Improving Image Classification Through Generative Data Augmentation

Abstract

As the industrial adoption of machine learning systems continues to grow, there is incredible potential to use this technology to revolutionize how medical diagnostic imaging is performed. The ability to accurately classify the information contained within a medical image is of critical importance for clinical implementation. Successful application of machine learning classification algorithms has traditionally relied on the availability of copious amounts of labelled training data. Unfortunately, medical datasets are typically small due to privacy constraints and the large cost associated with annotating the data. To ameliorate this limitation, a training scheme is developed in this thesis which can operate on small-scale datasets by using a generative adversarial network to augment the dataset with synthetic images. Through quantifying the uncertainty in the classification network, training samples are selected to maximize the performance of the classifier while minimizing the amount of required data. Furthermore, privacy constraints are preserved as the images sampled from the generative adversarial network are inherently anonymized. The experimental results demonstrate the efficacy in this approach and viability for application in the medical domain.