基于双通道特征融合网络的脑部图像配准

BRAIN IMAGE REGISTRATION BASED ON DUAL-CHANNEL FEATURE FUSION NETWORK

针对传统卷积神经网络的特征融合部分没有充分考虑各通道特征重要性以及特征利用率低等局限性问题,提出一种基于多层级配准场融合策略的双通道特征融合网络用于脑部图像配准。设计基于编码-解码的卷积网络对浮动图像和固定图像进行配准场估计;设计双通道特征融合模块对同级特征进行特征融合,基于分组卷积、全局平均池化等运算对输入特征进行通道赋权,利用空间变换网络对多级特征进行空间变换。在公开数据集的配准结果表明,提出的双通道特征融合网络能有效提高配准精度。

Aimed at the feature fusion part of the traditional convolutional neural network without fully considering the limitations of the feature importance of each channel and the low feature utilization rate, a dual-channel feature fusion network based on a multi-level registration field fusion strategy is proposed for brain image registration. A coding-decoding convolutional network was designed to perform registration field estimation on moving image and fixed image. A dual-channel feature fusion module was used to perform feature fusion on peer features. Based on operations such as group convolution and global average pooling, the input features were channel weighted, and spatial transformer network was used to spatially transform the multi-level features. The registration results on public data show that the proposed dual-channel feature fusion network can effectively improve the registration accuracy.