A Comparative Study of Supervised and Unsupervised Neural Networks for Oral Lesion Detection

Abstract

The main aim of the research presented is to compare supervised and unsupervised learning for oral lesion detection. Previous work has essentially focused on supervised neural networks. The study firstly compares the performance of ResNet-50, VGG16, SCAN and FOC on a publicly available benchmark set for oral lesion detection. The trained approaches are then evaluated on a generalization set comprised of realworld data. ResNet-50 provides the highest classification accuracy at 92.43%. The four techniques have also had a generalization set applied, with Resnet-50% providing the best classification accuracy for the generalization set at 82.13%. The unsupervised techniques (SCAN and FOC) have the smallest decrease in accuracy when applied to the generalization set. Future work includes further analysis of SCAN and FOC in terms of their good generalization capabilities as well as potential hybridization of the two techniques