Artificial Intelligence Helps Diagnose Oral Potentially Malignant Disorders: A Systematic Review and Meta-Analysis

Introduction

Oral potentially malignant disorders (OPMDs) are a group of oral mucosal lesions with an elevated risk of progression to malignancy. These disorders comprise diverse diseases with distinct clinical manifestations, histopathological characteristics, and etiological factors (Warnakulasuriya et al 2021). The most important common feature is that all these lesions can lead to oral cancer. Oral cancer is the sixth most common cancer worldwide, with 377,713 new cases and 177,384 deaths every year (Sarode et al 2020; Sung et al 2021). The 5-y survival rate is 34%, as oral cancer is mainly diagnosed at advanced stages (III to IV), probably due to low compliance of patients from lower socioeconomic status (SES) (Abati et al 2020). Lower SES is also correlated with a higher risk of oral cancer (Conway et al 2015). If oral cancer is diagnosed at earlier stages (I to II), the 5-y survival rate increases to 83% (Silverman et al 2010). However, secondary prevention is the most important; if lesions that can develop into oral cancer are diagnosed and treated, many lives can be saved (Feller and Lemmer 2012; Gelband et al 2015).

Visual examination remains the most widely used and cost-effective method for the initial diagnosis of OPMDs. However, its diagnostic accuracy is limited by the subjectivity of clinical assessment and the examiner’s level of experience, which can result in both overdiagnosis and underdiagnosis (Tripathy et al 2023). Warin et al (2022) demonstrated that general practitioners diagnose OPMDs with relatively low sensitivity (0.68; 95% confidence interval [CI], 0.62 to 0.75). Consequently, adjunctive diagnostic techniques such as brush cytology, chemiluminescence, and toluidine blue staining have been introduced to improve diagnostic objectivity. While these methods offer enhanced discrimination between oral cancer and OPMDs, their utility in the accurate identification of OPMDs alone remains limited. At present, biopsy followed by histopathological evaluation continues to be the gold standard for definitive diagnosis and for assessing malignant transformation potential (Jeddy et al 2017). Nonetheless, the invasive nature of biopsy may cause patient discomfort and anxiety, and its application is constrained in low-resource settings due to limited access to specialized care, infrastructure, and trained personnel (Kharche et al 2024). These limitations underscore the urgent need for the development of a novel, noninvasive, cost-effective, and reliable diagnostic modality for OPMDs that facilitates the preliminary screening of patients who are typically underserved or have limited access to health care services.

Recent advances in artificial intelligence (AI) present promising opportunities for the development of novel diagnostic tools to support clinical decision-making. Notably, progress in AI architectures for analyzing unstructured medical data has accelerated in recent years (Szabó et al 2024). The widespread availability and capability of taking high-resolution images of smartphones enable convenient documentation and monitoring of patients, particularly in community and primary care settings. Integrating AI-based image analysis with intraoral photography represents a potentially valuable diagnostic opportunity, offering a low-cost, user-friendly, and accessible solution, particularly beneficial in low-resource settings in which access to specialist care and adequate health care infrastructure is limited.

The aim of this study was to investigate the diagnostic performance of AI-based decision-making tools on regular intraoral photographs that support clinicians in the diagnosis of OPMDs.

Methods

The present meta-analysis was conducted according to the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) protocols while adhering to the Cochrane Handbook (Cumpston et al 2022). The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO) under registration number CRD42023475408.

Data Synthesis

Two-by-two contingency tables were extracted from the studies, containing true-positive, false-positive, false-negative, and true-negative values. The classical approach in diagnostic meta-analysis is to calculate pooled sensitivity and specificity. Some studies also included images of oral cancer to investigate the classification of oral cancer alongside OPMD and disease-free cases. In these cases, sensitivity and specificity were calculated based on the ability of the model to differentiate between OPMD and non-OPMD. The bivariate model of Reitsma et al and Chu et al was fitted (Reitsma et al 2005; Chu and Cole 2006) This approach takes into account the dependency between sensitivity and specificity. Diagnostic odds ratios (DORs) were also calculated for each study.

Sensitivity and specificity of the included studies were plotted as summary estimates of sensitivity and specificity and the corresponding 95% confidence and prediction regions on a hierarchical summary receiver-operating characteristic (HSROC) plot. The confidence region contains the pooled sensitivity and specificity (more exactly, 1-specificity) in 95% of the cases. The prediction region contains the true sensitivity and specificity (more precisely, 1-specificity) of a new study in 95% of the cases; hence, it provides excellent insight into heterogeneity.

All statistical analyses were performed in R (Team RC. R: A Language and Environment for Statistical Computing; Vienna, Austria R Foundation for Statistical Computing, 2021), and the code was heavily based on the online tool described by Freeman et al (2019). The CIs of proportions were calculated with the Wilson scoring method, whereas the CIs of the likelihood ratios were calculated with the Koopman asymptotic score method (Koopman 1984), both using the R package DescTools (Signorell A. 2017). Figure 1 summarizes the outcomes and calculation methods.

OPEN IN VIEWER

Figure 1.

Calculation method of (a) sensitivity, (b) specificity, (c) positive predictive value (PPV), (d) negative predictive value (NPV), (e) diagnostic odds ratio (DOR), (f) positive likelihood ratio (LR⁺), and (g) negative likelihood ratio (LR^–).

Results

Search and Selection

The systematic search of 5 databases identified 11,819 articles. The selection process is illustrated in the PRISMA flowchart (Fig 2). After removal of duplicates, 8,264 articles remained for title and abstract selection. Of these, 125 were eligible for full-text selection. For 22 articles, only conference abstracts were available and were excluded from the analysis. The full text of 3 articles could not be retrieved. One hundred sixteen articles did not meet the eligibility criteria, with the reasons for exclusion shown in Appendix 6. Finally, 6 articles were eligible for quantitative and qualitative analysis (Jurczyszyn and Kozakiewicz 2019; Lin et al 2021; Sharma et al 2022; Warin et al 2022; Keser et al 2023; Achararit et al 2023). No other eligible articles were found through citation chasing or manual search (Appendix 7). The inter-reviewer agreement assessed by Cohen’s kappa was 0.91 and 1.00 in the title-abstract and full-text selection phases, respectively.

OPEN IN VIEWER

Figure 2.

PRISMA selection flow chart.

Six retrospective articles were included in this meta-analysis. Four of the studies were conducted in Asian countries (China, Thailand, and India), and the remaining 2 studies were conducted in Europe (Poland and Turkey). Binary and multiclass classification models were evenly used in the studies. Binary classification models distinguish between 2 classes, while multiclass classification models categorize data into 3 or more classes. The teaching setting was well described, and all 6 eligible studies provided essential information on how many times it went through the data (epochs), how much data were processed at once (batch size), and how quickly the model adjusted its parameters (learning rate). This information is important for understanding how the model was trained, allowing researchers to reproduce the result and compare performance or improve the model in future work. A total of 697 images were tested with AI-based architectures from a total amount of 3,983 images. Different types of OPMD were tested (Jurczyszyn and Kozakiewicz 2019; Warin et al 2022) in the included articles. Only 1 type was investigated (Achararit et al 2023; Keser et al 2023). The baseline characteristics of the participants recruited in the articles are summarized in the Table 1.

OPEN IN VIEWER

Table 1.

Baseline Characteristics of the Included Articles.

Author, Year	Study Design	Classification Type	OPMD Types	Number of Images					Architecture Name	Architecture Type	Sensitivity (95% CI)	Specificity (95% CI)
				Total n of Images	n of Training Images	n of Validation Images	n of Testing Images	n of Tested OPMD Images
Jurczyszyn and Kozakiewicz, 2019	rs	Binary	Lichen planus; leukoplakia	n/a	n/a	n/a	63	42	—	PNN	0.88 (0. to 0.95)	0.90 (0.70 to 0.97)
Lin et al, 2021	rs	Multiclass	Not specified	1,448	869	269	310	52	HRNet-W18	CNN	0.90 (0.77 to 0.96)	0.99 (0.98 to 1.00)
Sharma et al, 2022	rs	Multiclass	Not specified	329	280	24	25	10	VGG19	CNN	0.78 (0.45 to 0.94)	0.81 (0.57 to 0.93)
									VGG16	CNN	1.00 (0.65to 1.00)	0.78 (0.55 to 0.91)
									MobileNet	CNN	0.56 (0.27 to 0.81)	0.88 (0.65 to 0.97)
									InceptionV3	CNN	0.90 (0.60 to 0.98)	0.60 (0.36 to 0.80)
									ResNet-50	CNN	0.00 (0.00 to 0.30)	1.00 (0.81 to 1.00)
Warin et al, 2022	rs	Multiclass	Leukoplakia; erythroplakia; white striae with mild, moderate, and severe dysplasia; hyperkeratosis; lichen planus	980	588	196	196	60	DenseNet-169	CNN	0.95 (0.90 to 0.98)	0.98 (0.96 to 0.99)
									ResNet-101	CNN	0.89 (0.82 to 0.93)	0.96 (0.93 to 0.98)
Achararit et al, 2023	rs	Binary	Lichen planus	1089	881	98	110	55	Xception	CNN	0.85 (0.74 to 0.92)	0.92 (0.81 to 0.97)
									ResNet-152 V2	CNN	0.91 (0.79 to 0.96)	0.80 (0.69 to 0.88)
									EfficientNetB3	CNN	0.75 (0.64 to 0.84)	0.95 (0.83 to 0.99)
Keser et al, 2023	rs	Binary	Lichen planus	137	109	14	14	7	CranioCatch	CNN	1.00 (0.65 to 1.00)	1.00 (0.65 to 1.00)

95% confidence intervals (95% CIs) shown in the table were calculated and not given values.

CNN, convolutional neural network; OPMD, oral potentially malignant disorder; PNN, probability neural network; rs, retrospective.

DOR

The DOR was calculated based on the 2 × 2 confusion matrix. The DOR also showed high diversity, ranging from 41.19 to 842.33. The same AI-based architectures with the best sensitivity were selected from the studies; the DOR was 218.32 (95% CI, 59.98 to 794.60) (Table 2).

OPEN IN VIEWER

Table 2.

Best-Performing AI-Based Tools in Terms of Sensitivity from the Eligible Articles.

Study	Architecture	Sensitivity	Specificity	PPV	NPV	DOR	LR+	LR−	PREV
Jurczyszyn and Kozakiewicz 2019	PNN	0.88 (0.75 to 0.95)	0.90 (0.70 to 0.97)	0.95*	0.79*	70.30	9.25 (3.02 to 33.34)	0.13 (0.06 to 0.28)	0.67
Lin et al 2021	HRNet-W18	0.90 (0.77 to 0.96)	0.99 (0.98 to 1.00)	0.94*	0.98*	788.50	76.00 (26.17 to 223.81)	0.1 (0.04 to 0.22)	0.14
Achararit et al 2023	ResNet-152 V2	0.91 (0.79 to 0.96)	0.80 (0.69 to 0.88)	0.77*	0.92*	41.19	4.49 (2.84 to 7.49)	0.11 (0.04 to 0.26)	0.42
Warin et al 2022	DenseNet-169	0.95 (0.90 to 0.98)	0.98 (0.96 to 0.99)	0.96*	0.98*	842.33	43.07 (20.07 to 93.7)	0.05 (0.02 to 0.11)	0.31
Sharma et al 2022	VGG16	1.00 (0.65 to 1.00)	0.78 (0.55 to 0.91)	0.64*	1.00*	Inf	4.5 (2.16 to 11.11)	0 (0 to 0.47)	0.28
Keser et al 2023	CranioCatch	1.00 (0.65 to 1.00)	1.00 (0.65 to 1.00)	1.00*	1.00*	Inf	Inf (2.82 to Inf)	0 (0 to 0.35)	0.50
Overall		0.93 (0.88 to 0.95)	0.95 (0.85 to 0.98)	NA	NA	212.39 (56.39 to 800.00)	16.69 (5.72 to 48.68)	0.08 (0.05 to 0.13)	0.0447

*

Calculated with study-level prevalence.

95% CI, 95% confidence intervals; AI, artificial intelligence; DOR, diagnostic odds ratio; Inf, infinity; LR⁺, positive likelihood ratio; LR⁻, negative likelihood ratio; NA, not applicable; NPV, negative predictive value; PPV, positive predictive value.

HSROC

HSROC visualized data heterogeneity for the best-sensitivity AI-based models. The heterogeneity of the data was considered moderate (Fig 3).

OPEN IN VIEWER

Figure 3.

Hierarchical summary receiver operating of the best performing artificial intelligence–based architectures in terms of sensitivity.

Overall sensitivity, specificity, PPV, NPV, DOR, and LR⁺, LR⁻, and HSROC were calculated and visualized for the best-performing AI-based architectures in terms of specificity (Appendices 11 to 13).

Discussion

To the best of the authors’ knowledge, this is the first meta-analysis to systematically evaluate the diagnostic accuracy of AI-based tools specifically for OPMDs using intraoral images. The studies analyzed demonstrated that AI-based diagnostic tools could help diagnose OPMDs in everyday practice. This meta-analysis showed high sensitivity, specificity, NPV, LR⁺, and LR⁻ for the proposed AI-based models in the diagnosis of OMPDs. However, these findings need to be interpreted with caution due to the low number of eligible articles. This finding demonstrates that AI-based diagnostic models could have the potential of distinguish images of OPMDs from normal mucosal lesions, likely benign or malignant. Nevertheless, further studies are needed on this topic to strengthen this finding.

Identifying these lesions may be more complex in clinical practice due to factors such as lesion localization and other types of diseases, including paraneoplastic mucosal diseases, pemphigus vulgaris, or mucous membrane pemphigoid and certain types of Candida infection (do Carmo et al 2014; Binnie et al 2024). One of the critical responsibilities of a dentist is oral cancer screening, a procedure that can prove challenging in the absence of sufficient expertise. As OPMDs can lead to oral squamous cell carcinoma, early diagnosis and adequate treatment of OPMDs can decrease the risk of malignant transformation, thus reducing the number of new cases of oral cancer (Warnakulasuriya 2020).

The systematic search also identified several articles that investigated OPMDs or other oral lesions (González and Quintero Rojas 2021; Tanriver et al 2021) However, these could not be used in the meta-analysis because they boosted their teaching datasets with images gathered from search engines such as www.google.com/imghp. In the authors’ opinion, they cannot be certain that photographs collected from the internet are OPMDs, which can significantly affect the results.

The data analysis of 6 studies emphasized the ability of AI-based diagnostic tools to diagnose OPMDs. However, the diagnostic performance varied widely, with sensitivity ranging from 0.00 (95% CI, 0.00 to 0.30) to 1.00 (95% CI, 0.65 to 1.00) and specificity from 0.60 to 1.00. This variability likely reflects a combination of factors, including differences in AI model architectures, training procedures, dataset quality, and image characteristics. For example, in the article by Sharma et al 2022, ResNet50, a deep and complex convolutional neural network with the highest number of layers and parameters among the models analyzed, showed a sensitivity of 0% (95% CI, 0.00 to 0.30) (Sharma et al 2022). The researchers selected training parameters (50 epochs, batch size of 16, and learning rate of 0.001) for all AI-based models. It is likely that, for ResNet50, the data quality, quantity, or the uniform training settings were suboptimal. Beyond this specific case, broader variability across studies can be attributed to heterogeneity in data sources (eg, single center vs multicenter), inconsistency in image acquisition protocols, and lack of standardized preprocessing and augmentation methods. Furthermore, the authors did not report essential information regarding lesion size, severity, or subclassification, all of which can significantly affect diagnostic performance. Larger or more clinically evident lesions are generally easier to recognize, whereas subtle lesions, such as simple leukoplakia, may be misclassified as benign hyperkeratosis, leading to underdiagnosis. The absence of a unified evaluation framework, including differences in validation strategies and performance metrics, further contributes to this inconsistency and underscores the need for comprehensive benchmarking and reporting standards.

In this meta-analysis, the proposed AI-based architectures demonstrated an exceptionally high NPV of 1.00, suggesting strong performance in correctly identifying disease-free cases. However, PPV was notably low at 0.42, indicating that a substantial proportion of positive predictions were false positives. This limitation has important implications for clinical usability. A low PPV reduces clinicians’ trust in AI outputs and may result in unnecessary patient anxiety, redundant follow-up visits, and inefficient use of health care resources. While false positives might not always lead directly to invasive interventions such as biopsies, they do increase the burden on clinical workflows and may contribute to overdiagnosis. In settings with limited access to specialists, a high false-positive rate could divert attention from higher-risk patients or contribute to alarm fatigue. Nevertheless, this may have a beneficial effect in uncertain cases, as such encounters can motivate patients to adhere to regular follow-up visits and enhance early detection efforts. Therefore, despite the promising safety profile indicated by the high NPV, the clinical utility of these models remains limited without further refinement. Improving PPV through better dataset stratification, inclusion of lesion-level annotations, and advanced model calibration techniques is critical to enhancing diagnostic precision and real-world applicability. Despite these challenges, AI technologies hold significant promise in improving access to specialist care for at-risk populations.

A major limitation in the current literature is the substantial methodological heterogeneity across studies. Key differences exist in dataset composition, image acquisition protocols, preprocessing techniques, and data augmentation methods. Moreover, studies use varied validation strategies, some relying exclusively on internal cross-validation, while others use external test datasets, often without standardized dataset-splitting procedures or consistent performance metrics. The lack of benchmarking frameworks and inconsistent reporting of model parameters, such as hyperparameters, feature selection methods, and overfitting-prevention strategies, further impedes reproducibility and cross-study comparisons. Notably, variability in AI architecture selection, training configurations (eg, batch size, learning rate, number of epochs), and evaluation criteria can significantly influence reported performance. A more systematic discussion and control of these methodological differences are essential to ensure interpretability, comparability, and reproducibility across studies and to establish best practices in the development of AI-based diagnostic tools. These issues collectively pose significant challenges to the generalizability and clinical translation of AI-based diagnostic models for OPMDs (Srivastava et al 2014; Wong et al 2021).

The diagnostic performance and clinical utility of AI-based tools for detecting OPMDs are influenced by several critical limitations, including restricted dataset diversity, risk of overfitting, and inadequate consideration of demographic factors. Many models are trained on data from single institutions or homogeneous populations, limiting their generalizability to diverse real-world settings. Insufficient representation of rare OPMD subtypes, along with small sample sizes and poor dataset stratification, increases the risk of overfitting, leading to diminished performance on external data. Furthermore, important demographic variables such as age, sex, and lifestyle-related risk factors are often overlooked, despite their relevance to lesion characteristics and malignancy risk. These methodological shortcomings can introduce diagnostic uncertainty, algorithmic bias, and decreased clinician confidence. These challenges underscore the urgent need for diverse, well-annotated datasets, rigorous external validation, and clinician-centered AI model development to ensure equitable, accurate, and clinically applicable diagnostic support for OPMDs (Virgens et al 2025).

In addition to these methodological concerns, inconsistency in intraoral image quality presents another critical limitation. Differences in image resolution, illumination, focus, angulation, and anatomical framing introduce significant variability, which may adversely affect model performance and generalizability. Inconsistent image quality, labeling practices, and diagnostic criteria across studies further hinder AI reliability and interpretability, often necessitating reliance on conventional diagnostic methods. However, most studies fail to report standardized criteria for image acquisition or quality control. The absence of such standards compromises diagnostic reliability and hinders reproducibility. Establishing and adhering to image quality benchmarks, such as minimum resolution thresholds, uniform anatomical orientation, and exclusion of artifacts, would enhance dataset consistency and improve model training outcomes. Future work should incorporate objective image quality assessment protocols or validated scoring systems to ensure that only diagnostically acceptable images are used for AI development and evaluation (Virgens et al 2025).

The integration of AI-based diagnostic tools into clinical workflows also raises important ethical considerations. One significant concern is the risk of misdiagnosis, which can occur due to algorithmic bias, poor generalizability, or limitations in training data. Such errors may lead to inappropriate management decisions, potentially compromising patient safety. Moreover, there is a risk that clinicians may become overly reliant on AI outputs, particularly when the decision-making process of the system lacks transparency or interpretability. This overreliance could diminish the role of clinical judgment, especially in ambiguous cases. To ensure responsible implementation, AI systems must be rigorously validated, used as adjunctive tools rather than standalone solutions, and integrated with appropriate clinician training and oversight mechanisms. Addressing these ethical and practical concerns is essential to ensure the safe, equitable, and accountable use of AI in the diagnosis of OPMDs (Kharche et al 2024).

The analyzed articles used binary and multiclass classification AI-based models. Binary classification architectures, which can distinguish between only 2 categories, can be used in screening by general practitioners or even in self-screening applications. In this case, the exact differential diagnosis is less crucial than determining whether the lesion warrants referral to a specialist, as discussed in the article by Figueroa et al (2022). In this context, 2 categories can be predefined, namely, “nonsuspicious” and “suspicious” or “nonreferral” and “referral.”

AI-based models with multiclass classification could be used in special care to aid in solving differential diagnostic problems because these architectures can deal with multiple predefined lesions and conditions. According to the current state of science, biopsy is still the gold standard, and further development of multiclass classification models and clinical trials is needed to improve diagnostic performance.

AI-based tools demonstrate potential not only in the classification of OPMDs but also in predicting their risk of malignant transformation, facilitating oral cancer screening and detection, and supporting clinical decision-making in cancer management (Hegde et al 2022; Umer 2024) In a systematic review and meta-analysis, Sahoo et al (2024) evaluated the diagnostic performance of AI-driven tools in detecting OPMDs and oral cancer. A total of 55 studies were included in the qualitative synthesis, encompassing the use of histopathological slides, intraoral photographs, and optical coherence tomography (OCT) images. Of these, 3 studies focused specifically on the identification of OPMDs—2 using intraoral images and 1 using histological images—reporting pooled sensitivity and specificity values of 0.96 (95% CI, 0.92 to 1.00) and 0.93 (95% CI, 0.90 to 0.96), respectively.

Li et al (2024) conducted a systematic review and meta-analysis to comprehensively evaluate the diagnostic performance of AI-based tools in the detection of OPMDs and oral cancer. The study assessed various imaging modalities, including OCT, autofluorescence, and conventional clinical photographs. The analysis of 17 studies revealed that OCT demonstrated the highest diagnostic accuracy (OR = 154.064; 95% CI, 38.83 to 611.34), followed by clinical images (OR = 61.83; 95% CI, 33.42 to 114.39) and autofluorescence (OR = 27.09; 95% CI, 17.85 to 41.11). However, meta-regression analysis showed no statistically significant differences between these subgroups. Following outlier removal using a graphic display of heterogeneity plot, diagnostic performance rankings shifted, with clinical images exhibiting the highest accuracy (OR = 77.77; 95% CI, 25.83 to 234.15), followed by OCT (OR = 56.99; 95% CI, 24.54 to 132.38) and autofluorescence images (OR = 33.91; 95% CI, 16.32 to 70.44), with no statistical significance (P = 0.637; Li et al 2024).

Warin et al (2022) showed that the sensitivity of the diagnosis of OPMDs differs among general practitioners (0.68; 95% CI, 0.62 to 0.75) and oral and maxillofacial surgeons (0.74; 95% CI, 0.61 to 0.87). There are other possible methods in the evaluation of OPMDs: autofluorescence imaging, reflectance spectroscopy, and vital staining technique. Sharma et al (2021) investigated 63 lesions in 45 patients with the Identafi® 3000 device and toluidine blue staining and compared the results with histological examination. The overall sensitivity and specificity of the violet light were 0.73 (95% CI, 0.59 to 0.86) and 0.46 (95% CI, 0.27 to 0.67), of green–amber light were 0.78 (95% CI, 0.62 to 0.90) and 0.15 (95% CI, 0.04 to 0.35), and of toluidine blue were 0.51 (95% CI, 0.34 to 0.68) and 0.85 (95% CI, 0.65 to 0.96), respectively. A statistically significant association was observed only between the toluidine blue staining and histological results (P = 0.04) (Sharma et al 2021).

Adeoye et al (2021) used a time-to-event AI-based model to examine data from 716 patients with diagnosed oral leukoplakia and oral lichenoid lesions. They used anamnestic, clinical, and pathological data to estimate malignant transformation potential. They highlighted the possibility of applying this AI-based predictive tool in clinical settings, particularly in areas with limited access to specialists.

Research frequently leverages online datasets; however, many are limited in size and derived from commonly used sources such as Kaggle and GitHub. Sahoo et al (2024) revealed that of 55 studies, 26 relied on such databases, underscoring the need for globally integrated data-sharing frameworks to advance AI applications, particularly in complex diseases such as OPMDs and oral cancer, which demand diverse and representative imaging data. High-quality AI development depends on the effective curation of large, well-annotated datasets, with standardized patient selection and image segmentation to reduce variability. Global collaborative efforts, such as The Cancer Imaging Archive, play a crucial role in addressing these challenges by offering access to robust, extensively labeled datasets (Sahoo et al 2024).

The findings from this systematic review and meta-analysis provide qualitative and quantitative evaluations of the current use of AI-based tools in diagnosing OPMDs while highlighting potential directions for future research. Comprehending the capabilities of these AI tools may open new opportunities for early detection and prompt treatment of OPMDs, potentially improving patient outcomes and preventing the development of oral cancer.

Conclusion

Based on the analyzed studies, AI-based diagnostic tools offer a noninvasive approach that may be capable of screening for OPMDs with high negative predictive values. At its current stage of development, AI may help increase doctor–patient interactions within at-risk populations, potentially leading to more timely diagnosis of lesions. Nonetheless, the ultimate clinical decision must always rest with the physician.

Author Contributions

D. Horvath, contributed to conception and design data acquisition, analysis, and interpretation, drafted the manuscript; A. Fekete, contributed to data acquisition and analysis, critically revised the manuscript; P. Martinekova, contributed to conception and design, data analysis, critically revised the manuscript; S. Kiss-Dala, contributed to conception, data analysis and interpretation, drafted and critically revised the manuscript; E. Abram, contributed to data interpretation, critically revised the manuscript; K. Marton, contributed to data acquisition and interpretation, critically revised the manuscript; A. Zsembery, contributed to conception, critically revised the manuscript; P. Hegyi, contributed to conception and design, critically revised the manuscript; A. Brody, contributed to conception and design, data acquisition and analysis, critically revised the manuscript. All authors gave their final approval and agreed to be accountable for all aspects of the work.

Supplemental Material