| |
| The quantification of magnetic resonance spectroscopy (MRS) signals remains challenging due to the low signal-to-noise ratio (SNR) of data. Time-domain quantification methods highly require user interactions, which reduce the reproducibility of the data quantification results. The goal of our work is to design a systematic methodology for automated quantification of MRS signals with low SNRs. We used Hankel singular value decomposition (HSVD) algorithm in our signal estimation step, along with extraction and reduction filter (ER-filter) as a frequency selective technique in the preprocessing step in order to avoid the interferences from nuisance peaks. In the automatic model order selection problem of HSVD, we implemented three strategies based on reconstruction residue measurement or information theoretic criteria. The performances were evaluated in terms of detection rate and relative root mean squared error (RRMSE). The simulations were run on both synthesized and semi-synthesized data. We tested the strategies in two cases, i.e., without and with an interfering signal nearby the signal of interest. It is shown that the minimum description length with condition (MDLcon) based methodology we proposed performs the best and can obtain reliable estimation performance (RRMSE<40%) when signal SNR is larger than -18dB, with a detection rate above 72.44%. The performance is consistent when the interfering signal is 0.08ppm separated. Overall, the MDLcon based automated MRS signal quantification methodology provides an effective way for low SNR MRS signal estimation and detection. Our work may shed light on automatic MRS signal quantification in clinical applications when the corresponding metabolite concentration is low. |
|
| Keywords:Magnetic resonance spectroscopy, model order selection, Hankel singular value decomposition, ER-filter |
|
