Dual-Stream Pyramidal Attention Networks for Medical Image Registration

Image registration plays a crucial role in medical image analysis. Conventionally, it is approached using an iterative optimization technique, which can be quite time consuming. Recently, the emergence of deep learning techniques has improved deformable registration; however, many current methods struggle with significant local deformation variations, leading to inaccuracies in estimating the deformation field. In this study, we introduce LGANet, a novel method for deformable image registration. It incorporates a dual-stream pyramidal network along with an attention mecha-nism. The network progressively refines the deformation field from coarse to fine by leveraging feature interaction and fusion, in conjunction with an image decomposition mod-ule. We evaluated our method on three publicly available datasets that include cross-patient and cross-modal CT-MR registration. The evaluation results demonstrated that our approach surpasses several state-of-the-art registration methods in terms of the Dice Similarity Coefficient (DSC), Hausdorff Distance at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$95^{th}$</tex> percentile (HD95), and average symmet-ric surface distance (ASSD) metrics. The source codes are accessible at ht tps: / / gi thub. com/huangzyong / LGANet-Registration.