Assessing the Role of Artificial Intelligence in Caries Detection and Clinical Decision-Making: A Scoping Review

Abstract

Background

Artificial intelligence (AI) support is expected to increase accuracy and improve treatment plans in dentistry. Nevertheless, AI’s ability to promote better oral healthcare is underexplored. This scoping review explores the influence of AI in supporting dental professionals with caries detection and decision-making regarding interventions.

Summary

Primary articles indexed on PubMed, Web of Science, Embase, Scopus, and Cochrane Library were searched until August 2025. Studies reporting the differences between participants’ caries diagnosis process and decision-making with and without AI were included. Studies only reporting algorithm accuracy, in vitro studies, or studies without an outcome related to caries detection with AI support were excluded. No time and language limits were imposed. Outcomes regarding the influence of AI on the diagnostics process and decision-making were retrieved and narratively summarised.

Key Messages

Thirteen publications were included. Number of participants ranged from 3 to 74, comprising dentists with varying experience and expertise and dental students. Results showed that AI enhances sensitivity, though its impact on specificity varies (10 studies). AI can promote unnecessary interventions for early-stage caries lesions (1 study). AI increased assessment time in two out of three cases (3 studies). AI’s cost-effectiveness is uncertain, as greater sensitivity did not lead to better economic outcomes (1 study). In conclusion, AI has the potential to improve diagnostics and influence treatment choices, but current evidence is limited and inconsistent regarding its impact on specificity, decision quality, and cost-effectiveness. Longitudinal studies in clinical settings with long-term follow-ups are needed to understand AI's impact on decision-making.

Introduction

Decision-making is critical in dentistry, particularly in the detection and management of dental caries, one of the most common oral diseases worldwide [1]. Dentists must collect information, diagnose, and create treatment plans based on clinical examination, patient input, and diagnostic complementary tools such as dental radiographs. The latter plays a key role across all areas of dentistry, as they allow a detailed view of what cannot be seen during clinical examination. Radiographs are complementary diagnostic methods that enable dentists to detect caries lesions, periodontal disease, and identify impacted teeth or bone lesions. With this information, treatment plans can be made according to the needs of each patient. Accurate interpretation of radiographs is key for effective caries detection and its subsequent treatment decisions; however, it comes with its challenges. A systematic review [2] summarised factors such as the type (e.g., bitewings and panoramic radiographs) and quality of images and the overall difficulty of cases, with higher complexity often resulting in increased errors [3]. In addition, studies show that accuracy can be influenced by clinical knowledge and experience, time pressure, cognitive load, and the education of dentists [4–7]. Their results indicate that time pressure reduces accuracy, while greater clinical experience improves specificity but not always sensitivity. High cognitive load seems to affect novice dentists, while experienced dentists’ cognitive load increases with the difficulty of a case [4, 7]. All these factors influence the diagnostic skills of dentists, emphasising the importance of continuous training in analysing dental radiographs.

To address these challenges, diagnostic systems based on artificial intelligence (AI) are increasingly gaining attention as a supporting tool for caries detection and other dental pathologies. Studies comparing these AI tools to experienced dentists show that such technologies can successfully interpret X-rays with high diagnostic accuracy, potentially leading to more optimised treatment plans [8, 9]. This enhanced accuracy can be critical for identifying early anomalies. For instance, AI can detect early signs of dental caries, periodontal disease, and other conditions that might be overlooked by the human eye [10–13]. This emphasises AI’s potential as a valuable diagnostic tool, which is likely to become more prevalent. On the other hand, while AI can increase diagnostic accuracy, an increasing reliance on technology may reduce the importance of the dentist’s clinical judgement. This could result in a shift of diagnostic and treatment decisions towards the output of the AI programmes, potentially leading to overtreatment. Evidence has shown that dentists generally opt for more invasive treatments for conditions that might not progress if observed or treated non-invasively [14]. AI may contribute to these invasive decisions by flagging initial lesions or presenting errors. In the recent past, few empirical studies have been conducted to determine the influence of AI on the decision-making process in diagnostics. For example, dentists and dental students were evaluated in their diagnosis outcomes with and without AI-based programmes [15, 16]. While these studies suggested that AI can improve diagnostic performance, other studies showed that invasive treatment decisions increased [10].

By analysing the existing literature, this review aims to provide an overview of the current state of research, identify gaps, and suggest future studies on how AI technologies influence dental professionals’ treatment decisions. The main goal of this review is to analyse AI’s influence on dentists’ caries detection and related treatment decisions, focussing primarily on radiograph analyses while also considering other imaging modalities allowing caries detection. A secondary objective is to assess the benefits and challenges of these technologies for clinical practice. In this context, impact or influence will be defined as any measurable change in the decision-making process connected to AI. This could cover changes in diagnostic accuracy, speed of decision-making, or differences in treatment outcomes when AI-based tools are used versus when they are not.

Methods

The study protocol was established and registered in the Open Science Framework before starting the review process (https://osf.io/5ycp9). The protocol and the review were conducted according to the JBI methodologies for scoping reviews [17]. The “Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews” (PRISMA-ScR) [18] was used for reporting. The addressed review question was: How does AI influence dental professionals’ decision-making in image diagnostics? The search was updated on August 14th, 2025.

Eligibility Criteria

Publications reporting on the influence of AI on the decision-making process of dental professionals and students were included. This covers how AI tools influence diagnostic accuracy, speed of diagnosis, and treatment decisions. Studies examining and describing the decision-making processes of professionals using AI vs. those not using AI were included. The primary outcomes of interest were as follows:

• 1.AI’s influence on the diagnostic process, including diagnostic performance
• 2.AI’s influence on time during the diagnostic process
• 3.AI’s influence on treatment decisions

Secondary outcomes included the following:

• 1.Cost-effectiveness of image diagnostics with and without AI
• 2.AI’s advantages and disadvantages in diagnostics

In the protocol, self-reported confidence levels were mentioned as a secondary outcome measure. However, none of the included articles reported on this outcome. No time or language restrictions were applied. Editorials, technical standards, theses, expert opinions, websites, and books were excluded. Publications solely focused on the technical development or validation of AI tools without assessing their impact on dental professionals’ decision-making and publications that did not compare AI-assisted and non-AI-assisted decision-making processes were excluded. Although in the protocol it was not specified which conditions studies had to address, after completing full-text study selection, the review was focused specifically on AI-supported caries detection. This approach allowed for a clear and interpretable synthesis, as AI’s impact may vary across different dental conditions. Consequently, studies that did not report on caries detection were excluded, ensuring that the findings could be meaningfully summarised. The table containing outcomes of other studies retrieved by our search is available in online supplementary Tables 2 and 3 (for all online suppl. material, see https://doi.org/10.1159/000550238) and at https://osf.io/5ycp9. Studies were included regardless of the specific imaging modality employed. Most of the included studies analysed bitewing radiographs, while some used periapical radiographs and one utilised photographic images. Although no specific inclusion criterion was applied regarding the imaging modality, all studies involved image types capable of identifying caries within their respective diagnostic ranges.

Search Strategy

A preliminary search of MEDLINE, the Cochrane Database of Systematic Reviews and JBI Evidence Synthesis was conducted. Due to the exploratory nature of the research question, a scoping review was chosen as the appropriate methodology to map and summarise the existing literature and provide an overview of the topic. The final search strategy aimed to locate both published and unpublished studies. PubMed, Web of Science, Embase, Scopus, and Cochrane Library were searched by the first reviewer, supported by an experienced librarian affiliated with Radboud Medical Library, to identify articles on the topic. In addition, grey literature was searched through Google Scholar. No time and language restrictions were imposed. All identified studies were uploaded into EndNote X9 software (Thomson Reuters, Philadelphia, PA, USA) to eliminate duplicates. The search strategy for each database is shown below (Table 1).

Table 1.

Search strategy

Database	Search string
PubMed	“Artificial Intelligence”[Mesh] OR “Artificial Intelligence”[tiab] OR “Computer Reasoning”[tiab] OR AI[tiab] OR “Machine Intelligence”[tiab] OR “Computational Intelligence”[tiab] OR “Computer Vision System*”[tiab] OR “Knowledge Acquisition*”[tiab] OR “Knowledge Representation*”[tiab] AND “Diagnostic Imaging”[Mesh] OR “diagnostic imaging” [Subheading] OR “Diagnostic imag*”[tiab] OR “Medical imag*”[tiab] OR “radiograph*”[tiab] OR “roentgenograph*”[tiab] OR X-ray[tiab] OR X-rays [tiab] OR X-ray*[tiab] OR Bitewing*[tiab] OR Bite Wing*[tiab] OR OPG[tiab] OR Orthopantomogra*[tiab] OR Pantomogra*[tiab] OR Panoramic*[tiab] OR “Computer Assisted Radiographic Image Interpretation*”[tiab] AND “Dentistry”[Mesh] OR “Dentists”[Mesh] OR “Tooth Diseases”[Mesh] OR Dental[tiab] OR Dentist*[tiab] OR Oral[tiab] OR Mouth[tiab] OR Prosthodontist*[tiab] OR Periodontist*[tiab] OR Tooth*[tiab] OR “Carious Lesion*”[tiab] OR “Carious Dentin*”[tiab] OR Caries[tiab] OR Tartar*[tiab] OR Endodont*[tiab] OR Orthodon*[tiab]
Web of Science	TS=(“Artificial Intelligence” OR “Computer Reasoning” OR AI OR “Machine Intelligence” OR “Computational Intelligence” OR “Computer Vision System*” OR “Knowledge Acquisition*” OR “Knowledge Representation*”) AND TS=(“Diagnostic imag*” OR “Medical imag*” OR “radiograph*” OR “roentgenograph*” OR X-ray OR X-rays OR X-ray* OR Bitewing* OR Bite Wing* OR OPG OR Orthopantomogra* OR Pantomogra* OR Panoramic* OR “Computer Assisted Radiographic Image Interpretation*”) AND TS=(Dental OR Dentist* OR Oral OR Mouth OR Prosthodontist* OR Periodontist* OR Tooth* OR “Carious Lesion*” OR “Carious Dentin*” OR Caries OR Tartar* OR Endodont* OR Orthodon*)
Embase	exp Artificial Intelligence/OR “Artificial Intelligence”.ti,ab,kf. OR “Computer Reasoning”.ti,ab,kf. OR AI.ti,ab,kf. OR “Machine Intelligence”.ti,ab,kf. OR “Computational Intelligence”.ti,ab,kf. OR “Computer Vision System*”.ti,ab,kf. OR “Knowledge Acquisition*”.ti,ab,kf. OR “Knowledge Representation*”.ti,ab,kf. AND exp diagnostic imaging/OR “Diagnostic imag*”.ti,ab,kf. OR “Medical imag*”.ti,ab,kf. OR “radiograph*”.ti,ab,kf. OR “roentgenograph*”.ti,ab,kf. OR X-ray.ti,ab,kf. OR X-rays.ti,ab,kf. OR X-ray*.ti,ab,kf. OR Bitewing*.ti,ab,kf. OR Bite Wing*.ti,ab,kf. OR OPG.ti,ab,kf. OR Orthopantomogra*.ti,ab,kf. OR Pantomogra*.ti,ab,kf. OR Panoramic*.ti,ab,kf. OR “Computer Assisted Radiographic Image Interpretation*”.ti,ab,kf. AND exp dentistry/OR exp dentist/OR exp tooth disease/OR Dental.ti,ab,kf. OR Dentist*.ti,ab,kf. OR Oral.ti,ab,kf. OR Mouth.ti,ab,kf. OR Prosthodontist*.ti,ab,kf. OR Periodontist*.ti,ab,kf. OR Tooth*.ti,ab,kf. OR “Carious Lesion*”.ti,ab,kf. OR “Carious Dentin*”.ti,ab,kf. OR Caries.ti,ab,kf. OR Tartar*.ti,ab,kf. OR Endodont*.ti,ab,kf. OR Orthodon*.ti,ab,kf
Cochrane Library	[mh “Artificial Intelligence”] OR (Artificial NEXT Intelligence):ti,ab,kw OR (Computer NEXT Reasoning):ti,ab,kw OR AI:ti,ab,kw OR (Machine NEXT Intelligence):ti,ab,kw OR (Computational NEXT Intelligence):ti,ab,kw OR (Computer NEXT Vision NEXT System*):ti,ab,kw OR (Knowledge NEXT Acquisition*):ti,ab,kw OR (Knowledge NEXT Representation*):ti,ab,kw AND [mh “Diagnostic Imaging”] OR (Diagnostic NEXT imag*):ti,ab,kw OR (Medical NEXT imag*):ti,ab,kw OR (radiograph*):ti,ab,kw OR roentgenograph*):ti,ab,kw OR X-ray:ti,ab,kw OR X-rays:ti,ab,kw OR X-ray*:ti,ab,kw OR Bitewing*:ti,ab,kw OR Bite NEXT Wing*:ti,ab,kw OR OPG:ti,ab,kw OR Orthopantomogra*:ti,ab,kw OR Pantomogra*:ti,ab,kw OR Panoramic*:ti,ab,kw OR (Computer NEXT Assisted NEXT Radiographic NEXT Image NEXT Interpretation*):ti,ab,kw AND [mh “Dentistry”] OR [mh “Dentists”] OR [mh “Tooth Diseases”] OR Dental:ti,ab,kw OR Dentist*:ti,ab,kw OR Oral:ti,ab,kw OR Mouth:ti,ab,kw OR Prosthodontist*:ti,ab,kw OR Periodontist*:ti,ab,kw OR Tooth*:ti,ab,kw OR (Carious NEXT Lesion*):ti,ab,kw OR (Carious NEXT Dentin*):ti,ab,kw OR Caries:ti,ab,kw OR Tartar*:ti,ab,kw OR Endodont*:ti,ab,kw OR Orthodon*:ti,ab,kw
Scopus	(“Artificial Intelligence” OR “Computer Reasoning” OR ai OR “Machine Intelligence” OR “Computational Intelligence” OR “Computer Vision System*” OR “Knowledge Acquisition*” OR “Knowledge Representation*”) AND (“Diagnostic imag*” OR “Medical imag*” OR “radiograph*” OR “roentgenograph*” OR X-ray OR X-rays OR X-ray* OR bitewing* OR bite AND wing* OR opg OR orthopantomogra* OR pantomogra* OR panoramic* OR “Computer Assisted Radiographic Image Interpretation*”) AND (dental OR dentist* OR oral OR mouth OR prosthodontist* OR periodontist* OR tooth* OR “Carious Lesion*” OR “Carious Dentin*” OR caries OR tartar* OR endodont* OR orthodon*)

Selection Process

Two reviewers (D.S.H. and M.S.C.) screened all retrieved studies (after deduplication) based on titles and abstracts using Rayyan [19]. Potentially eligible studies were thoroughly evaluated according to the inclusion criteria based on the full text. A third independent reviewer (T.P.C.) resolved disagreements on inclusion/exclusion at each stage of the selection process. References of selected studies were manually checked for additional relevant studies. In cases where multiple publications were available for the same study and reporting on the same sample, only the most recent publication was considered for inclusion.

Data Extraction

Data extraction was performed by the same two reviewers (D.S.H. and M.S.C.) independently using a pre-defined extraction sheet. Discrepancies between reviewers were resolved through discussion, with a third reviewer (T.P.C.) consulted if needed. Data extraction focused on gathering available evidence reporting on the influence of AI on diagnostic performance, including the time taken and treatment decision-making of dentists and dental students. Results reporting on the differences in diagnostic performance with and without AI, as well as differences in treatment decisions, were the focus of data extraction. In addition, results reporting on the influence of AI on cost-effectiveness were also extracted, as they may display advantages or disadvantages that come with the use of AI in dental diagnostics. The goal was to describe the evidence’s scope, gain a deeper understanding of the impact of AI when used in radiographic diagnostics, and shape future research. Four authors were approached twice by email for additional data; however, no responses were received.

Data Items

A data extraction sheet was created and adjusted based on the data from the included studies. The collected information was categorised into three main areas: general information, study characteristics, and outcomes. General information included the author, title, date, country, and publication type. Study characteristics focused on the type of AI tool used, the study’s aim, and the number and kind of participants. Data for the primary and secondary outcome measures were extracted as presented in Table 2.

Table 2.

Data items for outcome measures

Primary outcome measures
Diagnostic performance	Sensitivity, specificity, and diagnostic accuracy with and without AI (if not provided in articles, authors were approached for additional data. Otherwise, other metrics were presented in a narrative summary below and the supplementary files)
Diagnostic time	Average time taken, or average difference with and without AI
Treatment decisions	Proportions of non-invasive and invasive treatment decisions with and without AI
Secondary outcome measures
Cost-effectiveness	Incremental cost-effectiveness ratio in image diagnostics with and without AI.
Advantages and disadvantages	Articles stating AI’s dis-/advantages based on their outcomes

Critical Appraisal

Consistent with the scoping review methodology and PRISMA-ScR guidelines, a formal critical appraisal or risk of bias assessment was not conducted. This review aimed to map the existing literature rather than evaluate study quality. Therefore, all relevant studies were included regardless of methodological quality.

Data Synthesis

Study characteristics and results were presented descriptively to summarise AI’s influence on different aspects of decision-making. Additionally, advantages and disadvantages were presented to identify gaps and biases regarding AI support. Results were presented in a narrative format with tables and figures to summarise key findings such as diagnostic performance metrics and recurring key perspectives in the literature, including statements about AI, opinions, advantages, and disadvantages.

Results

Study Selection

After retrieving the articles from five databases and removing duplicates, 6,636 articles were screened based on title and abstract. Thirty-two records were eligible for full-text screening, of which 20 were excluded. Four selected records were registered protocols, one was a correspondence letter, six articles did not report on AI’s influence on the diagnostic performance of dental professionals, three reported on the performance difference between the AI model and dental professionals, and one reported solely on AI’s performance. Furthermore, five articles did not report on caries as a condition. Disagreement after full-text screening was limited to one article and excluded after consulting the third reviewer (TPC). Two articles were included after the last update on August 14th, 2025. The search results and the study inclusion process are presented in a PRISMA-ScR flowchart (Fig. 1) [18]. The PRISMA-ScR checklist for this study is attached in online supplementary Table 1.

Fig. 1.

Flowchart of the study selection.

Study Characteristics

Finally, thirteen studies were included in this scoping review, spanning publication years from 2020 to July 2024. The studies were randomised controlled trials (n = 6), machine learning studies (n = 4), and cross-sectional studies (n = 3). General dentists, specialised dentists, and dental students served as participants, with numbers ranging from three to seventy-four. Low participant numbers were typically recruited for machine learning studies, as their primary goal was to develop deep learning models and evaluate their applicability. Table 3 presents general information on the included studies. Studies were performed in Europe (n = 8), America (n = 1), and Asia (n = 4). In general, studies considered the dentists’ approaches (n = 8), while four studies were performed with dental students, and one used a mixed-participant approach. Ten studies used bitewings as diagnostic images, and two studies used periapical radiographs and focused on AI’s applicability on photographs. All collected data are available in online supplementary Table 2.

Table 3.

Included articles

Author, year	Country	Participants	Condition and type of image	Software	Study type	Aim	Methods	Results
Liu et al. [20] (2024)	China	3 junior dentists (less than 1 year of experience)	Caries detection in periapical radiographs	ResNet50 + SAM	Machine learning study	Validation of a system for automatic caries detection in periapical radiographs Compare the model to other systems Determine the influence of AI on caries detection by dentists	Model was trained on 4,278 periapical radiographs Two experienced endodontists did annotations The model was then compared to VGG19 and ResNet-50 and to the dentists who annotated the radiographs Three junior dentists assessed 602 radiographs first without then with AI They did this twice with 7 months in between	AI support resulted in an improvement of diagnostic performance
Castner et al. [3] (2024)	Germany	22 dentists	Caries detection in bitewings	dentalXrai Pro 1.0.4	Randomised controlled trial	Compare gaze behaviour during caries detection with and without AI, with a focus on expert interaction with AI support systems	20 bitewings were assessed by 22 dentists (10 with, 10 without AI) Time to first fixation, fixation count, duration, frequency with and without AI were compared	Task time: Longer with AI (109.19 s vs. 87.1 s), not significant Fixation duration: More time on UI with AI (5481.76 ms vs. 2348.39 ms, p = 0.0003) Fixation count: Higher on UI with AI (20.86 vs. 10.63, p < 0.001) Gaze transitions: More transitions between AI findings and bitewing with AI User interactions: AI was used 8.4 times on average, for 24.31% of task time Fixation frequency/duration: No significant difference between AI on/off UI focus: Increased attention to AI elements with support
Arsiwala-Scheppach et al. [21] (2023)	Germany	22 dentists	Caries detection in bitewings	dentalXrai Pro 1.0.4	Randomised Controlled Trial	Compare gaze behaviour during caries detection with and without AI	22 dentists assessed 20 bitewings (10 with, 10 without AI) Time to first fixation, fixation count, duration, frequency with and without AI were compared Dentists verbally reported the presence of caries (enamel/dentin)	Fixation duration: Fixation was longer on teeth with lesions/restorations and shorter on teeth without lesions with AI-support, but differences are not statistically significant Time to first fixation: Shorter median times across all teeth groups with AI support, but differences are not statistically significant Fixation count: Higher fixation counts on teeth with carious lesions/restorations with AI support Non-AI users followed a more systematic search strategy
Ramezanzade et al. [22] (2023)	Denmark	25 dental students	Prediction of pulp exposure in bitewings	ResNet-50	Randomised controlled trial	Evaluate the effect of AI support versus standard radiographic and clinical information on dental students' pulp exposure prediction ability Evaluate the model's performance	The AI model was trained to predict pulp exposure 25 dental students assessed 292 bitewings, divided into 4 different settings Bitewing Bitewing + clinical data Bitewing + AI predicting pulp exposure Bitewing + clinical data + AI predicting pulp exposure	The AI model had a better performance than the participants, but the participants, when given AI support, benefited only slightly
Ayan et al. [16] (2023)	Turkey	40 dental students	Caries detection in bitewings	YOLOv5xs	Cross-sectional study	Determine the caries diagnostic performance of dental students after training with AI Evaluate and compare the model	Two groups of dental students assessed 15 bitewings each (without AI) After 1 week, one group (group 2) received training in caries diagnostics with AI Then both groups assessed 15 bitewings each (without AI)	Training with AI resulted in an improvement in diagnostic performance Training with AI resulted in increased labelling time
Portella et al. [23] (2023)	Brazil	1 undergraduate student, 1 newly graduated dentist, 1 paediatric dentist	Caries detection in photographs	VGG-19	Machine learning study	Evaluate the effectiveness of a CNN in detecting healthy teeth and early carious lesions on occlusal surfaces using photographs Determine the influence of AI on caries detection by dentists	Phase I –2,481 posterior teeth (2,459 permanent, 22 deciduous were classified via ICDAS –ICDAS 0 and 2 surfaces photographed –CNN performance compared to the reference examiner –Three participants assessed images without a CNN Phase II: Participants reassessed photographs using CNN-generated algorithms after 7 days	AI support resulted in an improvement in diagnostic performance
Schropp et al. [24] (2023)	Denmark, UK	74 dental students	Caries detection in bitewings	AssistDent	Randomised controlled trial	Evaluate the educational impact of AI on students’ ability to detect enamel-only proximal caries in bitewings Evaluate whether proximal tooth overlap interferes with caries detection	Students assessed bitewings in two rounds, 25 in each round Control group without AI in both rounds Experimental group with AI in the first, without AI in the second round	AI support did not influence dental students’ ability to detect proximal enamel caries, but positive agreement improved over time Proximal tooth overlap interfered with caries detection
Schwendicke et al. [23] (2022)	Germany	23 dentists	Caries detection in bitewings	dentalXrai Pro 1.0.4	Randomised controlled trial	Evaluate cost-effectiveness with and without the use of AI during caries detection using bitewings	23 dentists assessed 20 bitewings 10 with and 10 without AI support The presence of caries was determined, and treatment decisions were made The proportion of true and false positives and negatives detected was counted Treatment decisions were assigned to each detection Costs were assessed based on German public/private fee catalogues Markov model: Simulated lifetime outcomes for posterior permanent teeth in 31-year-olds Monte Carlo simulations: Used to assess incremental cost-effectiveness at different willingness-to-pay thresholds	Higher accuracy with the support of AI did not lead to higher cost-effectiveness, because more invasive treatment suggestions lead to higher costs
Li et al. [25] (2022)	China	3 junior dentists (3 years of experience)	Caries and apical periodontitis detection on periapical radiographs	ResNet-18	Machine learning study	Develop a DL model to detect dental caries and periapical periodontitis on periapical radiographs automatically Compare the model to other systems Evaluate dentists’ performance with and without the tool	Model was trained on 4,129 periapical radiographs Three senior dentists did annotations Junior dentists assessed the same 400 radiographs first without, and 2 weeks later with the support of a model They repeated this after 2 months Outcomes for caries and apical periodontitis were measured	AI support resulted in an improvement in diagnostic performance
Lee et al. [26] (2021)	Korea	3 dentists	Caries detection in bitewings	U-net	Machine learning study	Develop a DL model to detect dental caries on bitewing radiographs automatically Evaluate the dentist’s performance with and without the model	Model was trained on 304 bitewings Model and three junior dentists then assessed 50 bitewings Junior dentists assessed 50 bitewings, and a few weeks later, revised the decision with the support of the model (time in between not specified)	Model can improve early caries diagnosis, but more training data are needed for stability and precision AI support resulted in improvement of diagnostic performance, especially for initial and moderate caries
Mertens et al. [10] (2021)	Germany	22 dentists	Caries detection in bitewings	dentalXrai Pro 1.0.4	Randomised controlled trial	Evaluate diagnosis and treatment suggestions during caries detection with and without AI	22 dentists assessed 20 bitewings 10 with and 10 without AI support The presence of caries was determined, and treatment decisions were made	AI support increased the accuracy for enamel caries AI support increased non-invasive and invasive treatment decisions
Devlin et al. [15] (2021)	UK	23 dentists	Caries detection in bitewings	AssistDent	Cross-sectional study	Evaluate enamel-only proximal caries detection with and without AI	24 bitewings were assessed with enamel-only caries by two groups One group with the support of AI, one group without	AI support increased sensitivity for enamel caries Experimental arm participants were less likely to correctly identify healthy proximal surfaces as non-carious
Devlin et al. [27] (2020)	UK	24 dental students	Caries detection in bitewings	AssistDent	Cross-sectional study (pilot study)	Evaluate enamel-only proximal caries detection with and without AI	24 dental students were divided into a control group and an experimental group Both groups assessed 24 bitewings to detect enamel caries The experimental arm used AssistDent	AI support resulted in dental students' ability to detect enamel caries

UI, user interface.

Results Synthesis

The outcomes of the included studies are summarised in Figure 2. Of the thirteen included studies, ten reported on the impact of AI on diagnostic performance [10, 15, 16, 20, 22–27], and three studies reported on the influence of AI on time during the diagnostic process [16, 21, 28]. Only one study reported on treatment suggestions made with AI [9], and one on cost-effectiveness [29]. Regarding dentition type, only three studies specified whether deciduous and/or permanent teeth were included. One study analysed photographs of 22 deciduous and 2,459 permanent teeth [23], while two studies examined only permanent dentition [21, 25]. The remaining ten studies did not report the type of dentition evaluated. No analytical summary of evidence was performed, as this review aimed to assess the scope of the influence AI can have on dental professionals’ decision-making in image diagnostics.

Fig. 2.

Outcome measures studies reported on regarding AI’s influence on dentists and/or dental students during image diagnostics.

Influence of AI on Diagnostic Performance: Articles Reporting on Accuracy, Sensitivity, and Specificity

In total, n = 9 articles reported on sensitivity and/or specificity and/or accuracy, with different ranges of change between AI-supported and non-supported groups (Fig. 3). Details on individual studies are provided in Table 4, with additional information available in online supplementary Table 3. AI-supported caries detection exhibited increased sensitivity from 60.3% to 76.5%, specificity from 75.17% to 81.0%, and accuracy from 67.5% to 78.8% for one study [25] and sensitivity from 67.2% to 80.7%, specificity from 72.9% to 80.8%, and accuracy from 69.6% to 80.6% for another study [20]. A randomised controlled trial revealed that AI-assisted detection increased dentists’ sensitivity from 72.0% to 81.0%. No significant improvements were observed in specificity and accuracy, but the ROC AUC rose from 85.0% to 89.0%, indicating a marginal benefit in diagnostic performance [10]. Likewise, another study found that AI-supported detection improved sensitivity in diagnosing initial caries among three junior dentists, showing an average increase from 80.1% to 88.3% for all dentists. However, this study did not report specificity and accuracy [26].

Fig. 3.

Range of performance differences between AI-supported and non-AI-supported caries detection, based on nine studies reporting one or more of accuracy, sensitivity, or specificity. Accuracy: median improvement 5.00% (IQR: 0.05–10.52%). Sensitivity: median improvement 14.84% (IQR: 8.8–25.5%). Specificity: median improvement 3.16% (IQR: 0.12–5.96%), with one outlier showing a decrease of −10.90%.

Table 4.

Averages for accuracy, sensitivity, and specificity

Author, year	Participants	Type of image	Accuracy without AI	Accuracy with AI	Sensitivity without AI	Sensitivity with AI	Specificity without AI	Specificity with AI
Liu et al. [20] (2024)	3 junior dentists	Caries detection in periapical radiographs	69.55%	80.58%	67.17%	80.68%	72.90%	80.83%
				↑ 11.03%		↑ 13.51%		↑ 7.93%
Ramezanzade et al. [22] (2023)	25 dental students	Prediction of pulp exposure in bitewings	65.5%	67.50%	55.00%	59.50%	68.5%	69.00%
				↑ 2.00%		↑ 4.50%		↑ 0.50%
Ayan et al. [16] (2023)	40 dental students	Caries detection in bitewings	84.00%	89.00%	52.00%	76.00%	69%	75.00%
				↑ 5.00%		↑ 24.00%		↑ 6.00%
Portella et al. [23] (2023)	1 undergrad. student, 1 novice dentist, 1 paediatric dentist	Caries detection in photographs	70.73%	80.73%	​	​	​	​
				↑ 10.00%
Li et al. [25] (2022)	3 junior dentists	Caries on periapical radiographs	67.50%	78.75%	60.33%	76.50%	75.17%	81.00%
				↑ 11.25%		↑ 16.17%		↑ 5.83%
Lee et al. [26] (2021)	3 dentists	Caries detection in bitewings	​	​	80.10%	88.31%	​	​
						↑ 8.21%
Mertens et al. [10] (2021)	22 dentists	Caries detection in bitewings	93.00%	89.00%	72.00%	81.00%	97.00%	97.00%
				↓ 4.00%		↑ 9.00%		0.00%
Devlin et al. [15] (2021)	23 dentists	Caries detection in bitewings	85.30%	83.40%	44.30%	75.80%	96.30%	85.40%
				↓ 1.90%		↑ 31.50%		↓ 10.90%
Devlin et al. [27] (2020)	24 dental students	Caries detection in bitewings	​	​	50.00%	80.00%	​	​
						↑ 30.00%

One study reported that dentists evaluating enamel-only proximal caries exhibited the most significant differences in sensitivity and specificity. The results showed that AI-supported dentists achieved an average sensitivity of 75.8%, compared to 44.3% in the non-AI group, indicating a 31% improvement in detecting early carious lesions. However, the AI-supported group demonstrated lower specificity (85.4%) compared to the control group (96.3%), suggesting that AI users were 11% less likely to accurately identify healthy surfaces as non-carious [15]. A previously conducted pilot study that involved 24 dental students showed a comparable increase of 30% in sensitivity, and the mean number of false positives per image was slightly higher in the experimental group (2.64) compared to the control group (2.46) [27]. In a different study, dental students were trained by an AI-supported programme for caries detection using bitewing radiographs. An increase in sensitivity (from 52.0% to 76.0%), accuracy (from 84.0% to 89.0%), and specificity (from 69.0% to 75.0%) was observed in the AI-trained group. However, the untrained group showed higher specificity (82.0%) than the AI-trained group (75.0%) [16]. Meanwhile, another study used photographs of healthy teeth and early carious lesions on occlusal surfaces. Three participants assessed the images without AI assistance, and their accuracy was recorded as 54.0% for the undergraduate student, 77.0% for the newly graduated dentist, and 81% for the paediatric dentist. After 7 days, the participants reassessed the images with AI support, improving accuracy: 68.0% for the undergraduate student, 88.0% for the newly graduated dentist, and 86.0% for the paediatric dentist. However, this study does not report specificity or detailed classification metrics, such as false positives and false negatives, limiting a more comprehensive assessment of the participants’ diagnostic performance [23].

No significant differences in performance metrics were reported in a study that aimed to evaluate the effect of AI support on dental students’ ability to predict pulp exposure. A total of 25 dental students assessed 292 bitewing radiographs with or without AI-supported predictions. The results indicated a minor, non-significant benefit from AI support: accuracy increased from 65.0% to 68.0%, sensitivity improved from 53.0% to 60.0%, and specificity fluctuated between 69.0% and 70.0% [22].

To summarise, sensitivity shows the most significant improvement. The median increase is 9%, with half of the cases experiencing an improvement ranging from 7% to 21.73%. In two studies, AI enhances sensitivity by over 30%, indicating that AI support effectively detected enamel caries [15, 27]. Regarding accuracy, most cases fall within a 1.42–10.77% increase range, although in one study, it decreases by approximately 4% [10]. Specificity, however, sees little to no change. The median difference is 1%, with most cases falling within a small range of −2% to 5%. Notably, one study reported a decrease in specificity by 10.9% [15] (Fig. 3).

Influence of AI on Diagnostic Performance: Article Reporting on Other Metrics

The previously mentioned AssistDent tool was also utilised in a randomised controlled trial to evaluate the impact of AI assistance on dental students’ ability to detect enamel-only proximal caries in bitewing radiographs. Seventy-four dental students evaluated bitewings in two rounds of 25 bitewings each. One group used the AI software in the first round but not in the second, while the other group assessed both rounds without AI. In the first session, the positive agreement was similar between groups (41.6% for the AI-supported group vs. 48.4% for the non-AI group). However, negative agreement was significantly higher in the AI-supported group (86.0% vs. 80.1%). In the second session, no significant differences were found in positive agreement (57.7% vs. 55.3%) or negative agreement (85.9% vs. 89.4%) [24].

Influence of AI on Treatment Decisions

Notably, only one study reported on the effect of treatment decisions with AI. The study evaluated the impact of AI in caries detection and treatment decisions. The results indicated that AI did not increase treatment intensity for sound surfaces but affected decisions regarding enamel and early dentin caries. For enamel caries, there was a significant shift towards non-invasive therapy (32–36%) and invasive treatment (17–24%). In the case of early dentin caries, AI support increased invasive therapies (from 55% to 66%), whereas no significant changes were observed for advanced dentin lesions. The study concluded that AI-supported diagnosis increased in both non-invasive and invasive treatment decisions [10].

Influence of AI on Gaze Patterns and Time during Diagnostics

Two studies explored the impact of AI on dentists’ gaze behaviour and time taken during caries detection. The study involved twenty-two dentists from a dental hospital who assessed eighteen bitewings with and without AI support. Various gaze metrics were compared, including time to first fixation, fixation count, duration, and frequency. Results showed that with AI, dentists had a significantly higher median fixation count on teeth with lesions or restorations (163 vs. 138) and on teeth with restorations (68 vs. 47). The median fixation duration was also longer on teeth with caries (412 ms) and restorations (292 ms). While AI users fixated longer on teeth with lesions or restorations and shorter on teeth without them, these differences were not statistically significant. Similarly, AI users had shorter median times to first fixation across all tooth groups, but these differences were also not significant. AI support resulted in higher fixation counts on teeth with carious lesions and restorations, whereas non-AI users followed a more systematic search strategy [21]. Meanwhile, the other study presented no significant changes in fixation duration or frequency, suggesting no substantial alteration in gaze behaviour and a slight, non-significant increase in time taken to analyse a bitewing [28]. In contrast, results from a different study reported a substantial rise in labelling time for the AI group from 6.79 min in the pretest to 14.68 min in the post-test. In comparison, there was no significant difference in the non-AI group (pretest: 7.21 min, post-test: 7.32 min) [16].

Cost-Effectiveness

To evaluate the cost-effectiveness of AI-assisted caries detection, a study was conducted, which referred to the randomised controlled trial by Mertens et al. (2021). True and false positives and negatives were recorded, and costs were estimated based on German public and private fee catalogues. Long-term outcomes were modelled using a Markov model, simulating tooth retention in 31-year-olds, and Monte Carlo simulations to assess cost-effectiveness across different willingness-to-pay thresholds. While AI-assisted detection demonstrated higher sensitivity, it also led to more frequent invasive treatments, resulting in higher costs. Ultimately, cost-effectiveness remained comparable between AI and non-AI approaches, with an average tooth retention of 49 years and a mean cost of EUR 330 per tooth (AI: EUR 250–409; no AI: EUR 248–410). It was concluded that although AI improved diagnostic accuracy, it did not lead to enhanced cost-effectiveness due to an increase in invasive treatment decisions [29].

Advantages and Disadvantages

Several points have been identified as potential advantages and disadvantages (Table 5). Some authors, for example, concluded that AI improves sensitivity [10, 16, 22, 23, 25, 26], particularly for small lesions and enamel-only caries [10, 15, 27]. Additionally, articles described AI as a valuable educational tool [16, 22], especially for inexperienced dental professionals. Furthermore, AI was linked with more effective image analysis [21]. While some studies suggested improvement in diagnostic performance, the results of others indicate that AI’s impact on the decision-making process remains unclear [22, 24], and its use does not necessarily translate to greater cost-effectiveness [29]. Several articles also express concerns about the limitations of AI. One reported issue was low specificity, which can lead to misdiagnoses [10, 15, 22, 24, 27], particularly in the early stages of caries diagnosis [10, 15] and in images involving overlapping approximal surfaces [22, 24].

Table 5.

Advantages and disadvantages mentioned by authors when AI was used during image diagnostics

Author, year	Positive observations (advantages)	Neutral observations	Negative observations (disadvantages)
Liu et al. [20] (2024)	Improves diagnostic performance, improves diagnostic consistency	​	​
Castner et al. [3] (2024)	Dentists used AI as a second reviewer	There is a slight (non-significant) increase in time taken to analyse a bitewing	Potential for fatigue due to increased time taken on task and disruption of viewing patterns
	Improves sensitivity	No change in fixation duration/frequency suggests that gaze behaviour does not change with AI
Arsiwala-Scheppach et al. [21] (2023)	More effective/efficient image examination (less time taken to notice, and more focus on caries and restorations)	​	​
Ramezanzade et al. [22] (2023)	Improves diagnostic accuracy, reduces the risk of overtreatment due to higher specificity	There is a slight improvement in diagnostic performance	Low specificity due to lack of trust in AI’s output
	helpful educational tool (statements based on AI model performance)	Unclear whether AI provides improvements in the decision-making process of dental professionals	Risk of misdiagnosis, due to incorrect interpretation of overlapping approximal surfaces, and artefacts
Ayan et al. [16] (2023)	Improves diagnostic performance, helpful educational tool	​	Decreased time efficiency
	prevents possible misdiagnosis (“The use of AI […] will also prevent possible misdiagnosis”)
Portella et al. [23] (2023)	Improves diagnostic accuracy	Using photography limits the conditions that can be detected	Inability to determine caries activity
	Helps with early caries detection, improves diagnostic consistency
Schropp et al. [24] (2023)	Potential to prevent unnecessary treatment	Training with AI did not improve caries detection significantly	Misdiagnosis due to AI’s struggle with overlapping surfaces
Li et al. [25] (2022)	Improves diagnostic performance due to higher sensitivity and specificity, improves diagnostic consistency	​	​
	prevents possible misdiagnosis
Schwendicke et al. [29] (2022)	Improves diagnostic performance due to increased sensitivity (Mertens et al.)	Higher diagnostic accuracy did not lead to higher cost-effectiveness	Risk of overtreatment due to increased non-/invasive treatment decisions (Mertens et al.)
Lee et al. [26] (2021)	Improves diagnostic performance due to increased sensitivity for initial and moderate caries	​	Risk of overtreatment due to false positives (False positives occurred in overlapping approximal surfaces)
			Risk of overreliance
Mertens et al. [10] (2021)	Improves diagnostic performance due to increased sensitivity	User-specific benefits and risks	Risk of overtreatment, due to increased non-/invasive treatment decisions
Devlin et al. [15] (2021)	Improves sensitivity in detecting enamel-only caries	​	Risk of unnecessary preventive treatment due to decreased specificity
Devlin et al. [27] (2020) (Pilot study)	Improves sensitivity in detecting enamel-only caries	No change in specificity due to a similar number of false positives	​
	Potentially leads to higher cost-effectiveness
	improves diagnostic consistency

Bold indicates it is based on individual study results.

Discussion

General Interpretation

This scoping review provides an overview of the role of AI in assisting dental professionals with image-based caries detection and treatment planning. Compared to previous reviews that primarily highlight the technical performance of AI algorithms, this study examines AI’s influence on clinicians’ decision-making by comparing diagnostic accuracy, efficiency, and treatment decisions with and without AI. Based on this review:

• 1.AI was found to enhance sensitivity for detecting caries, particularly for small lesions and enamel-only caries. Still, its effect on specificity varies across studies, raising concerns about possible misdiagnoses.
• 2.AI-supported diagnoses promoted reduced assessment time for experienced professionals but may increase the time for less-experienced dentists. Nevertheless, it is considered a valuable educational tool.
• 3.AI’s influence on treatment decisions after caries detection was discussed in only one study [10], making it difficult to draw conclusions. However, evidence of shifting towards more non-invasive and invasive interventions was revealed, suggesting that AI influences treatment recommendations significantly.
• 4.AI’s cost-effectiveness is also unclear due to only one study conducted within a single healthcare system reporting on the matter, which found that greater sensitivity does not necessarily translate into better economic outcomes. Therefore, the long-term cost-benefit ratio of AI in dental diagnostics requires further research across diverse healthcare systems to determine whether AI contributes to more efficient dental care or increases overall treatment costs.

To summarise, this review shows that AI promises improvement in caries detection, efficiency, and education in dentistry. However, its variable effects on specificity and a lack of studies focussing on treatment decisions and cost-effectiveness require more extensive research before conclusions can be made.

Limitations: Evidence

The current body of research on AI’s impact on dental professionals is limited, with a notable lack of studies investigating AI’s influence on treatment decisions and long-term patient outcomes. Additionally, the quality of existing studies varies, with inconsistencies in study designs and sample sizes. AI research has increased exponentially in recent years, with a focus on machine learning studies. To prove the applicability of their developed models, some studies used dental professionals’ diagnostic performance with and without AI, with as few as three participants, limiting their generalisability regarding AI’s influence on the decision-making process [20, 23, 25, 26]. Studies focussing on diagnostic performance also show inconsistencies, with some failing to report specificity [23, 24, 26, 27], an indicator of whether AI support leads to correctly identifying healthy surfaces as healthy. Without measuring specificity, it remains unclear whether AI contributes to unnecessary treatments. As mentioned before, no conclusions can be drawn from the existing literature regarding the use of AI for treatment decisions, as only one study is available. However, their results indicate that using AI increases non-invasive and invasive treatment decisions [10]. Moreover, most included studies did not specify the dentition type, making it difficult to assess whether diagnostic performance and treatment decisions differ across patient groups. Future research should examine whether AI-supported decisions maintain consistency across populations such as children and older adults, as lesion progression and diagnostic characteristics vary between deciduous and permanent dentitions. In addition, AI models may inherit biases from non-representative training datasets, leading to reduced generalisability across diverse populations. Age-related physiological differences and variations in disease presentation further highlight the need for inclusive datasets and bias-mitigation strategies to ensure both diagnostic consistency and equity in care.

Above all, the absence of standardised research protocols and longitudinal studies makes it difficult to assess the reproducibility of studies and the long-term effects of AI-supported diagnosis and treatment planning on patient outcomes. Current research is primarily cross-sectional, providing a snapshot of AI’s impact rather than insights into benefits or challenges over time. Additionally, among the included studies, only a few provided details regarding the participants’ level of experience. Some reported whether participants were dental students or practising dentists, and a limited number specified further distinctions, such as junior dentists with less than 1 year of experience [20], three junior dentists with approximately 3 years of experience [25], or a mixed sample including one undergraduate student, one newly graduated dentist, and one paediatric dentist [23]. None of the studies stratified their outcomes by years since graduation or academic level within the undergraduate curriculum. Similarly, no studies performed statistical adjustments for clustering or repeated measures that could arise from multiple assessments per participant. Calibration or specific training of participants before the diagnostic tasks was rarely mentioned, except for one study that explicitly described an AI-assisted training session [16]. This lack of consistent reporting on participant experience, calibration, and study design characteristics limits the interpretability and generalisability of findings. Future research should include stratified analyses and transparent reporting of participant expertise and training procedures to better understand how AI support affects different practitioner groups. Addressing the above-mentioned research gaps should be a priority in future research, particularly through longitudinal studies assessing the long-term effects of AI on clinical outcomes. Beyond technical performance, the trust and acceptance of dental professionals and patients in AI will also be crucial for successful implementation. Even if AI enhances diagnostic accuracy, clinicians’ willingness to integrate these tools into daily practice may depend on their perceived reliability, ease of use, and alignment with clinical workflows. Again, little about these aspects is addressed in the included studies.

Limitations: Review Process

The review process itself has some limitations. A formal quality assessment of the included studies was not conducted due to the heterogeneity in study designs, methodologies, and reporting standards. The vast differences in research designs made applying a uniform quality appraisal framework difficult. Additionally, while no language restrictions were applied during screening, all eligible studies were published in English. Potentially relevant non-English evidence may exist but was not captured. Another limitation concerns the inclusion of caries as the only condition. Results may differ for other diagnostic tasks and should not be generalised beyond caries detection. Future research can build on this work to explore AI’s impact across other areas of dentistry. Many machine learning studies were excluded because they did not report on AI’s impact on decision-making in a clinical setting or provide a direct assessment of dental professionals’ diagnosis performance when supported by AI compared to the conventional image analysis (without AI support). Many of these studies focused on AI development and model quality, which fell outside the scope of this review. While these exclusions were necessary for maintaining the review’s focus on dental professionals’ decision-making, they may have caused some insights on diagnostic performance to be left out. However, given the strong emphasis in the included studies regarding AI’s impact on diagnostic performance, it is unlikely that missing these studies would have changed the overall conclusions. Finally, the scope of this review was limited to studies published in specific databases until August 14th, 2025. The speed at which AI research is published may have led to relevant studies not being included. Even though no language restriction was imposed, non-English publications may have been overlooked due to the limited inclusion of databases.

Implications

For research: AI has been a highly relevant topic recently, highlighting that there is space to evaluate and improve research methods critically. Future research should expand beyond diagnostic performance to investigate AI’s role in decision-making and treatment decisions. It is crucial to come closer to an answer as to whether AI support causes overreliance, misdiagnosis, or overtreatment and whether these effects vary across different dental pathologies. Therefore, longitudinal studies in a clinical setting with long-term follow-ups and standardised protocols are necessary to evaluate AI’s influence on the decision-making process and, as a result, on patient care. Moreover, future studies should explore the impact of a larger adoption of AI-supported tools on oral healthcare improvement and the possible costs for healthcare systems. It is also important to include patients’ perspectives on AI-assisted diagnosis and treatment planning, addressing their trust, understanding, and acceptance of AI involvement in their care. Patient-centred outcomes will be essential to evaluate the true benefit of AI integration in practice.

For clinical practice: Integrating AI into dental diagnostics offers benefits and challenges. AI can improve diagnostic accuracy and consistency. However, concerns about overreliance, misdiagnosis, and cost-effectiveness must be addressed more thoroughly before the large-scale implementation. Dental professionals must view AI as a complementary tool rather than a replacement for clinical judgement or justification for treatment decisions. Additionally, before implementing AI tools, training programmes could help provide users with the necessary skills and address the risks of overdiagnosis and overtreatment, which are the likely outcomes of finding more lesions and oral conditions at early stages. The international clinical practice guidelines on when and how to intervene in any oral condition should be reemphasised in those training programmes and dental education. Moreover, training programmes should include education on the potential biases embedded in AI tools and strategies to critically appraise AI suggestions, empowering professionals to maintain autonomy in clinical decision-making.

For policymakers: Statements have been made by the World Dental Federation regarding the benefits and challenges of AI in dentistry, including recommendations for researchers, educators, model developers, and policymakers [30]. In statements such as these, ethical considerations must not be overlooked. Data privacy, algorithm transparency, and accountability in case of errors are key to AI integration. Therefore, ethical frameworks should guide the development of AI tools and their implementation in clinical settings. Clear responsibilities must be defined among developers and clinicians to ensure patients’ rights and safety. Moreover, it is self-evident that guidelines and policies must be updated as research comes to new findings to provide proper baselines for effective AI implementation into dental practice.

Conclusion

To answer the question “How does AI influence clinical decision-making in dental imaging?” this scoping review highlights that AI-supported diagnostic tools in dentistry show promise in improving sensitivity and efficiency, particularly for detecting early-stage lesions. However, the variability in specificity, the limited evidence regarding its influence on treatment decision-making, and the lack of long-term evaluations call for caution. While initial results suggest that AI may be a valuable adjunct to clinical judgement, its full integration into dental practice requires addressing unresolved issues around bias, clinician trust, patient outcomes, and cost-effectiveness. Therefore, it is crucial to understand how AI impacts clinical decision-making in dentistry over time, beyond initial diagnostic accuracy. Clinical studies should capture how dentists adapt to AI, whether they become overreliant or resist its use, and how these changes affect long-term patient outcomes stratified over different patient populations. By following decisions and outcomes over time, researchers can assess whether AI leads to genuinely better care or unintended consequences like overtreatment. Hence, more evidence needs to be generated that can subsequently guide AI’s safe and cost-effective integration and take full advantage of its potential.

Future studies should therefore prioritise the following:

• 1.Standardised outcome measures, including diagnostic performance metrics (such as sensitivity, specificity, accuracy, and AUC), decision-making indicators (such as diagnostic confidence, treatment choice, and decision time), and clinical outcomes (such as lesion progression or treatment success), to enable meaningful comparisons between studies.
• 2.Clinical study designs incorporating factors such as age, general health, dietary habits, oral hygiene status, and caries activity, with long-term follow-up to better evaluate clinical decision-making, specifically treatment decisions and subsequently patient outcomes.
• 3.Stratified analyses, for example, by clinical experience of participants and dentition type used in datasets, to clarify how AI affects different practitioners and patient groups.
• 4.Assessing the influence of dataset diversity and bias mitigation strategies to ensure equity and generalisability.
• 5.Investigating practice implementation, such as clinician acceptance, workflow integration, and cost-effectiveness.

By addressing these areas, the focus can shift from “Can AI detect?” to “How does AI change what we do?” and assess the real-world impact on oral healthcare delivery.

Acknowledgments

We acknowledge Marloes Ijff from the Radboud Medical Library for her assistance in developing the search strategy for this review.

Conflict of Interest Statement

Dr. Maximiliano S. Cenci was a member of the journal’s Editorial Board at the time of submission. Dilek Sezen-Hulsmans and Rutger E. Sonneveld are employed by Prime Dental Alliance and Radboudumc and carried out this work as part of a collaboration between the institutions. The other authors have no conflicts of interest to declare.

Funding Sources

No funding was obtained to support this review.

Author Contributions

Dilek Sezen-Hulsmans: writing – original draft, review, and editing, project administration, methodology, investigation, formal analysis, visualisation, data curation, and conceptualisation. Tatiana Pereira-Cenci: methodology, conceptualisation, and writing – review and editing. Rutger E. Sonneveld and Bas A. C. Loomans: writing – review and editing and supervision. Maximiliano S. Cenci: writing – original draft, review, and editing, supervision, project administration, methodology, investigation, formal analysis, data curation, and onceptualisation.

Funding Statement

No funding was obtained to support this review.

Data Availability Statement

All data generated or analysed during this study are included in this article and its online supplementary material. Further enquiries can be directed at the corresponding author.