基于注意力残差网络的跨模态医学图像配准

Cross-modality Medical Image Registration Based on Attention Residual Network

针对2D-3D医学图像配准方法时间长且容易陷入局部极值的问题,提出一种将注意力机制与残差网络融合的跨模态图像配准方法,使用深度残差网络自动提取图像特征,预测配准变换参数,在卷积块中嵌入混合域的注意力机制,提高网络对重要特征的关注度。根据变换参数的特点,设计分组回归的方式提高配准精度。实验结果表明,上述方法预测位移误差均值为0.07mm,角度误差均值为0.04°,优于其方法;配准时间仅需40ms,远低于传统方法。所提配准方法避免了传统方法循环迭代的过程,有效提高配准效率,满足医学图像配准的实时性和精度需求。

In order to solve the problem that the 2 D-3 D medical image registration method is time-consuming and is easy to fall into the local extremum,a cross-modality image registration method combining attention mechanism and residual network is proposed.A deep residual network was used to automatically extract image features,and then predict the registration transformation parameters,in which the attention mechanism of the mixed domain was embedded in the convolution block to improve the attention of important features.In addition,according to the characteristics of the transformation parameters,a group regression method was designed to improve registration accuracy.The experimental results showed that the mean translation error of the proposed method is 0.07 mm and the mean angular error is 0.04°,which was superior to other methods.The time consumption was less than 40 ms,which was much less than the traditional method.The proposed method is free of the loop iteration in traditional methods and effectively improves the registration efficiency,so it meets the real-time and accuracy requirements of medical image registration.