Cramér-Rao Bound Optimized Subspace Reconstruction in Quantitative MRI

TL;DR

This work tackles subspace reconstruction in quantitative MRI by incorporating preservation of the Cramér-Rao bound (CRB) into basis design. It introduces CRB-SVD, an approximate compressed CRB-based basis optimization that integrates orthogonalized derivatives of the signal into a joint energy-CRB objective and uses SVD on a concatenated matrix to produce a CRB-aware basis. Across simulations and in vivo neuroimaging of qMT and MRF-FISP, CRB-SVD achieves improved parameter accuracy and precision at smaller subspace sizes, enabling computational savings. The approach provides a practical upgrade to standard SVD-based subspace reconstruction and offers guidance on parameter weighting and applicability across sequences.

Abstract

We extend the traditional framework for estimating subspace bases that maximize the preserved signal energy to additionally preserve the Cramér-Rao bound (CRB) of the biophysical parameters and, ultimately, improve accuracy and precision in the quantitative maps. To this end, we introduce an \textit{approximate compressed CRB} based on orthogonalized versions of the signal's derivatives with respect to the model parameters. This approximation permits singular value decomposition (SVD)-based minimization of both the CRB and signal losses during compression. Compared to the traditional SVD approach, the proposed method better preserves the CRB across all biophysical parameters with negligible cost to the preserved signal energy, leading to reduced bias and variance of the parameter estimates in simulation. In vivo, improved accuracy and precision are observed in two quantitative neuroimaging applications, permitting the use of smaller basis sizes in subspace reconstruction and offering significant computational savings.

Figures (6)

• Figure 1: (a) Distinguishing one model parameter from another depends on the components of its signal derivative, $\mathbf{j}_i$, that are orthogonal to the span of all other signal derivatives, $\mathbf{J}_i$Scharf1993. (b) Depiction of a representative signal $\mathbf{s}$ for the inversion recovery MRF-FISP sequence Jiang2015a (assuming the scaling $M_0=1$), its derivative with respect to the longitudinal relaxation rate $\partial\mathbf{s}/\partial R_1$, and the corresponding orthogonalized derivative $(\partial\mathbf{s}/\partial R_1)_\perp$, where the components parallel to the derivatives with respect to all other model parameters were removed (i.e., $M_0$ and $R_2$). Note we plot the absolute value and $(\partial\mathbf{s}/\partial R_1)_\perp$ is scaled up to have a unit length, which emphasizes that the first segment is most important for encoding $R_1$.
• Figure 2: (a) Signal energy loss vs $\lambda$ normalized by the total signal energy in the qMT test dataset. (b) Cramér-Rao bound (CRB) loss calculated using the approximate compressed CRB ($\Delta B_{ac}$; Eq. \ref{['eq:acrbloss']}) and the exact compressed CRB ($\Delta B_{ec}$; Eq. \ref{['eq:ccrbloss']}), averaged over all fingerprints and orthogonalized derivatives in the test dataset. (c) CRB loss for $m_0^s$ and $R_\text{x}$ individually at $N_c=15$. (d) Average ratio of approximate compressed CRB to exact compressed CRB over the test dataset (Eq. \ref{['eq:ratio']}). The $\lambda$ values shaded in gray offer significant CRB improvements at a small cost to the signal fidelity for all $N_c$.
• Figure 3: Representative scatter plots of the approximate compressed CRB of the semi-solid spin pool fraction ($B_{ac}(m_0^s)$) vs the exact compressed CRB ($B_{ec}(m_0^s)$) for different numbers of coefficients ($N_c$) and CRB-loss weightings ($\lambda$), where each dot represents a fingerprint in the qMT test set. Note each subplot's axes are limited to the same range for improved visualization. Linear regressions across all data points in each subplot (including those not plotted within the axis limits) are shown in purple in reference to the identity line in red. Correlations between $B_{ac}$ and $B_{ec}$ are improved with increasing $N_c$ and $\lambda$.
• Figure 4: Bias (a--b) and standard deviation (c--d) of non-linear least squares (NLLS)-based $m_0^s$ and $R_\text{x}$ estimates of a typical white matter fingerprint derived from the traditional SVD ($\lambda=0$; blue) and CRB-SVD ($\lambda=0.5$; red) bases as a function of the number of coefficients $N_c$. Both metrics are normalized by the ground-truth $m_0^s$ and $R_\text{x}$ values, and the reference lines in (c--d) indicate the uncompressed CRB (Eq. \ref{['eq:tcrb']}) and exact compressed CRB (Eq. \ref{['eq:ccrb']}). Lower bias and variance more closely resembling minimum variance unbiased estimation is observed with the CRB-SVD basis, particularly for $N_c<18$, where the traditional SVD basis performs especially poorly.
• Figure 5: Comparison of non-linear least squares fits of the semi-solid spin pool's fractional size $m_0^s$ (a) and magnetization exchange rate $R_\text{x}$ (b) extracted from the traditional SVD ($\lambda=0$) and proposed CRB-SVD basis ($\lambda=0.5$) for different numbers of coefficients ($N_c$), where $N_c=30$ serves as a gold standard. The CRB-SVD basis improves the precision of the $m_0^s$ (magnifications) and the accuracy of the $R_\text{x}$ maps for small $N_c$.