Data augmentation for medical image analysis: a Systematic Literature Review

Recent advances in deep learning models have been largely attributed to the quantity and diversity of data gathered in recent years. Data augmentation is a strategy that enables practitioners to significantly increase the diversity of data available for training models, without actually collecting new data. Data augmentation techniques such as cropping, padding, and horizontal flipping are commonly used to train large neural networks for image analysis, and in more recent years adaptive techniques, such as GAN-based or Model-based approaches, have been proposed to increase the effectiveness of data augmentation strategies. Different data augmentation strategies are likely to perform differently depending on the type of input and visual task. For this reason, it is conceivable that medical imaging may require specific augmentation strategies that produce plausible data samples and allow effective regularization of deep learning model. Data augmentation may also be used to enhance specific classes that are under-represented in the training set, e.g. to generate artificial lesion samples. The goal of this thesis is to do an extensive systematic literature review in order to answer the following research questions: (i) which study designs are used to evaluate the effect of data augmentation, (ii) what types of data augmentation are used in the medical domain, (iii) what are their effects on the performance of deep learning-based methods for medical image analysis, (iv) what types of data augmentation are not adopted in the medical domain? In order to answer these research questions, 273 papers published in reputable venues have been retrieved and analyzed to highlight trends in recent literature. The key findings of the systematic literature review have been complemented by practical experiments on a chest x-ray dataset, particularly to explore lesser-used data augmentation techniques in the literature.