Abstract
The most common and widely known type of head and neck cancer is the Oral or mouth neoplasm, of which Oral Cavity Squamous Cell Carcinoma (OCSCC) is the most popular. Despite its impact on Mortality, it is always diagnosed at a late stage due to the inefficiency of the screening models at the early detection stage. Early detection of OCSCC has more than 83% survival rate, although the rate of early detection currently is 29%. Partnering with OCSCC early detection, the deep learning model aids in detecting patterns of oral cancer cells. Sequel to that, this paper proposes using ensemble pre-trained deep learning models while unifying the ensemble heads with more shared layers for the early detection of OCSCC from microscopic images. Various pre-trained deep learning models are evaluated using transfer learning while using the Augmentor library to establish high-quality microscopic oral cancer image datasets. The proposed approach obtained a 0.1-0.6% improvement compared with transfer learning methods using 100x magnification and 400x magnification, thus illustrating the robustness of the model for low-quality and high-quality images. Noting that the dataset used in this paper is a newly released competition dataset, a comparison was made with only the article that used the same data when writing this paper. The result obtained proves that the proposed methodology is a promising method for detecting and classifying OCSCC.