Evaluation of normalization technique on classification with deep learning features

Abstract

Cancer is one of the diseases with the highest mortality rate in the world. Dysplasia is a difficult-to-diagnose precancerous lesion, which may not have a good Hematoxylin and Eosin (H&E) stain ratio, making it difficult for the histology specialist to diagnose. In this work, a method for normalizing H&E stains in histological images was investigated. This method uses a generative neural network based on a U-net for image generation and a PatchGAN architecture for information discrimination. Then, the normalized histological images were employed in classification algorithms to investigate the detection of the level of dysplasia present in the histological tissue of the oral cavity. The CNN models as well as hybrid models based on learning features and machine learning algorithms were evaluated. The employment of the ResNet-50 architecture and the Random Forest algorithm provided results with an accuracy rate around 97% for the images normalized with the investigated method.