A Hybrid Ensemble Model Using DenseNet121 and EfficientNetB2 for Improved Oral Cancer Detection
Sadhana Ravishankar, P. Varalakshmi · 2025
A significant global health burden is posed by oral cancer, with over 377,000 new cases and 177,000 deaths reported annually; in this study, early detection in men is facilitated through the analysis of labeled endoscopic images. Solving the complexity in identifying the oral cancer through endoscopic image of affected region. The approach includes developing the model such as CNN, DenseNet121, EfficientNetB2, VGG19, and Swin Transformer that suit perfectly for medical datasets. The results are analyzed with different metrics and Grad-Cam is used to showcase the cancer zone detection. These models achieve high accuracy and precision in oral cancer detection with endoscopic images. Different data augmentation, optimization techniques, and ensembling methods are used to overcome overfitting and underfitting conditions. Models such as EfficientNetB2 with Voting classifier attained the accuracy of 98.82% and DenseNet121 with Stacking classifier gives the accuracy of 97% for oral cancer detection which helps the oncologist to start the early treatment for the betterment of patients.