SC-RegNet: Unsupervised Self-Calibrated 3D Biomedical Image Registration Network
Jian Jun Yang, Jun Wu, Rui Zhang, Meng Wang, Yujie Tang, Lei Qu · 2021 China Automation Congress (CAC) · 2021
Deformable image registration is critical in many biomedical image analysis tasks. Traditional registration algorithms involve a time-consuming iterative optimization process. Most deep learning-based registration networks follow the U-Net shape encoding-decoding architecture, and their downsampling layers are shallow, making it difficult to obtain sufficient semantic information with a small receptive field. In this work, we extend the self-calibrated convolution applied to tasks such as 2D target detection to the encoding stage of 3D image registration networks. The proposed SC-RegNet can get more semantic information to make up for the existing registration networks in an unsupervised training manner while not introducing additional parameters. We evaluated SC-RegNet on two public available human Brain MRI datasets: LPBA40 and IBSR18, and a private mouse brain dataset. The experimental results show the superiority of SC-RegNet over the SOTA image registration methods with no extra time-consuming.