Comparison between Transfer Learning and Data Augmentation on Medical Images Classification

Abstract

Image classification is a hot research topic in today’s society and an important direction in the field of image processing research. In this paper, we examined the classification improvement of two strategies in images data set with small samples: the first approach is data augmentation using the Generative Adversarial Networks (GANs). (GANs) are a mechanism for the production of artificial data with a distribution close to the distribution of real data. The second approach by using transfer learning methods to overcome the problem of small number of training data. In this study the techniques of transfer learning were preferred over other machine-learning algorithms because of the excellent classification accuracy of pre-trained models, which saves time by avoiding training problems and scratch checks of model weights. We have used different Measures to evaluate the different classifiers on medical images dataset using Classification Based GAN Augmentation (CBGA), and three transfer learning method (TL-VGG), Inception (TL-INC) and Resnet 50 (TL-RE). Through experimenting the two strategies on a different datasets, we observed that using the transfer learning approach is significantly better than using classification-based- on data augmentation on the same dataset. This approach saves not only considerable time, but also competitive performance accuracy.