Brain tumors are one of the most lethal diseases in the world. The segmentation of brain tumor is of great significance for physician in formulating appropriate diagnostic and treatment plans, not only accurate but also efficient 3D segmentation algorithms are urgently demanded in clinical practice. Nowadays, several 3D convolution neural networks have achieved impressive segmentation performance. However, these architectures come with extremely high computational overheads due to the extra depth dimensionality in 3D convolution, which may make these models prohibitive from practical large-scale clinic application. In this work, we aim at designing a more efficient and lightweight network without accuracy reduction for real-time segmentation of magnetic resonance images. To this end, we propose a multi-branch sharing network which consists of novel multi-branch sharing units. Different from other works, our proposed multi-branch sharing units focus the information sharing and communication between grouped layers by leveraging a Multiplexer operation, which can reduce the computational cost significantly while maintaining decent performance. Extensive experimental results on the BraTS2018 challenge dataset show that the proposed architecture achieve real-time inference while maintaining high accuracy for 3D brain magnetic resonance image segmentation.

脑肿瘤是世界上最致命的疾病之一。脑肿瘤的分割对于医师制定适当的诊断和治疗计划具有重要意义，在临床实践中不仅急需准确而且有效的3D分割算法。如今，几个3D卷积神经网络已经实现了令人印象深刻的分割性能。但是，由于3D卷积中额外的深度维度，这些体系结构的计算开销非常高，这可能会使这些模型无法在实际的大规模临床应用中使用。在这项工作中，我们旨在设计一种效率更高，重量更轻的网络，而不会降低准确性，以实时分割磁共振图像。为此，我们提出了一个由新的多分支共享单元组成的多分支共享网络。与其他工作不同，我们提出的多分支共享单元通过利用多路复用器操作集中了分组层之间的信息共享和通信，可以在保持良好性能的同时显着降低计算成本。在BraTS2018挑战数据集上的大量实验结果表明，所提出的体系结构实现了实时推理，同时保持了3D脑磁共振图像分割的高精度。这样可以在保持良好性能的同时显着降低计算成本。在BraTS2018挑战数据集上的大量实验结果表明，所提出的体系结构实现了实时推理，同时保持了3D脑磁共振图像分割的高精度。这样可以在保持良好性能的同时显着降低计算成本。在BraTS2018挑战数据集上的大量实验结果表明，所提出的体系结构实现了实时推理，同时保持了3D脑磁共振图像分割的高精度。