Rudy Rizzo1 and Roland Kreis1
1Department of Radiology and Biomedical Research, University of Bern, Bern, Switzerland
Synopsis
Potential
problems arising from restricting the fitting algorithm in MR spectroscopy to a
limited parameter space of physically meaningful values are investigated via Monte-Carlo
approach and theoretical considerations. Three parameter-space configurations
are compared to evaluate potential bias in the estimated mean cohort concentration
for a simulated cohort study typical for conditions for MRS of hippocampus. The
bias found for restrictions to positive concentration values can be ameliorated
based on an estimate of the distribution width, e.g., based on Cramer-Rao bounds.
Loosening parameter space restrictions can also eliminate bias while
maintaining the benefit of parameter restrictions for the search algorithm.
Introduction
Most
fit packages for in vivo MR spectroscopy offer ways to restrict fit parameters
to remain in a space of physically meaningful values either by enforcing prior
knowledge relations and/or setting bounds for the available parameter space. However,
random spectral noise and other random factors inherently lead to a Gaussian
distribution of values, which is the basis for unbiased averaging to obtain the
cohort mean of estimated values. Excluding natural variance-related parts of
solution space may hence lead to estimation bias. In particular, limiting
concentrations to positive values (i.e., positive peak areas) for low-concentration
metabolites or in general for spectra with limited SNR (e.g. from a small VOI
inside hippocampus) may lead to overestimation of the cohort average. This can
be prevented by excluding metabolites from evaluation based on Cramer Rao
Bounds (CRLBs) criteria, which also includes dangers for bias1 and which
prevents estimation of cohort averages independent of the cohort size. Here we
investigate the size of the effect using Monte-Carlo simulations plus CRLBs calculations and propose a correction term for zero-bounded fits as well as
alternative fitting bounds.Methods
Spectra of a simplified metabolite mixture
were simulated by NMR-scopeB2 with SNR and linewidth realistic for hippocampal
spectra at 3T3 (Fig. 1). A semiLASER sequence with TE=35 ms, spectral
width 4000 Hz, 4096 datapoints was simulated including real pulse shapes. Voigt
lines were imposed as lineshapes with T2=130 ms and Gaussian width of
5.5 Hz. Two cohorts of 2000 cases each were designed with small concentration
differences including ground truth (GT) values for GABA about equal to 1 or 0.5
times its CRLB value and zero GT value for Lactate (Lac) (Fig. 1). Fit (and CRLBs calculation) was performed using FitAID4 with a Levenberg-Marquardt
algorithm without (LM) and with lower area bounds at zero (LMB0) or at minus one CRLB (LMB-1CRLB). The expected distributions of estimates and
the calculation of the proposed bias correction term are based on formulas for
a truncated Gaussian distribution5 where its natural width is
assumed to be equal to the mean estimated CRLB, as illustrated in Fig. 2. Results & Discussion
Estimates of cohort averages are affected
by detectable bias in a 0-bounded LM fit if the relative CRLB for single
measurements is >25% and bias becomes substantial for relative CRLB >50%. In our synthetic examples, GABA and Lac show the
biggest effect as they were simulated in low or absent concentration. The distributions
of estimated concentrations are presented in Fig.3. The proposed corrections work
well, but not perfectly, because the fitting algorithm does not yield zero
concentrations for those cases that should come out with negative values if not
bounded in parameter space. Indeed, the corrected estimated means move close to
GT values (Fig. 4) but are only partly included within the confidence intervals
around GT ($$$± 2*SEM$$$, standard error of the mean $$$= CRLB / \sqrt{N}$$$, N: cohort size).
- COHORT 1: GABA $$$
= 0.524 ∈ 0.500 ± 0.024$$$; Lac
$$$ = 0.056 ∉ 0.000 ± 0.013$$$.
- COHORT 2: GABA
$$$ = 0.263 ∈ 0.250 ± 0.024$$$; Lac $$$
= 0.084 ∉ 0.000 ±
0.013$$$.
Further optimization can be done when accounting
for a non-zero $$$\mu_{TL}$$$ contribution. Using an iterative
look up function, $$$\mu_{TL}$$$
can be seen to be close to 0 and can
be incorporated in the correction function (Fig. 5). The relationship between $$$\mu_{TL}$$$ and the distribution characteristics would need
to be investigated for each fit package.
Extending the allowed parameter space with
a lower boundary of -1 CRLB for each estimated metabolite would reduce the bias
substantially (Fig. 4) and seems to largely retain the benefits of parameter
space restrictions (i.e., prevention of grossly wrong local $$$\chi^2$$$ and easing the
path for $$$\chi^2$$$ minimization.)
Conclusions
- Restricting the parameter space
for concentrations to positive values leads to bias in cohort averaging. Large
effects are seen for cases where the mean relative CRLB in single spectra is
>50% of the GT, but it can also realize with smaller CRLB in small cohorts.
- Instead of refraining from
reporting results for metabolites with large fitting uncertainties, cohort
averages can be corrected for the expected bias.
- The proposed correction term
can substantially reduce this bias for large cohort sizes while for small
cohorts care has to be used to prevent incidental findings due to small numbers
and t-tests would not be appropriate for group comparisons.
- Here, it was assumed that the
CRLB is a valid proxy for the width of the cohort distribution. This assumption
can be extended if data for true repeatability of estimates is available and
might be deduced from total variance analysis of a set of strongly-represented metabolites.
- The broadening term $$$\mu_{TL}$$$
for the values expected near zero, which may
help to increase the accuracy of the correction, is expected to be fit-package
dependent.
- Lowering the area parameter bound
from zero to -1 CRLB essentially eliminates the bias without correction term
needed and is proposed as compromise between strict zero-bounded and free fitting.
- Limiting the fit-parameter
space to meaningful values remains appropriate for evaluation of single
subjects or small groups but should be reconsidered for larger cohorts.
Acknowledgements
This work is supported by the Marie-Sklodowska-Curie Grant
ITN-39 237 (Inspire-Med).References
- R. Kreis, The trouble with quality filtering based on relative Cramer-Rao lower bounds, Magn. Reson. Med. 75, 15-18 (2016).
- Z. Starcuk Jr, J. Starcukova,
O. Strbak and D. Graveron-Demilly: Simulation of coupled-spin systems in the
steady-state free-precession acquisition mode for fast magnetic resonance (MR)
spectroscopic imaging. Meas. Sci. Technol. 20, 10, 104033 (2009).
- N. Allaili, R.
Valabregue, E.J. Auerbach, et al. Single-voxel 1H
spectroscopy in the human hippocampus at 3 T using the LASER sequence:
characterization of neurochemical profile and reproducibility. NMR Biomed.
28(10), 1209-1217 (2015).
- D. G. Q. Chong, R. Kreis, C. S. Bolliger, et
al. Two-dimensional linear combination model fitting of
magnetic resonance spectra to define the macromolecule baseline using FiTAID, a
Fitting Tool for Arrays of Interrelated Datasets. Magn. Reson. Mater. Physics,
Biol. Med. 24, 147-164 (2011).
- N. Johnson, S. Kotz, N. Balakrishnan,
Continuous Univariate Distributions, Second edition, Wiley, 1994.