Fusing CNNs and Transformers for Deformable Medical Image Registration
Da Hu · 2022 2nd International Conference on Computer Science, Electronic Information Engineering and Intelligent Control Technology (CEI) · 2022
Deformable medical image registration is fundamental for many medical image studies. In the field of deformable image registration, convolutional neural network (CNN) based methods cannot effectively deal with large deformation settings due to the local receptive fields. Transformers are good at modeling long-range dependency, but the lack of convolutional induction bias can lead to limited ability of grasping local details. In this paper, we propose a novel parallel architecture combining CNNs and Transformers, FTNet, to tackle this issue. FTNet contains convolutional branch and Transformer branch that extract features independently, with the proposed fusion strategy, the encoded features from both branches are efficiently fused. Extensive experiments on two brain image datasets demonstrate that our approach achieves superior registration accuracy against other state-of-the-art methods while maintaining desired diffeomorphic properties of deformation fields.