Data Augmentation for Improved Brain Tumor Segmentation

Abstract

Deep neural networks (DNN) oblige large preprocessed samples of training annotated images for successful training, which makes the approach costly particularly in the biomedical imaging domain. The data augmentation technique is regularly used by researchers to enlarge the volume of training data, creating, and producing augmented data capable to train the network about the essential properties of uniformity and stoutness. The use of conventional methods of data augmentation in most training system scenarios strictly restrict its capabilities and negatively impact the output accuracy. In this paper, we propose an automatic data augmentation technique for synthesizing high-quality brain tumor images using generative adversarial network architecture to facilitate deep learning-based methods to be trained with the limited preprocessed samples more competently. The tumor segmentation has been performed through geodesic active contour via a level set formulation. The proposed technique has been validated with different modalities of magnetic resonance imaging brain image obtained from BRATS13 datasets. Simulational results showed an enhanced performance yielding a dice similarity coefficient of 0.942.