Abstract: Brain tumor is a common brain disease that heavily threatens human health,so accurate brain tumor segmentation is vital for clinic diagnosis and treatment of patients.Due to different shapes and sizes,unstable positions and fuzzy boundaries of brain tumors,it is a challenging task to achieving high-precision automatic brain tumors segmentation.Recently,U-Net has become the mainstream model for medical image segmentation due to its concise architecture and excellent performance.But it also has some problems,such as limited local receptive field,spatial information loss and insufficient use of context information.Therefore,we propose a new cascade U-Net model based on dynamic convolution and non-local attention mechanism,named CDAU-Net.Firstly,a two-stage cascade 3D U-Net architecture is proposed to reconstruct more detailed and high-resolution spatial information of brain tumors.Then,the expectation-maximization attention is added to the skip connection of CDAU-Net,and the tumor context information is better utilized by improving the network’s ability to capture long-distance dependencies.Finally,the normal convolution is replaced by the dynamic convolution with local adaptive ability in CDAU-Net,which can further enhance the local feature capture ability of the network.Extensive experiments are conducted on public dataset BraTS 2019/2020 and compared with other representative methods,and experimental results show that the proposed method is effective in brain tumor segmentation.The CDAU-Net obtained the Dice values of whole tumor,tumor core and enhancing tumor segmentation on BraTS 2019/2020 verification sets are 0.897/0.903,0.826/0.828 and 0.781/0.786,respectively,which achieves good brain tumor segmentation performance.

Key words: Brain tumor segmentation, U-Net, Cascade network, Dynamic convolution, Expectation-maximizing attention