A Fetal MRI 3d Reconstruction Pipeline with Volume-Slice Optimization Registration and Self-Supervised Generation Model

Abstract

The slice-to-volume implicit neural representation methods can predict the slice intensity of fetal MRI stacks. However, the slice outliers and severe motion artifacts remain to be solved, such as limitations of regional intensity differences and consistency discrimination of global voxel intensity distribution in 3D space. To this end, we propose a novel fetal MRI reconstruction pipeline with Joint Optimization Registration and Coarse-to-fine Generation model (JORCG). Firstly, the enhanced joint volume-slice optimization strategy is used for rigid registration and deblurring of slices. Then, a coarse-to-fine Radiation Diffusion Generation Model (RDGM) is used to construct high-quality volume, which includes self-supervised Consistent Implicit Neural Representation (CINR) network and Global Diffusion Discriminative Generation (GDDG) module. CINR utilizes a double-spatial voxel correlation learning mechanism to reduce regional voxel intensity differences. CINR enhances regional 3D voxels for association and complementarity to generate a coarse MRI volume. We also fine-tune the weighted slice reconstruction loss and add voxel regularization to improve the training convergence speed. GDDG can enhance volume consistency discrimination of global voxel intensity distribution. Diffusion theory transforms all the discriminant information of 3D voxels and eliminates the uncertainty of voxel intensity distribution. Notably, The self-supervised trained CINR serves as an intensity transformation network for the GDDG module. GDDG continuously optimizes the voxel noise distribution to make the uncertain noise useful for MRI intensities. The experiments on some real-world fetal MRI datasets demonstrate that JORCG achieves state-of-the-art results.

Keywords: Fetal MRI, Coarse-to-Fine, Implicit Neural Representation, Reconstruction Loss, Diffusion