AI-driven prediction of progression to oral squamous cell carcinoma using a multiresolution pathology model

Abstract

Potentially malignant oral lesions are visible mucosal changes with risk of progressing to squamous cell carcinoma. These lesions are currently graded as no, mild, moderate, or severe dysplasia per WHO guidelines. While widely used, this system is subjective, imprecise, and limited in distinguishing low-risk from high-risk lesions. To address this, we developed a multiresolution deep learning model using digital pathology to predict malignant transformation. Trained on 221 digitized whole-slide images (111 progressors, 110 non-progressors), our vision transformer (ViT) model outperformed traditional CNN-based models, achieving 80.0% accuracy, an F1-score of 77.3%, and an AUROC of 0.798. Importantly, AI-predicted progression aligned with histopathologic features such as abnormal keratinization, increased apoptosis, and nuclear changes characteristic of malignancy. While further validation in larger, prospective cohorts is needed, our model demonstrates promise as an adjunctive tool for identifying high-risk lesions, potentially improving risk stratification, and guiding treatment decisions.