I. Introduction
The location and appearance of brain tumors are key to the diagnosis and treatment of brain cancers. Such information is usually acquired by the non-invasive magnetic resonance imaging (MRI) and extracted by segmenting the tumor regions in the scanned images. However, accurate brain tumor segmentation is always challenging. For example, as one of the most aggressive brain tumors, glioma affects tens of thousands adults around the world [1]–[3]. Compared with some brain tumors like meningiomas, the extent of gliomas is more difficult to accurately define, due to their different shapes, sizes, and diffused locations which vary from patient to patient [4]. Moreover, gliomas often extend their tentacle-like structures to invade the healthy brain tissues rather than just replacing them, which results in subtle changes and fuzzy tumor boundaries [5]. The accurate manual annotations of gliomas require laborious efforts from the professional radiologists [6] and could burden the workload in clinics. Therefore, automatic brain tumor segmentation methods on MR images would be beneficial in speeding up the process and provide objective and repeatable measurements for the radiation therapy treatment.