In　recent　years,　computer-aided　diagnosis　has　successfully　been　applied　in　various　clinical　problems.　By　this　way,　physicians　can　detect　diseases　with　a　high　accuracy.　Deep　learning　has　emerged　as　an　effective　tool　for　computer　vision.　Therefore,　this　paper　proposes　a　new　deep　learning-based　segmentation　method　for　brain　tumors　on　MRI　images.　Brain　tumor　is　a　fatal　disease　which　may　occur　in　any　position　of　human　brains.　The　change　of　brain　tumors　with　shapes　and　sizes　makes　it　difficult　for　precise　segmentation.　We　design　a　novel　architecture　of　fully　convolutional　networks　to　automatically　segment　brain　tumors　and　the　patch-wise　training　trick　is　exploited　to　train　the　model,　which　could　capture　local　information.　The　experiments　are　performed　on　the　MCCAI　Brain　Tumor　Segmentation　challenge　2015　dataset.　The　proposed　method　achieves　an　average　dice　score　of　0.82　(0.76,　0.73)　for　the　whole　tumor　(core　tumor,　enhancing　tumor)　regions.　©　2020,　Springer　Nature　Singapore　Pte　Ltd.