and periodontitis 78 Identification of missing tooth, caries, filling, prosthetic restoration, endodontically treated tooth, residual root, periapical lesion, and periodontal bone loss 79 Tooth numbering and detection of dental implants, prosthetic crowns, fillings, root remnants, and root canal treatment 80 Detection, segmention, and labeling of teeth, crowns, fillings, root canal fillings, implants, and root remnants 81,82 Tooth detection and numbering 83,84 Tooth segmentation and classification 85,86 Image quality improvement Correction of blurred panoramic radiographic images 87 Reduction of metal artifacts on CBCT images 88 Improvement of the resolution of CT/CBCT images [89][90][91]

Multimodal image registration
Registration of CBCT with intra oral scan, 92 optical dental model scan, 93 or MRI 94 3), as demonstrated by an increased sensitivity from 59.6 to 73.3%. 101Orhan et al used 109 CBCT scans to test an AI software named Diagnocat (Tel Aviv, Israel; Table 3) and reported high detection accuracy and no significant differences in the lesion volumes measured by the software and a radiologist. 31Notably, the presence of endo-perio lesions, buccal-lingual cortical perforations, incomplete apex, endodontically treated teeth, and large lesions associated with multiple teeth detrimentally affected the model's performance.
Diagnosis of root fractures, especially vertical root fractures, is a challenging and experience-dependent task, commonly accomplished by combined clinical and radiographic examination.Root fractures are categorized as horizontal and vertical fractures.Horizontal root fractures frequently occur in the anterior teeth due to dentoalveolar trauma while vertical root fractures are common in endodontically treated teeth as a result of excessive root canal preparation or occlusive force.CNN models have been developed to automatically detect vertical root fractures on 2D and 3D radiographic images. 32,33Despite promising diagnostic accuracy, the models still have to overcome a relatively low accuracy on non-endodontically treated teeth and the potential impact of caries, fillings, dental restorations, and metal artifacts on their performance.
Automatic detection and classification of C-shaped canals in mandibular second molars have been seen as another field of AI application.5][36] Their performance has been shown similar or superior to both general dentists and specialists. 34,36 Dental implants CNN models were also developed to detect peri-implant bone loss and implant fractures on 2D radiographic images.Liu et al developed a CNN model to automatically detect peri-implant bone loss on periapical images.37 The model performed similarly to two general dentists but inferior to one specialist.Another CNN model on periapical images measured peri-implant bone loss ratio and classified the bone loss severity into normal, early, moderate, and severe.38 Lee et al developed CNN models on periapical, panoramic, or both images to detect implant fractures and to classify the fractured implants into horizontal or vertical fractures.43 The models achieved AUCs of 0.90-0.98 for the detection task and 0.75-0.87for the classification task.The highest detection and classification accuracies were achieved on periapical images, likely due to higher spatial resolution of periapical images compared with panoramic images.

Maxillofacial pathologies
Treatment options and prognosis for patients with pathologies in the maxillofacial region are directly associated with the timing and accuracy of diagnosis.The differential diagnosis of maxillofacial pathologies is a challenge for general practitioners, particularly for incidental findings on diagnostic images.A delayed diagnosis will lead to a longer disease course, more invasive surgical approach, and poorer treatment outcome, especially for malignant lesions.Several researchers tried to develop AI tools to improve the diagnostic accuracy of general practitioners for various maxillofacial pathologies to reach the level of specialists.Poedjiastoeti et al developed a CNN model on panoramic images for automatic detection of ameloblastomas and odontogenic keratocysts, with high diagnostic performance (sensitivity and specificity over 80%) being on par with five oral-maxillofacial surgeons. 50Another CNN model on panoramic images detected and classified ameloblastomas, odontogenic keratocysts, dentigerous cysts, and radicular cysts, and obtained high classification performance with an AUC of 0.94, sensitivity of 88.9%, and specificity of 97.2%, respectively. 51The model by Endres et al outperformed 14 oral-maxillofacial surgeons in detecting infections, granuloma, cysts, and tumors in the jaws on panoramic images. 49Lee et al developed CNN models, respectively, on panoramic and CBCT images to detect, segment, and classify odontogenic keratocysts, dentigerous cysts, and radicular cysts. 52The model on CBCT images (AUC = 0.91) outperformed the one on panoramic images (AUC = 0.85).Ariji et al et al developed a CNN model on contrast-enhanced CT images to identify and segment metastatic cervical lymph nodes in patients with oral cancer. 77The model outperformed two radiologists in identifying cervical lymph nodes while its segmentation accuracy should be improved.It has been reported that inexperienced oralmaxillofacial radiologists are prone to miss pathological changes of the parotid gland while interpreting CT images of the maxillofacial region, leading to underdetection of Sjögren's syndrome. 76Kise et al developed a CNN model to assess the texture features of the parotid gland on CT images for automatic diagnosis of Sjögren's syndrome. 76The model performed similarly to three experienced radiologists and superior to three inexperienced radiologists.
The maxillary sinus is the largest paranasal sinus and frequently involved in dental surgical procedures due to its close proximity to the teeth in the posterior maxilla.][104] However, it has been reported that inexperienced dental practitioners were less likely to accurately diagnose sinus pathologies on radiographic images. 1055][56][57] The models obtained favorable performance on both detection and segmentation tasks.Murata et al reported that their CNN model performed similarly to two radiologists and outperformed two dental residents in the diagnosis of maxillary sinusitis. 55The CNN model by Hung et al obtained high accuracy for detecting and segmenting mucous retention cysts and mucosal thickening of the sinus on both ultra-low-dose and standarddose CBCT images with AUCs ranging from 0.84 to 0.93. 56

Temporomandibular joint
Diagnosis of temporomandibular joint (TMJ) disorders requires sufficient clinical experience.Undetected TMJ problems can result in patients suffering for a long time and undergoing unnecessary examinations and even invasive treatment.Jung et al developed two CNN models on panoramic images using different pre-trained flameworks for automatic diagnosis of TMJ osteoarthritis.The models achieved excellent diagnostic accuracy superior to that of three general dentists and even three TMJ specialists. 69The CNN model by Kim et al obtained high accuracy for measuring cortical thickness of the mandibular condyle head on CBCT images. 71ishiyama et al developed CNN models to diagnose mandibular condyle fracture on panoramic images, and reported high diagnostic accuracy with AUCs of nearly 0.9. 70

Other diagnostic purposes
Apart from the abovementioned diagnostic purposes, deep learning models can also be developed for automatic detection and classification of mandibular fractures, 72 diagnosis and prediction of osteoporosis, 74,75 detection of submandibular gland sialoliths, 73 differentiation of Stafne's bone cavity from mandibular radiolucent lesions 53 on 2D or 3D radiographic images.All these models obtained high accuracies mostly with AUC values over 0.9.

Reporting of the dental status
Charting of teeth, restorations, and present dental diseases is the first step in the routine assessment of dental patients.Any mistakes or oversights in the resulting dental records may lead to misdiagnosis and erroneous treatment decisions, such as extraction or endodontic treatment of the wrong tooth.As electronic dental health records are by now widely used in dental practice, automated charting using AI seems highly useful.Some studies reported excellent performance of CNN models for automated detection and numbering of deciduous and permanent teeth on panoramic images. 83,84Shaheen et al developed a CNN model on CBCT images for automated tooth segmentation and classification. 85The model achieved high accuracies for both segmentation and classification tasks, and has found its way into a commercially available software named Relu (Leuven, Belgium; Table 3).Fontenele et al reported that the presence of dental fillings in CBCT images negatively affected Relu's performance on tooth segmentation. 860][81][82] Commercially available systems including dentalXrai (Berlin, Germany), Denti.AI (Toronto, Canada), and Diagnocat (Tel Aviv, Israel) (Table 3) allow such charting in similar accuracy to practitioners. 781][42] These models achieved excellent classification accuracy and some even outperformed periodontists.Automatic implant classification models could be used to recognize and record the system of the placed implants in the dental recording systems, which can facilitate regular maintenance and future repairs.
Treatment planning AI has great potential to help dental practitioners with treatment planning and time-consuming tasks in the digital dental workflow.Segmentation, localization, and measurement of anatomical structures or pathologies on radiographic images as well as multimodal image registration are common manual steps required in the planning of oral and maxillofacial surgical procedures. 106o far, several AI applications have been proposed for automated landmark localization, [59][60][61][62][63][64] skeletal classification, [65][66][67] facial symmetry assessment, 68,107 and decisionmaking on tooth retention or extraction for orthodontic treatment 108,109 on 2D or 3D images.
Kunz et al developed a CNN model to automatically localize anatomical landmarks and measure their linear/ angular parameters on cephalometric radiographs. 60The mean absolute differences in the linear/angular analyses were 0.44-0.64mm/0.46-2.18°for the model and 0.35-0.88mm/0.55-1.80°for 12 orthodontists, which demonstrates similar performance.Bulatova et al 63 and Mahto et al 64 tested AI driven automated cephalometric analysis software applications named Ceppro (Seoul, Korea; Table 3) and WebCeph (Seongnam, Korea; Table 3), respectively.Ceppro achieved mean absolute localization differences ranging from 1.3 to 8.7 mm, with no significant differences between automated and manual localization for eleven out of sixteen selected landmarks.WebCeph obtained high agreement with intraclass correlation coefficients over 0.9 between automated and manual measurements on seven out of twelve cephalometric parameters.6][67] Lin et al developed a CNN model to assess facial symmetry before and after orthognathic surgery on CBCT images and reported high accuracy of 90%. 68nother group developed a CNN model for automatic detection of edentulous sites, nasal fossa, maxillary sinus, and mandibular canal, and measurement of the heights and widths of residual alveolar bone at the edentulous sites on CBCT images for dental implant treatment planning. 39The model's detection accuracy was high for edentulous sites (95.3%) and moderate for the mandibular canal (72.2%) and nasal fossa/maxillary sinus (66.4%).On the sites of maxillary premolars/molars and mandibular premolars, the automated bone height measurements were similar to the manual measurements (i.e., ground truth).The automated bone height measurements on the sites of maxillary/mandibular anterior teeth and mandibular molars as well as the automated bone width measurements on all tooth sites were significantly different from the manual measurements, with median measurement deviations of 1.7-11.3mm.The significant differences between automated and manual measurements might be due to the incorrect localization of the measuring points.
Assessment of the difficulty of planned third molar surgery is also a field of increased interest in AI research.Yoo et al developed a CNN model on panoramic images to classify the difficulty of third molar removal according to several parameters, such as the depth and angulation of the molar. 475][46] Choi et al developed a CNN model to determine whether lower third molars are truly in contact with or positioned buccally/lingually to the mandibular canal when they are shown as overlapped on panoramic images (CBCT readings served as ground truth), and to classify the non-contact molars as being buccally or lingually positioned. 46The model obtained accuracies of 72% for determining the true contact position and 81% for classifying the bucco-lingual position, outperforming six oral-maxillofacial specialists.Kim et al developed a CNN model on panoramic images to predict paresthesia due to damage of the inferior alveolar nerve during lower third molar removal, and reported high prediction accuracy with an AUC of 0.92. 48Apart from third molars, Vollmer et al attempted to develop CNN models on panoramic images to predict oroantral communication after tooth extraction. 58The prediction accuracy of the best model was similar to that of four oral-maxillofacial experts.
Multimodal image registration is a critical step in digital dental workflows where 3D images acquired from different imaging modalities, including CT, CBCT, MRI, intraoral, facial, and model scanning, are superimposed into the same coordinate frame to create a virtual augmented patient model.This is useful for treatment planning for dental implant placement, joint, salivary gland, orthognathic, and reconstructive surgeries. 110Multimodal image registration can be performed manually by aligning anatomical landmarks or semi-automatically by using the surface-based or fiducial marker registration approach.Although the semi-automatic approach is less time-consuming than the manual approach, its registration accuracy is affected by the quality of the acquired images, the presence of image artifacts, the deformation of the optical surface, and the distribution of the employed fiducial markers.Therefore, manual corrections are frequently required after semi-automatic image registration.In order to improve the efficiency and accuracy of multimodal image registration, a range of studies developed AI models to automatically register CBCTs with intraoral scans, 92 optical dental model scans, 93 or MRIs. 94ompared with conventional approaches, these models allow more accurate automated image registration in a significantly shorter time.

Image quality improvement
Due to the growing use of CBCT in daily dental practice for diagnosis and treatment planning, concerns have been raised regarding the increased risk of radiationinduced stochastic effects, particularly on radiosensitive organs such as salivary glands, thyroid glands, and eye lenses.While several low-dose CBCT protocols have been suggested and applied in practice, the perceived inferior image quality of low-dose scans may hamper their use, with clinicians nevertheless employing standard-or even high-dose scanning mode for certain imaging tasks. 111In addition, the presence of severe artifacts in CBCT images has been reported as one of the common reasons for re-exposure. 112Patient motion during scanning and metallic dental restorations are the main sources for the occurrence of movement or metal artifacts in CBCT images.
][89][90] CNN-based auto-positioning can reduce blurring occurred due to positioning errors in panoramic image by reconstructing the image with the corrected curvature. 87Hu et al 91 and Park et al 89 developed image denoizing tools using deep learning to remove noise from low-dose CT or CBCT for improving the image quality to be equivalent to high-dose scans.Hatvani et al developed a CNN-based method to enhance the resolution of teeth on cross-sectional CBCT images, which allows better visualization of the anatomical structure of teeth. 90CNN-based methods were also proposed to reduce metal artifacts in CT or CBCT images. 88,113,114hese methods identify and segment the areas of metal artifacts in the original images and then merge the original and corrected images to suppress the artifacts.

The future of AI in dentomaxillofacial imaging
Current applications of AI in dentomaxillofacial imaging mainly focus on improving diagnostic accuracy and easing the diagnostic and/or planning workflow.More and more AI applications were reported to be able to perform similarly to or even outperformed dentists (Table 4), oftentimes lifting general dental practitioners to levels equivalent to specialists.Notably, existing AI models are limited in their scope, mainly aiming to detect, segment, or classify anatomical structures or common pathologies.Rare variations or diseases have so far seldom been the focus.Given the difficulty practitioners have for diagnosing rare variations or diseases, AI models developed for such specific tasks could be truly clinically significant.With technical advances, the availability of larger data pools (including the pooling of different datasets and the uptake of federated learning, 115 and the increased usage of alternative labeling and training pathways in dentistry (such as weakly or self-supervised learning, 116 these difficult diagnostic tasks are expected to be tackled.In addition, it has been expected that AI at some point could "see" more on a certain image type than the human eye could.To do so, training of the AI on more sensitive sensor data as ground truth than the one later used during inference would be one option.Currently, this is not the case because the ground truth relies on the same sensor data and oftentimes involves human activity.The only way to achieve somewhat "superhuman" performance is by involving a larger number of practitioners to at least overcome the limitations of single dentists. 14radually, more and more AI models proposed have been tested by completely external image data acquired from different dental centers (Table 5) as suggested by the Artificial Intelligence in Dental Research guideline. 117Few models were able to achieve similar performance while most showed inferior performance on external images.Some studies have reported low cross-center generalizability of their models. 70,100It was noted that adding external images acquired from one dental center in the training dataset could increase the model's performance on the images from that external center but would decrease its performance on the images from the original center.These findings indicate that although cross-center training could improve the generalizability of AI models, the proportion of the images from different centers in the training dataset is also an influencing factor associated with the trained models' performance.Therefore, future studies should focus not only on internal testing but also on external testing.If external testing shows unfavorable outcomes, cross-center training should be considered to increase the model's generalizability.
The usefulness and efficacy of most proposed AI models in daily dental practice are still unclear based on current evidence.Although most studies reported that AI models could increase diagnostic ability of dental practitioners and reduce the time spent on time-consuming work in the treatment planning process, their true impact on real-world clinical practice is rarely discussed.In addition to a model's accuracy, future studies should focus more on its impact on treatment decision, clinical and patient-reported outcomes, and cost-effectiveness, which may be more important to patients, providers, and healthcare organizers. 118Schwendicke et al performed the first cost-effectiveness analysis of an AI application for caries detection on bitewings. 14They reported that AI showed significantly higher sensitivity than dentists, which allows more early caries to be detected, facilitates non-or micro-invasive management of the detected lesions, and thus avoids costly late retreatments.The high cost-effectiveness of dental AI applications implies that integrating AI into clinical practice has the potential to reduce healthcare cost burden, revealing their economic impact on healthcare systems.Only clinically relevant AI tools that are capable of fulfilling technical requirements with promising financial potential can attract healthcare stakeholders to continuously support their development, optimization, and application in dental medicine. 119,120Therefore, future research should assess the clinical, technical, and financial aspects of cost-effectiveness of AI applications in dental medicine to demonstrate their true usefulness in daily practice.
Moreover, the impact of AI on patient-provider interaction should not be ignored.A more positive attitude regarding dental AI applications was observed in younger and more educated individuals than in older and less educated individuals. 121Compared with younger individuals, the elderly are more sceptical towards such advanced healthcare technologies.Providers need to frame the usage of AI individually to retain trust into the care process.The output of AI should be able to help patients to objectify any diagnosis and to support visual recognition of a lesion, which can improve patientclinician communication and increase patients' trust in any derived management.

Conclusions and outlook
Personalized dental medicine should allow to provide the safest, most efficacious and efficient diagnostics and therapeutics tailored to individuals based on one's biological, social, and behavioral characteristics.Based on current evidence, true personalized dental medicine is   The model's diagnostic performance using only the root portion of the tooth was similar to the specialist and superior to the general dentist.Both the specialist and general dentist showed better diagnostic performance when reading panoramic radiographs compared with periapical images., which may facilitate a deeper understanding of the interaction of these multilevel data and hopefully bring us closer to a truer form of personalized dental care for patients in the near future. 3,7

Periodontal evaluation
Personalized dental medicine, AI and their relevance for DMFRHung et al

Dentomaxillofac Radiol, 52 , 20220335
Personalized dental medicine, AI and their relevance for DMFRHung et al

Table 1
Risk factors and risk indicators

Table 3
300mples of commercially available AI software for dental applications Hung et al to additionally provide volumetric information of the detected lesions.Krois et al100and Ekert et al30developed CNN models on panoramic images or image crops to detect apical pathologies and classify teeth into https://www.craniocatch.com/en/AI,artificial intelligence; CBCT, cone-beam computed tomography 7 of 22 Dentomaxillofac Radiol, 52, 20220335Personalized dental medicine, AI and their relevance for DMFR (where root canal fillings were less frequent).Based on their findings, cross-center training seems to be able to improve a model's generalizability.Hamdan et al reported that the diagnostic ability of eight dental practitioners to detect apical radiolucencies on periapical images increased with the aid of a commercially available AI software named Denti.AI (Toronto, Canada; Table

Table 4
Performance of the developed AI models in comparison to specialists/general practitioners

20220335 Personalized dental medicine, AI and their relevance for DMFR
Hung et alstill far from being a reality as most evidence-based up-todate clinical practice guidelines for the management of dental diseases still stratify individuals and lesions into risk and thus assign identical management strategies to all individuals in a certain risk group.A wide range of AI applications, including several commercially available software options, have been developed based on diagnostic images to assist clinicians in the diagnosis and treatment planning of various dentomaxillofacial diseases, with performance similar or even superior to that of specialists.Although these dental AI applications are seen to have the potential to enable a more precise, personalized, preventive, and participatory approach for the management of dentomaxillofacial diseases, almost all of them only work on image data obtained at a certain time point in the diagnostic or treatment process without considering other data such as individual characteristics and clinical assessment.Advanced technologies with improved data analytic approaches are expected to enrich these AI applications with diverse, multimodal, large, and complex data from the individual level (e.g., demographic, behavioral, and social characteristics; clinical data generated by records mining, clinical assessment, diagnostic imaging, omics technologies; and real-time consumer data from wearables and tracking devices), setting level (e.g., geospatial, environmental, and provider-related data), and system level (e.g., health insurance, regulatory, and legislative data)