A deep learning model using localized attention mechanism for brain MRI registration

Abstract

Registration methods based on deep learning have been gotten popular for its performance and efficiency by automatically extracting deep features of images. Currently, the registration performance is further improved by using cascade networks. However, there is few studies pay attention to the registration of local complex areas of the brain. This paper combines the advantages of cascade networks and attention mechanism to solve the problem of registration in complex small areas of the brain by introducing a localized attention block. Specifically, we introduce an attention mechanism module (called Res-non-local-block), which consists of the attention mechanism (non-local-block), allowing the block to more flexibly adapt to complex anatomical structures and image changes. We also extract deformation fields of different scales by downsampling the predicted deformation fields to further improve the local registration performance. Experiments on the public datasets LPBA40 and HBN demonstrate that our proposed method exhibits higher registration accuracy on local areas of the brain than several popular deep learning based on registration models.