High-resolution B0 mapping suffers from long scan time, and issues with phase-wraps. We present an acquisition and reconstruction technique that resolves both problems. We utilize “X” sampling in k-TE space, in which multiple phase-encoding lines are acquired exactly twice per TR. The echo spacing is shortest for central k-space and largest for outer k-space. A multi-scale reconstruction enables pixel-wise phase unwrapping. This technique may be particularly useful for quantitative susceptibility mapping (QSM), as it could a) shorten scan time while maintaining the sensitivity to high-order field variation and b) simplify phase-unwrapping, which are the key features of interest in QSM.
Figure 1 shows the proposed sampling scheme. Positive and negative halves of $$$k$$$-space are divided into $$$N$$$ bins. In each TR, $$$N$$$ phase-encoding (PE) lines are acquired, each exactly twice. The echo time difference (ΔTE) increases with the magnitude of $$$k_{PE}$$$.
Field estimation is performed at multiple spatial scales. The first (coarsest) scale corresponds to the central k-space bin, which has the shortest ΔTE. Figure 2 shows the first iteration. Step 1: Bin 1 is extracted and the other regions are zero-filled. The early-TE (red dots) and late-TE (blue dots) samples are inverse Fourier transformed to $$$I_{a,1}$$$ and $$$I_{b,1}$$$. Step 2: $$$\Delta B_{0}^{1}$$$ is estimated: $$$\Delta B_{0}^{1} = \frac{\angle (I_{a,1}\cdot I_{b,1}^{*})}{\gamma 2\pi \Delta TE_{1}}$$$. Note that $$$\Delta B_{0}^{1}$$$ is free of wraps when $$$\Delta TE_{1}$$$ is smaller than the reciprocal of the resonant frequency range. Step 3: The phase of $$$I_{a,1}$$$ and $$$I_{b,1}$$$ are adjusted using $$$\Delta B_{0}^{1}$$$ to match the echo times of bin 2: $$$I_{a,1}^{'} = I_{a,1}\cdot e^{+j\gamma 2\pi \Delta B_{0}^{1}\tau }$$$, $$$I_{b,1}^{'} = I_{b,1}\cdot e^{-j\gamma 2\pi \Delta B_{0}^{1}\tau }$$$ Step 4: Bin 1 is updated by the Fourier transform of $$$I_{a,1}^{'}$$$ and $$$I_{b,1}^{'}$$$ (red and blue squares). Bin 1 and bin 2, now with the same effective echo times, are combined for the next iteration. In the $$$i$$$th iteration, when $$$\Delta B_0^{i}-\Delta B_0^{i-1}$$$ is less than the reciprocal of $$$\Delta TE_i$$$, pixel-by-pixel phase-unwrapping can be performed perfectly using $$$\Delta B_0^{i-1}$$$ as a reference. After $$$N$$$ iterations, $$$\Delta B_{0}^{N}$$$ is the result of multi-scale estimation. As a final step, a non-linear optimization is performed using $$$\Delta B_{0}^{N}$$$ as the initial guess: $$\Delta B_{0}^{nonlin}\left ( \boldsymbol{r} \right )=\arg min_{\Delta B_0 \left ( \boldsymbol{r} \right )} \left \| \mathcal{F}_u \{\hat{M}\left ( \boldsymbol{r} \right ) e^{-j\left ( \hat{\psi _0}\left ( \boldsymbol{r} \right )+\gamma 2\pi \Delta B_0 \left ( \boldsymbol{r} \right )TE \right ) } \}-S\left ( \boldsymbol{k},TE \right )\right \|_2^2$$, where $$$\mathcal{F}_u$$$ represents the “X”-pattern Fourier under-sampling in $$$k$$$-TE space, $$$\hat{M}\left ( \boldsymbol{r} \right )$$$ is the estimated magnitude image, $$$\hat{\psi _0}\left ( \boldsymbol{r} \right )$$$ is the phase at TE = 0 estimated from the central $$$k$$$-space bin, $$$S\left ( \boldsymbol{k},TE \right )$$$ is multi-echo $$$k$$$-space data.
Data: 3D Multi-echo GRE data were synthesized using:$$m\left ( \boldsymbol{r},TE \right )=\rho \left ( \boldsymbol{r} \right ) e^{-TE/T_2^{*}\left ( \boldsymbol{r} \right )} e^{-j\left ( \psi _0 \left ( \boldsymbol{r} \right )+\gamma 2\pi \Delta B_0 \left ( \boldsymbol{r} \right )TE \right )}+n\left (\boldsymbol{ r},TE \right )$$, where proton density $$$\rho \left ( \boldsymbol{r} \right )$$$, $$$T_2^{*}$$$ map $$$T_2^{*}\left ( \boldsymbol{r} \right )$$$, low-resolution phase offset $$$\psi _0 \left ( \boldsymbol{r} \right )$$$, and field map $$$\Delta B_0 \left ( \boldsymbol{r} \right )$$$ were taken from a fully-sampled multi-echo GRE scan of a healthy volunteer. Acquisition parameters: 3 Tesla GE Signa HD23, spatial resolution = 0.5 mm (AP) x 0.5 mm (RL) x 1 mm (SI), TE = 5,10,15,20 ms, TR = 50 ms, BW = ±62.5kHz. $$$n\left ( r,TE \right )$$$ was i.i.d. bivariate gaussian noise scaled to make the white matter SNR equal 40. Simulated acquisition: We simulated a 5-bin acquisition ($$$N=5$$$, divided along $$$k_y$$$), and ten echo times: 5,7,9,11,13,15,17,19,21,23 ms. The $$$k_z$$$ phase-encoding direction was fully-sampled. Evaluation: Projection onto dipole field (PDF)4,5 was used to extract the high-order field variation originating from tissue susceptibility difference (“local field”). Both the total $$$\Delta B_0$$$ and the "local field" of the estimation were evaluated.
Results & Discussion
Figure 3 shows multi-scale estimation in a noiseless case. Fine structures were iteratively revealed in the reconstruction. Figure 4 shows the case with noise. Non-linear optimization has less error, indicating more noise-resistance than multi-scale estimation alone. In the "local field" evaluation, non-linear optimization reliably reconstructed the field around tissue structures (e.g. basal ganglia nuclei). Poor depiction of the field near through-plane veins (red arrows) was observed. This may affect susceptibility quantification near these veins.1. Haacke EM, et al. Quantitative susceptibility mapping: current status and future directions. Magnetic resonance imaging. 2015 Jan 31;33(1):1-25.
2. Haacke EM, Brown RW, Thompson MR, Venkatesan R. Magnetic resonance imaging: physical principles and sequence design. 1st ed. Wiley-Liss; 1999.
3. Wang Y, Liu T. Quantitative susceptibility mapping (QSM): decoding MRI data for a tissue magnetic biomarker. Magnetic resonance in medicine. 2015 Jan 1;73(1):82-101.
4. Liu, T, et al. A novel background field removal method for MRI using projection onto dipole fields (PDF). NMR in Biomedicine 24.9 (2011): 1129-1136.
5. de Rochefort L, et al. Quantitative susceptibility map reconstruction from MR phase data using bayesian regularization: validation and application to brain imaging. Magnetic resonance in medicine. 2010 Jan 1;63(1):194-206.