Detection of Pre-Cancerous Conditions in Oral Cancer Progression

Abstract

Cancer is one of the main causes of deaths globally, responsible for around 10 million deaths in 2020. This is attributable to the predominance of patients detected at late stages. mouth cancer is the most frequently diagnosed cancer in the world. The increased death rate is mostly related to late-stage diagnosis; prompt intervention on the identification of oral cancer or pre-cancerous lesions is expected to give improved results and reduce treatment expenses. The currently available diagnostic approaches include clinical examination, imaging, and the Light Gradient Boosting Machine (LightGBM) classifier. This research presents a unique method to differentiate among malignant and benign oral cancers and to detect their pre-cancerous stages. The proposed methodology introduces a hybrid DL model that incorporates MobileNetV2 along with a Support Vector Machine (SVM) classifier, and by using Histogram of Oriented Gradients (HOG) feature extraction. This extension utilizes deep feature representations acquired through MobileNetV2, by using HOG descriptors to identify edge and structural data pertinent to oral lesion patterns, in contrast with the original LightGBM-based framework that depends on manually extracted colour and texture features. The deep-HOG vector of features is used as input to an SVM-based classifier to improve the performance, particularly in multi-class situations. The study shows outstanding performance, more than the current state-of-the-art techniques. The proposed method provides an efficient and effective solution for the categorization of oral cancer.