Super-resolution Reconstruction of Fetal Brain MRI Through Learning of Multi-view Interpolation Weights

Motivation: Super-resolution reconstruction of isotropic fetal brain MR images is critical for prenatal examinations, but is hindered by fetal motion and misalignment of thick-slice scans. Goal(s): This study aims to quantify misalignment between slices and volumes while predicting isotropic high-resolution volumetric images. Approach: Our approach learns an implicit function to quantify misalignment between MR slices and volumes by obtaining interpolation weights from latent codes of multi-view motion-corrupted thick-slice stacks. The established model then interpolates the isotropic high-resolution image by utilizing predicted weights. Results: Our proposed method effectively mitigates the adverse effects of motion corruption in fetal brain MRI while substantially reducing reconstruction time. Impact: Our end-to-end fetal brain super-resolution approach bypasses traditional two-step iterative optimization paradigm, and substantially reduces reconstruction time. It eliminates the need for slice-to-volume registration, and shifts the focus of implicit neural representation from addressing appearance estimation issues to motion estimation.