RC-Block: Refinement Coefficient for Rectifying Deformation Field

Suwei Zhang, Tai Ma, Ying Wen

Abstract: In deformable image registration, learning-based methods have demonstrated impressive performance. However, previous methods mainly focus on enhancing the capability to predict the deformation field, without fully exploring the optimal ways to update the deformation field. As a result, errors can easily accumulate during the process of updating the deformation field, thus leading to inaccurate registration results. To solve this problem, we propose RC-Block (Refinement Coefficient Block) to find the optimal way to update the deformation field, thus achieving high accuracy registration results. The proposed module introduces a novel refinement coefficient, which can leverage information from the current scale to continuously rectify the previous deformation field. In addition, RC-Block is a plug-and-play module that can be seamlessly and easily integrated into most registration methods. Extensive experimental results show that our module can effectively improve existing advanced methods’ performance with minimal additional burden.

External IDs:dblp:conf/icmcs/ZhangM024