Abstract

Info

Author(s):: JI Yarong; WANG Yu; XIAO Hongbing; XING Suxia; School of Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China

Keywords:: Keywords: glioma TensorMixup Mixup data augmentation deep learning

Abstract:: Abstract: The automated segmentation of glioma and its subregions is of great significance for the diagnosis, treatment and monitoring of brain cancer. A data augmentation method called TensorMixup is proposed based on the improvement of traditional Mixup, and it is applied to the three-dimensional U-Net for glioma segmentation. The image patches with the size of 128×128×128 voxels are selected according to glioma information of ground truth labels from the magnetic resonance imaging data of any two patients with the same modality. A tensor in which all elements are independently sampled from Beta distribution is used to mix the image patches, and then the tensor is mapped to a matrix which is used to mix the one-hot encoded labels of the above image patches, so as to synthesize a new image and obtain its encoded labels. The model is trained using the synthetic data for improving the segmentation accuracy. The experiment on the proposed algorithm on BraTs2019 data set shows that the mean Dice coefficients are 91.32%, 85.67%, and 82.20% in the whole tumor, tumor core, and enhancing tumor region, respectively, which indicates that TensorMixup is feasible and effective for glioma segmentation.