Examination of ResNet and DenseNet Architectures in Early Diagnosis of Oral Cancer: An Evaluation

Oral cancer holds a significant position among head and neck cancers and is encountered quite frequently. Oral cancer, which is the eleventh most common type of cancer worldwide, causes approximately 177,000 deaths and 350,000 new cases every year [1, 2]. The most commonly observed type of oral cancer is Oral Squamous Cell Carcinoma (OSCC) [6] which comprises about 90% of the cases [7]. The survival rate for OSCC is low due to the frequent late diagnosis [19]. This also underscores the importance of early diagnosis. Convolutional neural networks (CNN) are highly preferred for their high performance in early diagnosis. In this study, the early diagnosis of oral cancer has been investigated through the utilization of CNN. Additionally, two models are selected for each of the two different CNN architectures. Classification is carried out with varying hyperparameters on these four models, and the resulting classification accuracies were examined. Furthermore, the two architectures are compared in terms of their performance, highlighting the differences in accuracy and efficiency. The accuracy values for the DenseNet and ResNet architectures in this classification problem are investigated. Models were selected with varying layer depths within each architecture to understand how the number of layers affected classification accuracy. Furthermore, these processes are carried out with different optimizers and epoch numbers, aiming to explore the influence of optimizer choices and varying epoch numbers on classification accuracy. As a result of the study, the highest accuracy rate was measured as 97.01%, achieved using the DenseNet201 architecture with the SGD optimizer.