Oral leukoplakia (OL) is the most common oral potentially malignant disorder of the oral cavity, demanding reliable diagnosis and risk stratification (1). It usually presents as a white plaque that cannot be scraped off and cannot be attributed to any other definable disease (1). OL affects approximately 2.6% of the population (2) predominantly men over 50 years— with a malignant transformation rate estimated in 9.5% (3).

Clinically, OL is classified as homogeneous or non-homogeneous. Homogeneous lesions appear as uniformly white plaques, with smooth, regular surfaces and sharp margins. In contrast, non-homogeneous OLs display mixed white/red patterns, and are further divided into speckled (alternating white and red areas), nodular (with small, polypoid exophytic nodules), or verrucous (characterized by a wrinkled or corrugated surface) (4). Regardless of the subtype, lesion size may range from well-defined millimeter-sized plaques to extensive patches covering broad areas of the oral cavity (4). Non-homogenous lesions are associated with higher degrees of dysplasia and with a higher risk of transforming into an oral squamous cell carcinoma (OSCC) (5, 6).

The diagnosis of OL is clinical, often mimicking other white or mixed white and red lesions (such as candidiasis, frictional keratosis, or lichen planus). Therefore its clinical diagnosis can be challenging, especially for general dentists or non-oral medicine practitioners (7). A comprehensive evaluation must start by excluding any traumatic or reactive causes, followed by an attempt to gently scrape the lesion to ensure it cannot be removed. If the lesion persists, the adjacent mucosa should be stretched to rule out leukoedema. Once these and other differential diagnoses are excluded, a biopsy must be performed, as histopathological examination is mandatory to provide definitive diagnosis and assess the presence and severity of epithelial dysplasia (mild, moderate or severe) (4).

Even with this systematic diagnostic workflow, the histopathological evaluation of OL has limitations. For example, incisional biopsies of large or multifocal lesions can miss critical dysplastic areas. Histologic similarities with other entities, such as lichen planus, can generate diagnostic doubts (8). Dysplasia grading also remains highly subjective, showing significant inter- and intra-observer variability. With no definitive molecular or immunohistochemical markers to confirm the diagnosis of OL, there is a pressing demand for novel, precise and objective diagnostic approaches (8). In this context, Artificial Intelligence (AI)-driven analysis and decision-support tools hold promise for improving consistency, accuracy, and lesion detection (9).

AI encompasses the theory and creation of computer systems capable of performing tasks with human-like intelligence, including visual perception, speech recognition, decision-making, and language translation (10). Machine Learning (ML) is a subfield of AI that focuses on algorithms that improve their performance through exposure to data. In supervised learning, for example, ML models detect patterns in labeled datasets and apply them to make predictions on new, unseen inputs without explicit programming (11). Deep Learning (DL), on the other hand, extends ML by using computer neural networks (CNN) inspired by the human brain (12). These systems of interconnected nodes or neurons are organized into an input layer, one or more hidden layers, and a final output later. During forward propagation, data passes through these layers to produce an output or prediction, whereas during backpropagation, the neural network is able to adjust its internal weights based on the error between its prediction and the true outcome. This iterative process of propagation and weight adjustment allows the network to self-correct and continually enhance its performance over time (13).

In recent years, numerous studies have investigated the application of ML and DL technologies to support clinicians in diagnosing OL (14); however, a comprehensive synthesis of this evidence is still needed. Therefore, this scoping review aims to systematically synthesize current evidence on the use of artificial intelligence technologies, particularly classification algorithms, for the diagnosis of OL and for the assessment of epithelial dysplasia severity in clinically diagnosed OL lesions.

The electronic search yielded 233 records. After deduplication removal (n = 67), 166 studies were screened based on the predefined eligibility criteria. Of these, 142 were excluded, leaving 24 articles for retrieval and full-text review. Eighteen articles were excluded at this stage for the following reasons: not meeting the “concept” criterion (n = 16), being a narrative review (n = 1), or having been retracted (n = 1). Initially, 6 primary publications were included in this review. However, 4 additional articles were identified during the search update in July 2025, bringing the final total to ten publications. The study selection process is illustrated in the PRISMA-ScR flow diagram (Figure 1).

OL carries a malignant transformation risk ranging from 1% to 40% (26), underscoring the need for accurate diagnosis and risk stratification, which remains challenging due to its non-specific presentation. Current diagnosis relies heavily on exclusion, combined with a “wait-and-see” approach supplemented by biopsy. However, this method has important limitations, including histological overlap with other oral pathologies and the dependence on the pathologist's expertise, which introduces additional variability and subjectivity into the diagnostic process (27).

Recent advancements in computer sciences and AI-based technologies offer an opportunity to overcome these diagnostic barriers, enhancing the precision and reproducibility of OL diagnosis. Both ML and DL models have demonstrated significant value in dentistry, particularly by improving the detection of caries, periodontal disease, apical periodontitis, and salivary gland diseases, among others (28–30). Overall, their diagnostic performance is able to match, and sometimes even surpass, that of experienced specialists, offering a promising complement to traditional diagnostic workflows (28–30).

This scoping review identified ten studies that explored the application of AI in OL diagnosis, most of which were published within the past five years, reflecting both the growing interest and rapid technological evolution of this field. This momentum is further supported by initiatives like the MimoSA UPLOAD and MimoSA ANNOTATE tools, which provide centralized platform for collecting and labeling oral lesion images from worldwide cohorts, thereby accelerating the development of AI algorithms capable to detect high-risk oral potentially malignant disorders (31). Collectively, the included studies spanned multiple regions, including India, China, Brazil, Germany, Poland, and the Netherlands, illustrating broad interdisciplinary and international engagement in the field. However, underrepresentation of other regions and populations may introduce bias into the current evidence landscape and limit global generalizability of results.

Despite the limited number of eligible investigations included in this work (16–25), the findings consistently support the utility of AI-based tools for OL diagnosis. When differentiating OL from NOM, most models achieved moderate-to-high diagnostic performance, with sensitivity, specificity and accuracy values typically exceeding 80%. Comparable results were observed when distinguishing OL from other mimicking conditions, such as OSCC, PVL and other white lesions. Notably, models tended to perform more robustly when identifying clinically complex or advanced lesions, a pattern observed across both spectroscopy- and image-based studies (18, 24). However, this does not necessarily indicate that AI models inherently struggle to recognize subtle presentations of OL per se, but rather limitations in training data diversity, underrepresentation of mild lesions within datasets, or differences in model generalization and pattern recognition capabilities (32, 33). Indeed, advanced disease presentations such as OSCC and PVL often display more distinct anatomical or spectroscopic alterations (34, 35), which AI algorithms can more readily recognize and learn to classify.

To improve performance across the full clinical spectrum of OL, future research should prioritize the inclusion of subtle OL cases during model training and validation. Expanding dataset variability through intentional sampling, multicenter collaborations, and open-access data-sharing initiatives could substantially reduce current biases. Furthermore, applying data augmentation strategies, such as image transformations or contrastive learning with clustering techniques, may help to balance underrepresented categories and enhance model generalizability to less conspicuous lesions (36, 37). Integrating molecular biomarker data, particularly those reflecting early epithelial alterations and subepithelial microenvironmental changes associated with malignant transformation (an aspect still largely overlooked in OL research) (26), could further improve OL detection and support risk stratification in the future. This multimodal strategy makes sense as it mirrors clinical reasoning, where visual, contextual and molecular information collectively inform diagnostic judgements.

Beyond data quantity and diversity, the type of data used also plays a critical role in shaping AI performance. In this review, the included investigations used a variety of data modalities including spectroscopy (16–18, 23), clinical photographs (21, 22, 24, 25) and histopathological images (19, 20). In some cases, these were combined with electronic health records and/or genomic data, which appeared to enhance the model's diagnostic performance (19, 22). For example, Muniz de Lima et al. reported a 30.7% increase in OL identification accuracy when clinical data were integrated with histopathological images, compared with image-only approaches (19). Supporting evidence from other medical fields, points in a similar direction: in dermatology, fusion of dermatoscopic images with clinical data has been shown to outperform unimodal models, especially when using advanced fusion methods such as cross-attention (38). Taken together, these findings highlight multimodal data integration as a promising and likely necessary direction for developing clinically robust AI tools for OL diagnosis in the future.

Diagnostic performance also varied according to the underlying algorithmic architecture. Early spectroscopy-based approaches that relied on traditional ML methods demonstrated promising sensitivity and specificity but often lacked correlation with lesion morphology or histopathological grade (16). Subsequent models combining FTIR and Raman analyses achieved higher class-specific accuracies; however, remained at a proof of concept stage and have yet to progress toward clinical validation (18). In contrast, image- and histopathology-based models employing advanced DL architectures, particularly CNNs and hybrid fusion systems such as E-MOD-Plus frameworks, achieved near-human or even superior diagnostic performance, outperforming general dentists in OL identification and some oral pathologists in epithelial dysplasia grading (20, 24). Nevertheless, to date, only one of these studies has validated its findings in an external population (20), underscoring once more the limited generalizability of current models and the need for further multicenter validation efforts.

Finally, it is important to note that none of the AI systems reviewed have been deployed in clinical settings to date. Most remain in preliminary phases of development, with only one study (20) having conducted external validation. The remaining models rely heavily on internal datasets, limiting their generalizability. This highlights a significant translational gap, making clear that these AI models are not yet suitable for use without expert clinical oversight. Moreover, integrating AI workflows into the diagnosis of OL introduces important ethical and practical challenges that warrant explicit discussion. For instance, high-resolution images may unintentionally expose identifiable facial features raising serious privacy concerns. Algorithmic bias is another concern, as most models have been developed using datasets form specific ethnic populations (primarily Indian, Chinese and German), which may hinder performance across diverse or underserved groups, contributing to diagnostic inequities.

Results from this scoping review underscore the considerable promise of AI-based tools as adjuncts for both the clinical and histopathological diagnosis of OL, with reported sensitivity, specificity and accuracy values frequently exceeding 80%. However, the current body of evidence remains at an early developmental stage, characterized by substantial methodological heterogeneity, small sample sizes, and limited external validation. Crucially, current evidence supports the use of AI only as a decision-support system and does not support its independent clinical use, as these tools must always be interpreted by qualified clinicians. To advance toward clinical translation, future research should prioritize standardized reporting frameworks and the external validation of high-performing algorithms in larger, more diverse, and geographically representative cohorts. Furthermore, efforts to enhance model performance across the full clinical spectrum of OL (particularly in subtle presentations), will be essential to improve diagnosis and timely intervention of OL.