Multi-modal Brain Tumor Segmentation Using Cascaded 3D U-Net

Automatic segmentation of brain tumor plays an important role in the process of diagnosis and treatment. In recent years, scholars have been actively studying 3D brain MRI images, in order to achieve the automatic segmentation of the lesion area and provide more objective basis for clinicians' diagnosis and treatment. This paper proposes a multi-modal brain tumor segmentation algorithm based on improved 3D U-Net. According to the difference of each mode of MRI images, a cascade network architecture is constructed, which gradually completes the segmentation of brain tumor regions from coarse to fine. The basic segmentation network is based on 3D U-Net, and the deep residual module is added to solve the gradient degradation problem in the process of network training. In the aspect of loss function, this paper introduces category weight into Tversky loss function to get weighted loss, which improves the segmentation accuracy of small targets. The experimental results show that on the BraTS2018 validation data set, the Dice coefficients of the proposed network in the whole tumor, tumor core and enhanced tumor are 0.883, 0.834, 0.746, respectively.