INTRODUCTION

Recent advancements in MRSI have aimed to achieve high resolution and increased spatial coverage encompassing the whole brain, including the cortical regions near the scalp. However, lipid signals originating from the scalp have continued to pose challenges in the acquisition and quantification of metabolite and macromolecule spectra. We have developed the Fast Lipid signal Processing (FLIP) algorithm¹ which removes subcutaneous lipid contributions in MRSI data based on scalp and brain structural information from MRI. While this approach is very fast for 2D MRSI, scaling up to high resolution 3D data posed difficulty as the number of independent mask voxels in the model grows very large, requiring long processing time. To reduce the 3D EPSI processing time, we aimed to develop a spatial basis with fewer terms without sacrificing lipid removal capacity. Considering the strong lipid signal attenuation with distance from relatively small receiver elements, the spatial domain could be reduced to the region near each receiver coil element, rather than solving the global whole-head geometry. We compared this approach, dubbed FLIP-COIL, with the FLIP algorithm as well as the Papoulis-Gerchberg (PG) algorithm², Hankel Singular Value Decomposition (HSVD)³, and L2 lipid removal⁴.

METHODS

All MR measurements were performed on a 3 T scanner (Skyra, Siemens, Erlangen, Germany) using a 16-channel head receive coil. MRSI (3D) data were acquired using an EPSI sequence⁵ (TE/TR/TR2=16/1551/511 ms, matrix = 50×50×18, FOV = 280×280×180 mm³, 76% GRAPPA fill factor (38/50 lines), no inversion lipid nulling, acquisition time = 17.8 min). Unsuppressed water MRSI data were acquired for coil combination and a concentration reference. T1-weighted MRI was performed using a MPRAGE sequence (TE/TR/TI = 3.98/2000/830 ms, matrix = 176×256×256, FOV = 176×256×256 mm3, GRAPPA factor = 2). Ten subjects (mean age = 33 yrs, F/M=4/6) were scanned, and all were consented in accordance with protocols approved by the University of Kansas Medical Center Institutional Review Board. EPSI data regridding was performed using the MIDAS software.⁶ All lipid removal algorithms were implemented in-house in Matlab (Mathworks, Inc., MA, USA).
The FLIP methodology has been described in our recent work¹. For the FLIP-COIL approach, an additional mask preparation process was used. First, the SNR was determined from the water reference MRSI and then interpolated to the MRI dimensions. An SNR threshold was then found for each coil such that 1500 to 2000 lipid voxels remained, which is ~1/10 of FLIP. Processing times were recorded, and spectra for each algorithm were fitted using LCModel to assess N-Acetylaspartate (NAA), total choline (tCho) and creatine (Cr) concentrations. Fitting performance was evaluated by the Cramer Rao Lower Bound (CRLB) for NAA in four supratrentorial transverse slices. For CRLB assessment, voxels were subject to criteria of tissue fraction ≥50%, linewidth ≤ 0.1 ppm and CRLB of NAA ≤ 20% (all methods). Metabolite concentration ratios were assessed in gray matter (GM) and white matter (WM) regions in a superior slice in 10 subjects, subject to the same voxel inclusion criteria.

RESULTS AND DISCUSSION

The efficiency of lipid removal and variations in the lineshape quality using FLIP-COIL are shown in Fig. 1. Overall, the FLIP based methods (FLIP, FLIP-COIL) performed better in lipid removal than other methods (Fig. 2). NAA/Cr and tCho/Cr values were similar with expected GM and WM differences using all algorithms, (Fig. 2). HSVD and L2 oucomes showed many missing outer cortical voxels indicating failed spectral fitting due to the poor spectral quality (i.e., residual lipid) (Fig. 2).
Spectral quality after lipid removal was evaluated in ten subjects using the Cramer Rao Lower Bound (CRLB) provided by LCModel (Fig. 3). In total, 10,298 voxels passed the quality criteria. The distribution of CRLB showed was most left-shifted, i.e., better spectral quality, with the original FLIP algorithm, while FLIP-COIL showed somewhat worse CRLB but still better than PGA, L2 and HSVD (Fig. 3).
Metabolite concentration ratios of NAA and tCho to Cr were assessed from 343 GM voxels (GM fraction: 67%) and 316 WM voxels (WM fraction: 91%) (Fig. 4). Central WM regions showed higher NAA/Cr and tCho/Cr than those in the outer GM with all algorithms, as expected. Generally, FLIP had the narrowest distributions and fewest outliers with comparable outcomes by FLIP-COIL. The PGA, HSVD and L2 algorithms generally exhibited broader distributions and more outliers (Fig. 4).
Processing times and methodology were previously reported¹ for FLIP, PGA, HSVD and L2. The newly developed FLIP-COIL approach improved the processing time from 40 min (for FLIP) to 11 min.
In summary, we have developed a new approach to accelerate processing of the spatial lipid signal removal for 3D EPSI. The coil-wise processing approach of FLIP-COIL out-performed other frequently used algorithms, and was largely comparable with the FLIP benchmark. By restricting the basis to match the expected signal components from each coil and drastically reducing the matrix size, FLIP-COIL could achieve approximately 4 times acceleration of processing speed with minimal impact on lipid removal performance of FLIP.

Acknowledgements

This study is partially supported by NIH (R01 AG060050). The Hoglund Biomedical Imaging Center is supported by the NIH (S10RR029577) and the Hoglund Family Foundation.