Structural Low-Rank Regularized Diffusion Model for Even-Odd Echo Separation in EPI Reconstruction

Motivation: In accelerated EPI, besides aliasing patterns caused by undersampling, eddy artifacts often arise, significantly degrading image quality. Goal(s): We aim to construct a method that can accelerate EPI (effectively suppress aliasing pattern) while suppressing eddy artifacts. Approach: Diffusion models effectively suppress aliasing patterns caused by accelerated imaging. Given that eddy artifacts arise from odd-even echo switching, we construct a diffusion model for separate reconstruction of odd-even echo acquisition signals. Additionally, we incorporate Structural Low-Rank prior into proposed diffusion model to ensure robustness. Results: Taking DWI data as an example, our method works well in removing both eddy artifacts and aliasing artifacts. Impact: We propose a diffusion model with Structural Low-Rank prior that couples odd-even echo acquisition signals, effectively suppressing eddy artifacts while reconstructing magnetic resonance images from undersampled $k$-space data.