Data Augmentation by Generative Adversarial Networks for White Blood Cell Image Classification

Deep learning based algorithms have shown a great success on different tasks including image classification. One of the requirements of implementing deep learning approaches is availability of large-scale datasets. However, the lack of big medical datasets due to the difficulties in recording these kinds of data, is one of the major problems in implementing deep learning approaches. Therefore, data augmentation has become an important step for increasing the number of data samples. Image rotating in different angles, horizontal and vertical flipping is one of the popular image data augmentation methods. However, the generated images are so similar to the original ones. Recently, Generative Adversarial Neural Networks (GANs) have been proposed as powerful methods for generating new data samples. In this article, we explore image augmentation by GAN structures to be used in leukemia diagnosis task. To this end, a deep convolutional GAN is considered for generating white blood cell images to increase the number of image samples of ALLIDB database. Then, a deep Convolutional Neural Network is applied on the augmented dataset to classify the images as normal or leukemia. Experimental results verify that by implementing GAN approach for image augmentation we can achieve to 84%, classification accuracy which is 10% improvement with respect to the common augmentation method.