Camer-Rao Lower Bound Analysis of 1H Magnetic Resonance Spectroscopic Imaging for Breast with Singular Value Decomposition, LCModel and QUEST
W Feng1, Y Xuan2, A Chu3, J Hu4*, (1) New York Presbyterian Hospital, Tenafly, NJ, (2) Wayne State University, Detroit, MI, (3) Yale New Haven Hospital, New Haven, CT, (4) Wayne State University, Detroit, MIWE-C-116-12 Wednesday 10:30AM - 12:30PM Room: 116
Purpose: LCModel and QUEST supply error estimate capability, however no standard basis available for unknown lipids mixture in breast. Singular Value Decomposition (SVD) with Linear Combination method was developed to implement desired CRLB function. The purpose of this study is to evaluate SVD, QUEST and LCModel quantification for breast MRSI with Camer-Rao Lower Bound (CRLB) analysis.
Methods: Spectrum for 6 metabolite were simulated with Vespa (STEAM SVS, 3.0T, TE=42ms, TM=10ms), mixture spectrum was simulate with known ratio. jMRUI 4.0 (QUEST function) and LCModel 6.3 was used in the study, metabolite basis was generated from individually simulated data. In-house SVD was developed with Matlab 7.6 to combine Hankel SVD and Linear Combination method. The estimated FID signals are constructed from the quantification parameters, residue error spectrum was calculated by FFT of the difference of original FID and estimated FID. In order to study the CRLB, known simulated spectrum without noise was analyzed with SVD, QUEST and LCModel. Then introduce random noise, quantification residue was analyzed, and estimated parameters with CRLB were compared with results without noise. Sample patient breast MRSI data was collected with Siemens TrioTim (PRESS CSI, 3.0T, TE=80ms).
Results: For simulated data, both Quest and LCModel can reliably quantify with SNR of 5 and signal as low as 17% of maximum; while SVD method can quantify with SNR of 10.
Representative breast MRSI data were analyzed, LCModel residue error is 13.3% of original maximum and Cho CRLB is 56%; while SVD can achieve 3.7% and 3.6% respectively.
Conclusions: For all methods, CRLB estimations increase with decrease the SNR. SVD is sensitive to noise, since it has no prior knowledge to reduce noise, if noise is larger than signal, fake peak can be introduced; while LCModel and QUEST can not quantify breast data well with limited simulated lipid basis.