Dual-scale fusion vision transformer model for vocal cord leukoplakia risk stratification: A multicenter study

This study aimed to develop an artificial intelligence (AI) system for accurate three-tier risk stratification of vocal cord leukoplakia (VCL), with particular focus on distinguishing high-risk lesions (severe dysplasia and carcinoma in situ) from low-risk and malignant lesions, while enhancing diagnostic performance among clinicians with limited experience.

This retrospective multicenter study analyzed 8510 laryngoscopic images from 743 patients across three tertiary hospitals in China. A dual-scale Vision Transformer (ViT) architecture was constructed, integrating multi-scale feature analysis with cross-attention fusion mechanisms. The model was rigorously evaluated through internal-external validation and a prospective reader study involving 12 endoscopists of varying expertise.

The AI system demonstrated superior performance in three-tier classification, achieving F1-scores of 0.883 (95% CI: 0.869-0.896) and 0.861 (95% CI: 0.844-0.878) for high-risk lesion identification in internal and external validation, respectively. AI assistance significantly improved junior endoscopists' sensitivity from 67.1% to 81.4% (P < 0.001), effectively narrowing diagnostic performance disparities with senior experts. The system maintained robust generalizability across institutions, with 90.5% accuracy under heterogeneous imaging protocols.

The proposed AI framework provides a clinically effective solution for reliable VCL risk stratification and reduces diagnostic variability between clinicians. Validated across multiple centers, this dual-scale ViT approach establishes a novel paradigm for laryngoscopic diagnosis and holds significant potential to standardize diagnostic workflows in resource-limited settings.