Non-rigid Registration Technique for Large Deformation Medical Image

Medical image registration faces significant challenges in dealing with non-rigid deformations. The problem such as folding of deformation displacement field and model degradation, all can lead to loss of registration accuracy. To address these problems, a deep learning-based unsupervised method(MulSc-Net) is proposed. On one hand, the MulSc-Net adopt a novel multi-scale registration strategy which can capture deformations at different scales with a larger receptive field via a fusion model of convolution and dilated convolution. On the other hand, an anti-folding constraint is introduced to ensure the continuity of displacement field, and a residual method is employed to prevent model degradation during training. In this work, the MulSc-Net model is evaluated on a lung CT dataset. The experimental results show that MulSc-Net can achieve better registration accuracy compared to current related methods.