Denoising in Parallel Imaging Via Structured Low-Rank Matrix Approximation

Derya Gol¹, Lee C. Potter¹

Interpolation approaches in parallel MRI exhibit a noise amplification effect that may be mitigated via regularization techniques which are computationally expensive. In this study, we propose a pre-processing technique based on structured low-rank matrix approximation via truncated singular value decomposition (TSVD), which is able to suppress noise and ghost artifacts efficiently. TSVD method has been previously used in parallel MRI to improve the conditioning of the system matrix and to reconstruct k-space via matrix completion. In contrast to previous work, here rank properties are used to denoise acquired data in a computationally simple preprocessing for GRAPPA reconstruction.