Water removal in MR spectroscopic imaging with L2 regularization

Liangjie Lin, Michal Považan, Adam Berrington, Zhong Chen, Peter B. Barker

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


Purpose: An L2-regularization based postprocessing method is proposed and tested for removal of residual or unsuppressed water signals in proton MR spectroscopic imaging (MRSI) data recorded from the human brain at 3T. Methods: Water signals are removed by implementation of the L2 regularization using a synthesized water-basis matrix that is orthogonal to metabolite signals of interest in the spectral dimension. Simulated spectra with variable water amplitude and in vivo brain MRSI datasets were used to demonstrate the proposed method. Results were compared with two commonly-used postprocessing methods for removing water signals. Results: The L2 method yielded metabolite signals that were close to true values for the simulated spectra. Residual/unsuppressed water signals in human brain short- and long-echo time MRSI datasets were efficiently removed by the proposed method allowing good quality metabolite maps to be reconstructed with minimized contamination from water signals. Significant differences of the creatine signal were observed between brain long-echo time MRSI without and with water saturation, attributable to the previously described magnetization transfer effect. Conclusions: With usage of a synthesized water matrix generated based on reasonable prior knowledge about water and metabolite resonances, the L2 method is shown to be an effective way to remove water signals from MRSI of the human brain.

Original languageEnglish (US)
Pages (from-to)1278-1287
Number of pages10
JournalMagnetic resonance in medicine
Issue number4
StatePublished - Oct 2019


  • L2 regularization
  • MR spectroscopic imaging
  • metabolite map
  • water suppression

ASJC Scopus subject areas

  • Radiology Nuclear Medicine and imaging


Dive into the research topics of 'Water removal in MR spectroscopic imaging with L2 regularization'. Together they form a unique fingerprint.

Cite this