A multi-scale large kernel attention with U-Net for medical image registration

Deformable image registration minimizes the discrepancy between moving and fixed images by establishing linear and nonlinear spatial correspondences. It plays a crucial role in surgical navigation, image fusion and disease analysis. Its challenge lies in the large number of deformed parameters and the uncertainty of acquisition conditions. Benefiting from the powerful ability to capture hierarchical features and spatial relationships of convolutional neural networks, the medical image registration task has made great progress. Nowadays, the long-range relationship modeling and adaptive selection of self-attention show great potential and have also attracted much attention from researchers. Inspired by this, we propose a new method called Multi-scale Large Kernel Attention UNet (MLKA-Net), which combines a large kernel convolution with the attention mechanism using a multi-scale strategy, and uses a correction module to fine-tune the deformation field to achieve high-accuracy registration. Specifically, we first propose a multi-scale large kernel attention mechanism (MLKA), which generates attention maps by aggregating information from convolution kernels at different scales to improve local feature modeling capabilities of attention. Furthermore, we employ large kernel dilation convolution in proposed attention to construct sufficiently long-range relationships, while keeping lower number of parameters. Finally, to further improve local accuracy of the registration, we design an additional correction module and unsupervised framework to fine-tune the deformation field to solve the issue of original information loss in multilayer networks. Our method is compared qualitatively and quantitatively with 24 representative and advanced methods on the 3 public available 3D datasets from IXI database, LPBA40 dataset and OASIS database, respectively. The experiments demonstrate the excellent performance of the proposed method.