An Attention Augmentation-Based Transformer Network for Unsupervised Medical Image Registration

Transformer-based models have achieved significant success in medical image registration in recent years. Since the self-attention operation has quadratic complexity, it usually causes huge computational overhead for these methods. So, how to provide higher quality registration while being efficient in terms of parameters and computational cost is a research hotspot. For this goal, we propose A²TNet, an attention augmentation-based transformer network, wherein the attention augmentation is achieved via combining the spatial attention and channel attention together. Meanwhile, a shifted window mechanism is introduced to further reduce the calculation complexity of the proposed attention module. Experiments carried out on two different brain MRI datasets, LPBA and Mindboggle, demonstrate that A²TNet can improve registration accuracy while effectively controlling complexity compared to existing deep learning registration models.