Prediction of Oral Cancer Using Deep Transfer Learning

Over the years, mouth cancer has become a serious health issue for people all over the world. However, if discovered and treated early, it is very curable. Although there have been several studies on the use of machine learning techniques for the prediction of oral cancer, most of these studies have focused on the analysis of histopathological images or clinical data. Only few studies have investigated the use of non-invasive imaging techniques, such as X-rays and intraoral cameras for the detection and prediction of oral cancer. This is a significant gap in the literature, as the use of non-invasive imaging techniques has the potential to make oral cancer prediction more accessible and affordable. Convolutional Neural Networks (CNN), a machine learning technique made specifically to analyse image data have demonstrated outstanding performance in image recognition and classification tasks, which makes them a potentially effective tool with a view to early oral cancer prediction. This research work focuses on designing a framework for the prediction of oral cancer using CNN architectures. A total of 4946 oral cancer images were obtained from Mendely, the dataset was taken through some data augmentation techniques like rotation, shearing, flipping and brightness adjustment. The augmented data was therefore trained with selected architectures (VGG16, InceptionV3, Mobile-netV2, and Resnet) of CNN for classification. A batch size of 64 and an epoch of 10 were used to train each model. Amongst all the architectural models used, ResNet has the highest accuracy, recall, AUC and F-1 score rates of 0.866, 0.967, 0.872 and 0.913 respectively while VGG16 has the highest precision rate of 0.877. This study shows that ResNet would give an accurate, easier and faster detection of oral cancer using clinical images.