A comparative study of single-stage and dual-stage classification models for OPMDs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early diagnosis of oral potentially malignant disorders (OPMDs) and oral cancer is critical for reducing mortality.
• Existing public datasets for oral mucosal diseases are limited in size and disease coverage.
• Deep learning shows promise for improving oral image recognition and diagnosis.

Purpose of the Study:

• To address limitations in current oral mucosal disease datasets.
• To develop and evaluate a novel dual-stage multi-classification approach for oral image analysis.
• To compare the performance of dual-stage versus single-stage classification models.

Main Methods:

• A high-quality dataset of 1,348 oral mucosal disease images was created and released.
• Ten pre-trained deep learning models were trained using single-stage and dual-stage classification pathways.
• Model performance was assessed using accuracy, precision, recall, F1-score, and AUC.

Main Results:

• The dual-stage Swin Transformer and DenseNet-169 model achieved 0.9029 accuracy and 0.9735 AUC.
• The best single-stage model (EfficientNet-B0) achieved 0.8710 accuracy and 0.9766 AUC.
• The dual-stage model demonstrated superior performance over single-stage models.

Conclusions:

• A publicly available, high-quality dataset of 1,348 oral mucosal disease images was established.
• A novel dual-stage classification model integrating Swin Transformer and DenseNet-169 was proposed.
• The proposed dual-stage model outperformed conventional single-stage models in key performance metrics.