Performance evaluation and clinical application exploration of a ViT-CNN ensemble model for multiclass oral mucosal disease classification: a pilot retrospective analysis based on public datasets

Objective To assess the performance of a Vision Transformer (ViT)-based deep learning model in classifying oral mucosal diseases (OMD) and to explore the value of the integrated model for clinical support.Design A pilot study that combines publicly available datasets with integrated modelling.Methods Developed an EfficientNet-B0 convolutional neural network and a ViT-B16 model, and tested three integration strategies: average method, weighted method and geometric average method. Evaluation metrics included accuracy, F1 score and inference speed. Diagnostic subject performance was compared between general dentists and oral specialists to validate model efficacy.Results The integrated model outperformed individual models, with the geometric average integration method achieving an accuracy of 94.32%. When used by general dentists, the diagnostic time was reduced from 241 s to 112.4 s, with an accuracy rate of 93%.Conclusions The ViT-enhanced integrated system can improve the classification efficiency of OMD and provide support for non-specialist doctors. However, validation with larger datasets is needed in future research.