Evaluation of Data Augmentation of MR Images for Deep Learning

Abstract: Data augmentation is a process to create new artificial data by altering the available data set. It has by many been considered a key factor for increasing robustness and performance of image classification tasks using deep learning methods, such as the convolutional neural network. This report aims to investigate the effect of data augmentation for the segmentation of magnetic resonance (MR) images using neural networks. Geometric data augmentation techniques such as rotation, scaling, translation and elastic deformation are evaluated for different data set sizes. Also, motion artifacts are simulated in MR images as an augmentation to increase segmentation performance on images containing real artifacts. A method for simulating these artificial motion artifacts is developed, where the k-space of the image is altered to simulate patient movement due to breathing. Results showed that geometric data augmentation improved performance of the network for smaller data sets, but lost importance as the amount of available training data grew bigger. Segmentation performance of images containing motion artifacts significantly increased if the neural network had been trained on a training set augmented with simulated artifacts.