Deformable Image Registration via Dual-channel Cross Constraint Network for Brain MR Image
Han Zhou, Hongtao Xu, Xinyue Chang, Wei Zhang, Qiufeng Chen, Lifang Wei · 2023
Most conventional deep learning-based registration methods utilize a single-stream encoder-decoder network to compute the deformation field between two 3D volumes. Nevertheless, these conventional methods face limitations in addressing deformation consistency due to insufficient constraint information and the neglect of semantic consistency. To address these issues, we propose an innovative novel dual-channel cross constraint network for medical image registration. This network leverages both grayscale and segmentation images, which share identical semantic information and feature representations. Two encoder-decoder structures designed for the computation of two deformation fields corresponding to the grayscale image and the segmentation image respectively. These deformation fields are generated by the dual-channel cross constraint network. In order to maintain semantic consistency at corresponding positions of the grayscale and segmentation images, we introduce a directional consistency constraint. The produced deformation fields also demonstrate orientation consistency at the corresponding positions. To enhance semantic consistency, we apply the cosine similarity function to restrict the variance among deformation fields. Four public datasets are used to verify the effectiveness of our method. The evaluation experiment illustrates that our method can achieve superior performance compared with the baseline method. It obtains 79.9%, 64.5%, 69.9%, and 63.5% on Dice scores for OASIS-1, OASIS-3, LPBA40, and ADNI, respectively.