Image Registration with Learned Regularization
Xinjie Lai, Yang Wan, Lvda Wang, Shihui Ying · 2023
Image registration is one of the important tasks in medical image processing. The accuracy of image registration greatly affects the subsequently intelligent analysis. In this paper, we focus on the medical image registration via deep learning, and propose the unsupervised deep learning framework based on model decoupling and regularization learning. Specifically, as a highly ill-posed inverse problem, we first decompose the image registration into two simpler sub-problems to reduce the complexity of model solving. Further, two light neural networks are constructed to approximate the solutions of the two subproblems, where the training strategy of alternative iteration is used. Secondly, we introduce regularization learning module to form an end-to-end deep learning architecture. Here, we approximate the regularization term via the ISTA net, and make the regularization constraints more consistent with the actual data distribution. It reduces the registration errors caused by the deviation between the pre-set prior knowledge and the actual data distribution, and then realizes a higher accurate and data adaptive registration approach. Finally, to validate the performance of the proposed algorithm, we compare it with the VoxelMorph algorithms on brain MRI images dataset LPBA40 and lung CT image dataset DIRLAB. The experimental results show that the proposed algorithms obtain the best performance.