Deformer: Towards Displacement Field Learning for Unsupervised Medical Image Registration

AbstractRecently, deep-learning-based approaches have been widely studied for deformable image registration task. However, most efforts directly map the composite representation to spatial transformation through convolutional neural network, ignoring its limited ability capture correspondence. On other hand, Transformer can better characterize relationship with attention mechanism, long-range dependency may be harmful task, where voxels too large distances are unlikely corresponding pairs. In this study, we propose a novel Deformer module along multi-scale framework The is designed facilitate mapping from by formulating displacement vector prediction as weighted summation of several bases. With predict fields in coarse-to-fine manner, superior performance achieved compared traditional and learning-based approaches. Comprehensive experiments on two public datasets conducted demonstrate effectiveness proposed well framework.KeywordsDeformable registrationDisplacement basesMulti-scale