Abstract:

Deformable registration plays a significant role in multiple medical image analysis tasks. However, the fixed receptive field of convolutional neural networks is difficult to fully capture the spatial context information of the brain, while the Transformer architecture can effectively capture global information but has a high computational cost. Therefore, this paper proposes a deformable registration network based on multi-scale dilated residual convolution and dual attention. This network adopts a multi-scale dilated residual convolution to simultaneously capture local details and extensive context information. The dual attention module is adopted to enhance the expressive ability of features. The decoder section introduces dynamic upsampling to finely reconstruct high-frequency details to ensure the topological integrity of the deformation field. Experiments were conducted on two public brain datasets. The results demonstrate that the proposed model achieves high registration effects in multiple indicators such as Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95) and percentage of non-positive Jacobian determinant, demonstrating superior registration precision and smoother deformation fields.

Key words: Deformable image registration, multi-scale, dual attention, dynamic upsampling