Glioma Subregions Segmentation with a Discriminative Adversarial Regularized 3D Unet

Automated segmentation of tumor and its subregions has significant clinical important in the diagnosis, monitoring, and treatment planning of brain cancer. In this paper, we proposed a new 3D Unet embedded with in a discriminative adversarial network form, designed for the segmentation of glioblastoma subregions including tumor core (TC), enhancing tumor (ET) and whole tumor (WT) from multiple Magnetic Resonance Imaging (MRI) sequences. The proposed network was trained end-to-end in the manner of generative adversarial learning fashion. Additionally, a bias correction method was applied to preprocess the MRIs before feeding to the proposed neural network. Through a 5-fold cross validation, the proposed method was proved to be effective in improving the generalization of the vanilla 3D Unet model, while maintaining sufficient segmentation accuracy.