An ensemble deep learning model with empirical wavelet transform feature for oral cancer histopathological image classification

Abstract

ABSTRACTOral squamous cell carcinoma (OSCC) has become quite prevalent across many countries and poor prognosis is one of the major reasons for the ensuing high mortality rate. It mainly occurs in sites such as tongue, tonsil, oropharynx, gum, floor and other parts of the mouth. For early detection, the widely used approach is biopsy, in which a small portion of the tissue is taken from the mouth and examined under a disinfected and secure microscope. However, these observations do not effortlessly distinguish between normal and cancerous cells. Diagnosis of OSCC is generally done by pathologists who mostly rely on their years of empirical experience from tissue biopsy sections. The possibilities of human errors increase while detecting the cells using microscopy biopsy images physically. With the growth of artificial intelligence, deep learning models have gained immense importance in recent years and have become one of the core technologies in numerous fields including the prediction of lung cancer, breast cancer, oral cancer, and various medical diagnosis. It not only enhances accuracy, but also fastens the image classification process, as a result, lowering human errors and workload. Here, we have made use of a customized deep-learning model for aiding pathologists in better OSCC detection from histopathological images. We accumulated and analyzed a complete set of 696 histopathological oral images, amongst them 80% have been taken in the training set, 10% of the images are included in the validation set, and the rest 10% for testing purposes. In this study, 2D empirical wavelet transform is used to extract features from the images; later an ensemble of two pre-trained models, namely Resnet50 and Densenet201 are used for the classification of images into normal and OSCC classes. The efficacy of the model is assessed and compared in terms of accuracy, sensitivity, specificity, and ROC AUC scores. The simulation results show that the proposed model has achieved an accuracy of 92.00%. Thus, this method may be utilized for assisting in the binary classification of oral histopathological images.