Abstract
Robust motion correction of fetal brain MRI slices is crucial for 3D brain volume reconstruction. However, conventional methods can only handle a limited range of motion. Hence, a deep learning model based on geometric constraints is proposed in order to predict the arbitrary motion of fetal brain MRI slices in a standard anatomical space, which consists of a global motion estimation network and a relative motion estimation network. In particular, the relative motion estimation network is used to estimate the relative motion between two adjacent slices, which is exploited as a geometric constraint. Then, sharing features between two networks make the model to learn more unique feature representations for global motion correction, and a weight-learnable strategy is employed to balance the contributions of two networks. With this design, the proposed method can estimate more complicated and large motions. Moreover, to build a large simulated fetal brain stack dataset with realistic appearance for successfully training a robust motion correction model, we introduced a control point-based method to simulate fetal motion trajectories at different gestational ages, between stacks and within 2D slices. The experimental results on a large number of fetal brain stacks demonstrate the state-of-the-art performance of our method.