Artificial intelligence in malignant risk assessment of oral lesions: an umbrella review

Early detection of malignant oral lesions is crucial for their effective management. While histopathological examination is the gold standard for diagnosis and treatment planning, it is time-consuming and often inaccessible in resource-limited settings where expert evaluation and rapid diagnosis are essential. Artificial intelligence (AI) approaches offer noninvasive and rapid alternatives. AI models can significantly facilitate early detection, improve prognostic assessment, and support personalized treatment strategies, ultimately enhancing clinical outcomes and advancing precision medicine in oral oncology.

The present umbrella review aimed to assess the effectiveness of AI models in detecting and predicting the malignant potential of oral lesions.

A comprehensive search was conducted in the Cochrane Library, PubMed, Scopus, and Web of Science databases using the keywords "artificial intelligence", "oral lesion", and "systematic review". The eligibility of identified articles was initially assessed based on predefined inclusion and exclusion criteria. Subsequently, the methodological quality of the eligible studies was appraised using the Joanna Briggs Institute (JBI) checklist, and studies meeting the quality standards were included for data extraction. Data extraction was performed independently by two reviewers, and the methodological quality of the included reviews was appraised using the AMSTAR 2 tool.

AI and deep learning models, particularly convolutional neural networks (CNNs) and hybrid architectures, demonstrated excellent diagnostic performance for oral squamous cell carcinoma (OSCC) and oral potentially malignant disorders (OPMDs). Histopathology-based AI models achieved the highest diagnostic performance, with sensitivities of up to 98%, specificities of up to 95%, and diagnostic odds ratios (DORs) reaching 460.83. Clinical photography and optical coherence tomography also showed high diagnostic accuracy, with sensitivities of approximately 91% and 88-90%, respectively, whereas autofluorescence imaging demonstrated comparatively lower performance. AI models achieved diagnostic sensitivity comparable to experienced clinicians while outperforming less experienced practitioners. However, heterogeneity in datasets, imaging modalities, and study methodologies limited direct comparisons across studies.

Despite methodological limitations in most included reviews, AI models, particularly convolutional neural networks applied to histopathology and clinical photography, show promising early results and potential to approach expert-level sensitivity in controlled settings.

Article Registered in the International Prospective Register of Systematic Reviews (PROSPERO) under registration number CRD420251021811.