DAU-Net: An unsupervised 3D brain MRI registration model with dual-attention mechanism

摘要

Medical image registration plays an important role in clinical treatment. However,the convolution-based registration frameworks fail to address the localexcessive deformation between images. Furthermore, the folding point in thedisplacement vector field (DVF) reduces the reliability of registration results.In this study, we propose a dual-attention mechanism-based U-shaped registrationframework (dubbed DAU-Net). Firstly, the multi-scale attention mechanismis introduced to extract the long-range dependence to deal with thelocal excessive deformation. Then, the channel attention mechanism is proposedto enhance the information fusion between channels, which not onlyfuses the features between different layers in the dual-attention network butalso improves the non-linear mapping ability of the registration network. Inthe end, the objective function with the folding penalty regularization term isdesigned to improve the smoothness of the DVF. The model is evaluated onLPBA40 and Mindboggle101 open datasets. The registration accuracy inLPBA40 and Mindboggle101 datasets has been increased by 2.9% and 3.1%,respectively, while the folding rate is reduced by nearly 40 times comparedwith VoxelMorph. Combined multi-scale attention mechanism with channelattention mechanism, the registration accuracy of DAU-Net is improved. Byutilizing the folding penalty regularization term, the folding rate is decreasedsignificantly.