基于双重注意力机制和迭代聚合U-Net的脑肿瘤MR图像分割方法

Brain tumor MR image segmentation method based on double attention mechanism and iterative aggregation U-Net

脑肿瘤核磁共振图像分割是脑肿瘤临床诊断的基础.针对传统U-Net网络中编码器无法从多尺度提取特征信息以及跳跃连接过程特征融合信息不全面等问题,提出了一种引入双重注意力机制以及迭代聚合的U-Net脑肿瘤分割算法.首先,在U-Net编码器部分,引入了卷积核注意力机制SKNet,网络可自适应选择卷积核尺寸,获取不同尺度的特征信息;其次在解码器部分添加了通道注意力模块CAM,使网络模型聚焦于重要的特征信息,减弱无关信息的干扰;最后在跳跃连接部分引入迭代聚合的思想,让网络将高级语义特征和低级语义特征进行融合,使得特征信息更加丰富全面,进而提高分割精度.在BraTS2019数据集上实验结果表明:肿瘤的整体区域(WT)、核心区域(TC)以及增强区域(ET)的Dice相似系数分别为80.51%,71.46%,71.32%,较原始模型分别提升了1.34%,4.34%,3.41%.与其他模型相比,该算法具有很好的分割性能.

Brain tumor MR image segmentation is the basis of the clinical diagnosis of brain tumors.In view of the problems that the encoder in the traditional U-Net network cannot extract feature information from multiple scales and the feature fusion information in the skip connection process is not comprehensive,a method that introduces double attention mechanism and iterative aggregation U-Net brain tumor segmentation algorithm is proposed.Firstly,in the U-Net encoder part,the convolution kernel attention mechanism SKNet is introduced,and the network can adaptively select the size of the convolution kernel to obtain different scale feature information.Secondly,a channel attention module CAM is added to the decoder part to make the network model focus on important feature information and reduce the interference of irrelevant information.Finally,iterative aggregation is introduced in the skip connection part.The idea is to let the network fuse high-level semantic features and low-level semantic features,so that the feature information is more abundant and comprehensive,and then the segmentation accuracy is improved.The experimental results on the BraTS2019 dataset show that the Dice similarity coefficient of the whole tumor region(WT),tumor core region(TC),enhancing tumor region(ET)are 80.51%,71.46%,and 71.32%,respectively,which are 1.34%,4.34%,and 3.41%higher than the original model.Compared with other models,the algorithm also has good segmentation performance.