The fine segmentation of brain tumor, which is instrumental in brain tumor diagnosis, treatment planning and prognosis, is becoming a research hotspot in medical images processing. However, manual segmentation is labor-intensive, time-consuming and easily affected by subjective and objective factors. Automatic segmentation methods based on deep learning have attracted significant interest in recent years. In this paper, a novel multi-scale mesh aggregation network (MSMANet) for brain tumor segmentation is proposed. Firstly, an improved Inception module is introduced to replace the standard convolution in the encoder to extract and aggregate effective information from different receptive fields. Secondly, a novel mesh aggregation strategy is proposed to gradually refine the shallow features and further alleviate the semantic gap. This strategy maximizes the aggregation of multi-level features at different scales and realizes the complementary advantages among features. Finally, the ability of network identification and convergence is improved by employing attention mechanism and deep supervision. Experiments were implemented on BraTS2018 to evaluate the proposed network. The dice similarity coefficient (DSC) of MSMANet in enhanced tumors, whole tumors and tumor cores are 0.758, 0.890 and 0.811, respectively. Experimental results indicate that satisfactory performance is achieved compared with the state-of-the-art methods.

脑肿瘤的精细分割有助于脑肿瘤的诊断、治疗规划和预后，正成为医学图像处理的研究热点。然而，人工分割费时费力，且容易受主客观因素的影响。近年来，基于深度学习的自动分割方法引起了人们的极大兴趣。在本文中，提出了一种用于脑肿瘤分割的新型多尺度网格聚合网络（MSMANet）。首先，引入了改进的 Inception 模块来代替编码器中的标准卷积，以​​从不同的感受野中提取和聚合有效信息。其次，提出了一种新的网格聚合策略来逐步细化浅层特征并进一步缓解语义差距。这种策略最大限度地聚合了不同尺度的多级特征，实现了特征之间的优势互补。最后，通过采用注意力机制和深度监督，提高了网络识别和收敛的能力。在 BraTS2018 上进行了实验以评估所提出的网络。MSMANet 在增强肿瘤、整个肿瘤和肿瘤核心中的骰子相似系数（DSC）分别为 0.758、0.890 和 0.811。实验结果表明，与最先进的方法相比，实现了令人满意的性能。在 BraTS2018 上进行了实验以评估所提出的网络。MSMANet 在增强肿瘤、整个肿瘤和肿瘤核心中的骰子相似系数（DSC）分别为 0.758、0.890 和 0.811。实验结果表明，与最先进的方法相比，实现了令人满意的性能。在 BraTS2018 上进行了实验以评估所提出的网络。MSMANet 在增强肿瘤、整个肿瘤和肿瘤核心中的骰子相似系数（DSC）分别为 0.758、0.890 和 0.811。实验结果表明，与最先进的方法相比，实现了令人满意的性能。