Semi-supervised learning with deep convolutional generative adversarial networks for canine red blood cells morphology classification

Abstract

Information of Red Blood Cell (RBC) morphology, obtained by analysing RBC images, is regularly requested by veterinarians to diagnose anaemic dogs. Machine learning techniques have been exploited to speed up the image classification. Recently, many researchers used deep learning techniques for classification; however, a large quantity of labelled data is necessary to extract performance with them. A lack of annotated data, due to time and costs for pathologist and their limited numbers, has become a difficulty. This limits the amount of annotated data and leads to a large number of unannotated data, preventing traditional deep learning algorithms from being effective. We show that a semi-supervised learning method, using the Generative Adversarial Networks (GANs) for canine RBC morphology classification, can solve the lack of labelled data, when we want to train a deep learning classifier. Our semi-supervised GAN can use both labelled and unlabelled data and showed that they can achieve the same level of performance as a traditional convolutional neural network, with a smaller number of labelled images. Furthermore, we showed that augmenting the limited numbers of a labelled images enhanced the overall performance. A key benefit of our method is reduced pathologist cost and time to annotate cell images for developing a deep learning classifier.