近几年,深度学习在生物医学图像处理中的应用得到了广泛关注。从深度学习的基本理论和医学领域应用出发,提出了一种改进的三维双路径脑肿瘤图像分割网络,用于提高核磁共振成像序列中对脑肿瘤各个区域的检测精度。所提算法以3D-UNet为基础架构,首先,使用改进的双路径网络单元构成类似于UNet的编码-解码器结构,该网络单元在保留原有特征的同时,还可以在脑肿瘤的纹理、形状和边缘等方面产生新特征,来提高网络分割精度;其次,在双路径网络模块中加入多纤结构,在保证分割精度的同时减少了参数量;最后,在每个网络模块中的组卷积之后加入通道随机混合模块来解决组卷积导致的精度下降问题,并使用加权Tversky损失函数替代Dice损失函数,提高了小目标的分割精度。所提模型的平均Dice_ET、Dice_WT和Dice_TC均优于3D-ESPNet、DeepMedic、DMFNet等算法。该研究结果具有一定的现实意义和应用前景。

In recent years, the application of deep learning in biomedical image processing has received widespread attention. Based on the basic theories of deep learning and medical applications, this paper proposes an improved three-dimensional dual-path brain tumor image segmentation network to improve the detection accuracy of brain tumors in nuclear magnetic resonance imaging sequences. The proposed algorithm is based on 3D-UNet. First, the improved dual-path network unit is used to form the encoder-decoder structure similar to UNet. While retaining the original features, the network unit can also generate new features in texture, shape, and edge of the brain tumor to improve the accuracy of network segmentation. Second, the multi-fiber structure is added to the dual-path network module, which reduces the amount of parameters while ensuring the accuracy of the segmentation. Finally, after the group convolution in each network module, the channel random mixing module is added to solve the problem of accuracy reduction caused by group convolution, and the weighted Tversky loss function is used to replace the Dice loss function to improve the segmentation accuracy of small targets. The average Dice_ET, Dice_WT, and Dice_TC of the proposed model are better than 3D-ESPNet, DeepMedic, DMFNet, and other algorithms. The research results have certain practical significance and application prospects.