<p>Primary intracranial tumors caused by glioma may result in death of the patients. Fortunately, glioma patients can be treated if the tumors are identified at an early stage. Deep learning realizes automatic segmentation of glioma from Magnetic Resonance Images (MRI). One of the most classic algorithms is the U-Net network, which is mainly due to its unique encoder and decoder structure and strong feature extraction capability. In this paper, we summarize the variants of U-Net algorithms comprehensively. Firstly, the main elements to improve segmentation performance are analyzed, including convolution operation and effective module. Dilated convolution provides technical support for multi-scale feature extraction, and various attention mechanisms make the model better at extracting important features. We find lots of variants that combine these two approaches to enhance the segmentation performance. The Dice score in WT and ET can reach up to 93.37 and 90.17, respectively. Moreover, the parameters of the lightweight model can even be reduced to 0.19 M. Cascade network and multi-task architecture are important approaches to handle class imbalance in glioma. Additionally, Dice Loss, Cross-Entropy (CE) Loss, and variant loss functions are extensively formulated. The multimodal loss function is also introduced for better performance. Combined with clinical requirements, optimal modality selection, modality fusion, and modality missing have emerged as new research areas. However, these algorithms still have some limitations, such as heavy dependence on datasets and poor interpretability, which hinder the promotion and application in clinical.</p>