基于双模特征强化注意力网络的新生儿脑组织图像分割

Neonatal brain tissue image segmentation based on a bimodal feature-boosted attention network

目的探讨基于双模特征强化注意力网络的新生儿脑组织图像分割方法及其应用价值。方法针对T1和T2双模态新生儿脑磁共振图像, 在编码路径采用双模特征提取、可分离卷积和空间-通道注意力, 在解码路径引入可分离卷积。采用数据互补性和完备性对提取的双模特征进行融合构建潜在统一表达, 使用跳跃卷积连接编码路径和解码路径来增强信息传送。结果基于双模特征强化注意力网络的新生儿脑组织图像分割方法能够准确分割白质、灰质和脑脊液, 超过现有基准方法。结论基于双模特征强化注意力网络的新生儿脑组织图像分割方法不仅改善图像分割的准确性和有效性, 还可促进磁共振成像技术在新生儿大脑生长发育和健康评估的临床应用。

Objective To explore a method for neonatal brain tissue segmentation based on bimodal feature-boosted attention network(BFAN)and its value in clinical use.Methods T1-and T2-weighted MR images of neonatal brain were processed with BFAN,whose encoder path was supported by bimodal feature extraction,separable convolution and spatial-channel attention,and the decoder path involved separable convolution.The extracted bimodal features were fused to build up latent common representation by using complementarity and completeness of the bimodal data.Skip convolution was used to connect the encoder and decoder paths to enhance information transmission.Results BFAN-based segmentation of neonatal brain tissue image could accurately segment the images into white matter,gray matter and cerebrospinal fluid,out-performing the benchmarking modality currently available.Conclusion BFAN-based neonatal brain tissue segmentation improves not only the accuracy and effectiveness of image segmentation,but also promotes clinical use of MRI techniques in assessing the brain development and health in neonates.