Joint affine and deformable three‐dimensional networks for brain MRI registration

• DOI: 10.1002/mp.14674
• Publication Date: 2020-12-20T16:46:10Z
• Authors: Zhenyu Zhu; Yiqin Cao; Chenchen Qin; Yi Rao; Di Lin; Qi Dou; Dong Ni; Yi Wang
• Keywords: Image registration; Similarity (geometry); Hausdorff distance

Purpose Volumetric medical image registration has important clinical significance. Traditional methods may be time‐consuming when processing large volumetric data due to their iterative optimizations. In contrast, existing deep learning‐based networks can obtain the quickly. However, most of them require independent rigid alignment before deformable registration; these two steps are often performed separately and cannot end‐to‐end. Methods We propose an end‐to‐end joint affine network for three‐dimensional (3D) registration. The proposed combines deformation methods; first one is obtaining second a subnetwork achieving nonrigid parameters subnetworks shared. global local similarity measures used as loss functions subnetworks, respectively. Moreover, anatomical devised weakly supervise training whole network. Finally, trained perform in forward pass. Results efficacy our was extensively evaluated on three public brain MRI datasets including Mindboggle101, LPBA40, IXI. Experimental results demonstrate consistently outperformed several state‐of‐the‐art with respect metrics Dice index (DSC), Hausdorff distance (HD), average symmetric surface (ASSD). Conclusions provides accurate robust without any pre‐alignment requirement, which facilitates