Artificial Intelligence-Based Multi-Model for Risk Assess ment of Oral Potentially Malignant Disorders Using Con focal Imaging of Exfoliative Cytology

Abstract

Early detection of oral potentially malignant lesions (OPMLs) is crucial for improving prognosis and survival rates. Traditional diagnostic methods, including biopsies, are invasive and often delay treatment. This study introduces an innovative, non-invasive artificial intelligence (AI) system that integrates three machine learning models to evaluate patients’ clinical data, the staining patterns of toluidine blue-stained images of lesions, and confocal microscopic imaging of exfoliative cytology samples. The system was deployed using a user-friendly web interface to provide risk assessment for OPMLs. Model testing on a pilot dataset (n=50) has demonstrated the following accuracy: 91% (clinical model), 94% (microscopic image-based model), 74.32% (stained image-based model), and 98% agreement between the system output and experts’ risk assessment. This AI- driven system holds promise as a diagnostic tool for OPMLs. Further validation of this tool using larger datasets is needed.