Hybrid transformer and convolution iteratively optimized pyramid network for brain large deformation image registration

In recent years, the pyramid-based encoder-decoder network architecture has become a popular solution to the problem of large deformation image registration due to its excellent multi-scale deformation field prediction ability. However, there are two main limitations in existing research: one is that it over-focuses on the fusion of multi-layer deformation sub-fields on the decoding path, while ignoring the impact of feature encoders on network performance; the other is the lack of specialized design for the characteristics of feature maps at different scales. To this end, we propose an innovative hybrid Transformer and convolution iteratively optimized pyramid network for large deformation brain image registration. Specifically, four encoder variants are designed to study the impact of different structures on the performance of the pyramid registration network. Secondly, the Swin-Transformer module is combined with the convolution iterative strategy, and each layer of the decoder is carefully designed according to the semantic information characteristics of different decoding layers. Extensive experimental results on three public brain magnetic resonance imaging datasets show that our method has the highest registration accuracy compared with 9 cutting-edge registration methods, which fully verifies the effectiveness and application potential of our model design.