This systematic review aimed to synthesize evidence on artificial intelligence (AI) methods for the detection, classification and segmentation of oral squamous cell carcinoma (OSCC), oral potentially malignant disorders (OPMDs) and oral epithelial dysplasia (OED) across imaging, histopathology, spectroscopic and molecular data modalities.

Diagnostic accuracy studies, spanning from 2021-2026, utilizing deep learning, machine learning classifiers, vision transformers and other AI modalities were included in the analysis. Four major databases (Scopus, MedLine/PubMed, Lilacs, Livivo) were screened for publications. The QUADAS-2 tool was utilized to assess quality.

Out of 272 articles shortlisted for abstract screening, 42 studies met the inclusion criteria. Regarding data modality, histopathological images were the most frequently utilised (35.7%), followed by clinical oral photographs (31%). Convolutional neural network (CNN) architecture was the predominant AI modality in 78.6% of the studies. Classification was the predominant task (57.1% of studies, alone or combined with detection/segmentation). Accuracy was the most commonly reported metric (88.1%), followed by specificity (47.6%), sensitivity (42.9%), AUC (35.7%), and F1-score (33.3%). Risk of bias was predominantly unclear for index-test (95.2% of studies) and flow/timing (83.3%), whereas patient selection was better differentiated (35.7% low risk, 57.1% unclear, 7.1% high) and the reference-standard domain showed the most favourable profile (61.9% low risk).

AI-based modalities hold a promising future for early diagnosis of oral cancer utilizing clinical and histopathology images as curated datasets. A consistent reporting framework and greater methodological rigour are needed to support clinical translation. However, standardised multi-centre studies are required before clinical implementation.