Deform U-Net: Unsupervised Deformable 3D Biomedical Image Registration Network
Yahu Yao, Rong Lu, Jun Wu, Zaiyang Tao, Lei Qu · 2024
We can focus on the challenges posed by traditional registration algorithms and the limitations of U-Net-based architectures in capturing sufficient semantic information due to their shallow coding layers with small receptive fields. We want to explore how newer architectures or modifications to existing ones could address these limitations. Inspired by introduced deformable convolutional networks, we plan to fuse deformable convolution into our proposed network structure. Building on its success in 2D target detection, we apply deformable convolution to 3D image registration networks. The Deform U-Net proposed in this paper can compensate for the shortcomings of existing registration networks by obtaining more semantic information with unsupervised training without Introduction of additional features. We conducted evaluations on two publicly available human brain MRI, specifically LPBA40 and IBSR18. On the LPBA40 and ISBR18 datasets, the DSC reached 68.82, 62.30, respectively. The experimental results show that the Deform U-N et image registration method outperforms the VoxelMorph image registration method.