摘要
卷积神经网络(CNN)模型已成为当前医学图像分割的主要研究手段。为进一步提升其分割精度,提出一种多尺度双重注意力网络(MDA-Net)架构。首先,为获得更多图像特征的感受野,在传统编码器基础上提出一种多尺度特征金字塔编码器模块,以提取丰富的特征信息;其次,提出多空洞残差双注意力模块以增强图像细节特征的表征能力;最后,利用不同大小的卷积核设计了自适应多尺度融合块以更好的捕捉特征上下文信息。将MDA-Net与U-Net、Attention U-Net、CE-Net、CS-Net及R2-UNet等先进的网络在脑部、肺部、纹理细胞等分割任务上进行比较测试,实验证明了方法的优越性能,其中脑部分割的相似性指数(MIOU)相较于U-Net,CS-Net分别提升0.24%、2.16%。此外,还通过在肺部数据集上进行的消融验证了每个组件的有效性。
CNN model has become the main research method of current medical image segmentation.To further improve its segmentation accuracy,a multi-scale dual attention network(MDA-Net)architecture is proposed.Firstly,to get more image characteristics of the receptive field,a multi-scale feature pyramid encoder module is proposed on the basis of traditional encoders to extract rich feature information.Secondly,a multi-hole residual dual attention module is proposed to enhance the characterization ability for image feature details.Finally,an adaptive multi-scale fusion block is designed by using different size convolution kernels to better capture the feature context information.MDA-Net was compared with U-Net,Attention U-Net,CE-Net,CS-Net and R2-Unet on segmentation tasks of brain,lung and texture cells,which experiment proves the superior performance of this method,the MIOU of brain segmentation is 0.24%and 2.16%higher than that of U-Net and CS-Net respectively.In addition,the effectiveness of each component is conducted through ablation studies on the lung data set.
作者
纪秋浪
王继红
杨晨
丁才富
王阳
Ji Qiulang;Wang Jihong;Yang Chen;Ding Caifu;Wang Yang(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China)
出处
《国外电子测量技术》
北大核心
2022年第6期65-71,共7页
Foreign Electronic Measurement Technology
基金
国家自然科学基金(62065003)
贵州省科技基金(黔科合-ZK[2022]重点-020和一般-105)项目资助。
关键词
医学图像分割
卷积神经网络
多尺度
金字塔
注意力机制
medical image segmentation
convolutional neural network
multi-scale
pyramid
attention mechanism
