LKCA-Net: Medical Image Registration Network Enhanced by Large Kernel Convolution Attention

Abstract

In medical image registration, current models cannot simultaneously achieve favorable registration performance and maintain a small number of computational parameters. To solve this problem, this paper proposes a method based on U-shaped structure. By introducing Convolutional Block Attention Module (CBAM) and multi-scale large kernel convolution (LK) into this structure (LKCA-Net), it enhances feature expression ability, reduces information loss and increases the receptive field in the registration process. This method adopts the attention mechanism at the early stage of the encoder to simulate the way human eyes preferentially focus on key target areas when initially observing images. Meanwhile, at the 1/4, 1/8 and 1/16 resolution scales of the encoding path, large kernel convolution modules with parallel 1×1×1, 3×3×3 and 5×5×5 branches are deployed, and more contextual information can be effectively captured under controllable computational overhead by means of grouped convolution. The experimental results show that compared with indicators including Dice and parameter size, the LKCA-Net model using attention mechanism and multi-scale large kernel is obviously superior to the currently popular TransMorph and VoxelMorph models in terms of computational parameters and registration accuracy.