Exploration of Heterogeneous Activation Functions on Oral histology Images through Convolutional Neural Network Approach

Abstract

In dentistry, Oral Squamous Cell Carcinoma (OSCC) is the most serious and potentially fatal condition. The patient's probability of survival would rise with early diagnosis and medication. Compared to other digital imaging types, histopathological specimens are more helpful in determining the presence of oral malignant images. Haematoxylin and eosin are two stains that can be used to differentiate between cancerous and normal cells. Malignant cells can be distinguished by their appearance, size, spread, and morphological traits. The factors that need to be considered are marginal adhesion, homogeneity of cell shape, and lumpy size. Given the complex and unpredictable nature of the images, OSCC classifying raises serious challenges. A robust strategy is needed to handle these problems and ensure reliable classification. To generate automated systems, the OSCC histopathology images are processed efficiently. To discern between images that are malignant versus those that are not in the histopathology image dataset, a novel Convolutional Neural Network (CNN) model with distinct activation functions is developed. Activation functions are crucial in cancer and non-cancer scenarios when classifying images. The flattened layer executes binary image sorting employing the Sigmoid activation function. The proposed CNN model is enhanced using ReLU, LeakyReLU, Tanh, Swish, PReLU, GeLU, and ELU in three hidden layers of CNN. The swish activation function offers superior categorization with low loss and increased accuracy. The results highlight activation function selection's significance in improving CNN diagnostic accuracy for OSCC and providing vital novel data to develop reliable and effective medical image classification systems.