Optimal Deep Learning Convolutional Neural Network-Based Oral Cancer Detection Model

Abstract

Oral cancers continue to pose a serious challenge to global health, particularly in developing regions such as India. Lack of automated systems for early visual screening leads to late diagnoses of many patients and subsequently results in high morbidity rates due to the high rate of late-stage diagnosis. For these reasons, finding more efficient means of early identification of oral cancers is essential; however, current methods of diagnosis either take a long time, are subjective (i.e., determined by a single individual), or require an expert in oral pathology to assess patient examination results. Therefore, this study's primary objective is to propose an advanced deep learning model for the detection of oral cancer using a combined deep convolutional neural network–long short-term memory architecture. The education will utilize the publicly available oral cancer dataset, which consists of various labelled images of oral squamous cell carcinoma. First, various preprocessing approaches, including cropping images and applying Contrast Limited Adaptive Histogram Equalisation to recover image excellence, will be applied to the images before extracting features from each image using a DCNN. Next, the most recent Hybrid Horse Herd Lion Optimisation (HHHLO) approach will be used to select the optimal features for each image. Thereafter, the features will be classified as belonging to a patient with oral cancer or without cancer using the combined DCNN-LSTM model, which effectively combines both spatial and sequential patterns. The model will be built and validated using four performance metrics: recall, accuracy, precision, and F1-score. Compared with previously developed systems, this model demonstrated significantly higher performance (97.5% maximum accuracy). The model presented here offers a reliable and effective means of identifying oral cancers at an early stage and assisting in making informed clinical decisions.