Achieving large coverage volume, while maintaining low lipid contamination of the data by manually placing sat bands is time-consuming and requires a highly-skilled operator. Previously, Osorio et al developed an OVS scheme using octagonal selection with cosine-modulated VSS pulses (19
). This allowed an increase in coverage within plane, but the curvature of the skull meant that it was only sufficient for single-slice acquisitions. Nevertheless, we found that using the octagonal sat band configuration together with manual sat bands (OCTAGONAL protocol) in place of a rectangular configuration (STANDARD protocol) improved the lipid suppression in the corners of the volume of interest in multi-slice MRSI acquisitions and allowed the operator to place a larger PRESS box closer to the edges of the brain, improving brain coverage.
Hovdebo and Ryner developed a technique that achieved non-cuboidal excited volume for single-voxel MRS by iteratively simplifying the polyhedral shape and using its faces to prescribe sat bands (26
). 3D sat band optimization based on image data for spectroscopic imaging was introduced by Li et al in their 2006 ISMRM abstract (18
) and later elaborated in a paper by Martinez-Ramon (28
While initially based on the same idea of maximizing the number of lipid pixel and minimizing the number of brain pixels covered by the sat bands, described in (18
), the technique used in our study had several major differences from the one implemented in (28
One difference of our algorithm was that optimization was performed in several steps on subsets of parameters. Resulting values from one step were used as starting conditions for the next one. This made the algorithm find the optimal solution faster and, more importantly, prevented it from getting stuck in a local minimum. (red, blue) shows how the algorithm converged on an optimal solution in a reasonable number of iterations. The other two curves (green, purple) show what would happen if all 36 unknowns were optimized simultaneously, using the same cost function and starting with the same initial configuration. While it was able to achieve some initial improvement, the optimization got stuck in a configuration that was likely a local minimum of the cost function and was not able to make much improvement after that. Also the cost function (eq. 1
) contained the regularization term that discouraged configurations where no brain or lipid pixels were covered and the sat bands were far away from the origin. This obviated the need for a separate coarse placement of sat bands as in (28
), since the initial distances to the sat bands could be calculated as a part of multi-step optimization with the same cost function.
Fig. 3 Convergence of the optimization algorithm: blue and red - number of lipid and brain pixels, covered by sat bands using a multi-stage algorithm; green and purple - number of covered lipid and brain pixels using an algorithm in which all parameters are (more ...)
Another difference of this technique was the use of linear interpolation of covered volume to achieve a continuous cost function, since the number of covered pixels was inherently discrete. Even with interpolation, the cost function was not differentiable, so the gradient-based optimization methods could get stuck in areas where the gradient was not continuous. Instead, we chose to use Nelder-Mead simplex method, which did not use the gradient of the cost function.
A major part of the study was development of the pulse sequence that could use additional calculated sat bands. While the technique in (28
) calculated the placement of 16 sat bands, only 8 could be used by the pulse sequence. In contrast, the pulse sequence used in the in-vivo experiments, described in this paper, used 9 calculated sat bands in addition to 10 fixed bands in octagonal configuration, implemented using cosine-modulated RF pulses. This allowed to effectively cover the top of of the head, as well as provide additional suppression in the corners of the PRESS box.
The study in (28
) presented MRSI data from six healthy volunteers and showed sat band placement placement very similar to that of a highly-trained operator. The techniques developed in our study were able to achieve more optimal placement of sat bands than human operators did. That made possible to acquire good quality 3D MRSI data from a much larger volume of the brain than the current protocols. Optimal placement of the sat bands allowed the acquisition of MRSI data from tissues near the edge of the brain that were previously difficult to cover due to the curvature of the skull. It also ensured effective lipid suppression and allowed the PRESS box to extend beyond the boundary of the brain. Automatic placement of the sat bands was also designed to reduce the variability in the quality of the scan that is inherent for manual prescription. Repeated acquisition of data from the same subject with different head position and different operators () shows very good reproducibility in terms of data quality and covered volume.
To verify that the technique works robustly in a clinical setting and allows for larger coverage volumes, we added an additional MRSI acquisition to the routine exams of patients with brain tumors. The data quality parameters (metabolite peak SNR, peak width, metabolite ratios) were calculated within the areas of healthy white matter to measure the effects of the technique itself on data quality as opposed to the variation in the parameters caused by the disease. These data show no compromise in data quality, while achieving significantly larger coverage volume. It was observed that the technique would bring the most benefit to the cases with multi-focal and heterogenous tumors. shows one such case, which would have required two standard MRSI acquisitions to cover both disease sites, while the AUTOSAT+OCTAGONAL covered both areas of interest at the same time.
One limitation of the study was that the authors had no control over the population of patients, scheduled to receive scans at our facility, so few patients in the study had challenging tumors like this. Also, due to experimental nature of the protocols with automatic sat bands, acquired data was not used by the physicians treating the patients. Future studies will assess the effects of the technique on radiological interpretation of tumor spectra.
The protocols with automatic sat band placement may add time to the patient exam. The optimization algorithm can be made to work with standard T1-weighted images, making it unnecessary to acquire an extra image series for sat band optimization. Calculation for the sat band placement took around 3 minutes (2.8 GHz Intel Xeon computer), which was usually less time consuming than adjusting sat bands manually and was usually run while another image series was being acquired.
The largest limiting factor encountered in our study was signal drop-off in the anterior of the lower slices, presumably due to field inhomogeneity caused by tissue-air interface in sinuses. This meant that tumors below the level of the eye-sockets could not be covered. This effect can be reduced by placing the PRESS box obliquely and performing high-order shimming over an oblique volume.
In conclusion, the technique described in this study has helped solve two of the most significant problems with brain PRESS MRSI acquisitions: limited brain coverage and difficulty in prescription. The improved coverage will be useful for evaluating heterogeneous and infiltrative tumors, as well as tumors at the periphery of the brain, which are difficult to evaluate with current protocols. It should make possible a more accurate assessment of the progression of tumors in serial studies. The use of this technique reduces the need for extensive operator training, thus facilitating wider utilization of MRSI in the clinical setting.