Two-Stage Implicit Representation Reconstruction with Iterative Registration and Double-Channel Encoder for Fetal MRI

Reconstructing 3D Magnetic Resonance Imaging (MRI) from multiple motion-corrupted stacks initially addressed challenges such as fetal motion and unstable maternal environments encountered in prenatal brain MRI diagnostics. However, The existing fetal MRI reconstruction methods still have the problem of structural information loss caused by poor reconstruction volume quality and insufficient stability. To this end, we propose a novel Two-stage Implicit Representation Strategy, TwIRS. The first stage is training by the 3D implicit representation network with iterative Slice-to-Volume Registration (SVR) and double-channel encoder. It includes an iterative registration module to solve severe shoe motion, and an implicit filed reconstruction module of dual-channel feature encoder fusion to solve the loss of structural information caused by low consistency and high noise. In the second stage, the previously trained implicit field reconstruction module is used for sampling to reconstruct a higher-resolution volume. Most experiments demonstrate that our method achieves state-of-the-art reconstruction quality.