Poster R10 - An Updated Spectral Fitting Strategy for Magnetic Resonance Spectroscopic Imaging

Introduction: 1H Magnetic Resonance Spectroscopic Imaging (MRSI) can “provide information on tumor metabolism and heterogeneity” not possible with single voxel 1H MRS, providing clinicians with novel information to determine tumor cellularity/grading, treatment planning, and treatment effectiveness (1). However, applying MRSI data in clinical decision-making can be difficult, as the 1H spectroscopic signal suffers from innately low SNR and spectral overlap, issues that are only exacerbated in the transition from MRS to MRSI, where SNR is lower due to worse artifact suppression and higher spatial resolution.
Therefore, the development of quantification algorithms dedicated to MRSI is vital. There are already several packages that offer an MRSI fitting mode (3-7). However, most packages were created for single-voxel MRS, and so fail to incorporate spatial prior information that can define the MRSI problem space (3-4). Moreover, the few spectral fitting packages that consider MRSI-specific prior information, have stopped development (6-7) or use a simpler physics model of the signal for simulation (5-7). For this purpose, we have developed our in-house ProFit-1D (2,8) software, which has demonstrated superior accuracy and similar precision for spectral fitting of MRS data in comparison to the golden standard (2,9), LCModel (6), for the fitting of MRSI data by establishing the spatial relationship for confounding fitting parameters (Phase & Gaussian line-width), while keeping concentration-related (Lorentzian line-shape & Amplitude) parameters unweighted to avoid bias. We compare our results with LCModel to establish a baseline comparison of ProFit’s fitting capabilities as it applies to MRSI.

Materials and

Methods: ProFit-1D: Figure 1 shows the original ProFit-1D algorithm (2,8). To incorporate spatial prior information, the ProFit algorithm was first used on 7T proton 2D FID-MRSI datasets (12,13) to approximate the spatial distribution of the confounding parameters. The spatial covariances are determined using variogram estimation (7,10) (figure 2). They are then used to weigh our custom cost function component, σ^2 ∑_(s~t)▒‖X_s-X_t ‖_W^2 , incorporated to maximize the joint probability of a voxel and its neighbors. This cost function is combined with the R3 component used in Profit-1D (2) to create the final cost function: R_N=∑_m^M▒R_3^2 +σ^2 ∑_(s~t)▒‖X_s-X_t ‖_W^2 . This is used to optimize the parameter vector X using non-linear least square optimization for a site of voxels N with 4x4 voxels in a block-ICM based approach (11). This block-ICM step replaces the second iteration in the profit-1D algorithm, after which the Profit-1D algorithm continues a voxel-by-voxel fitting for final spectral constraints and baseline optimization. A schematic of the proposed algorithm is provided in figure 3.
Database: Data shown is from a patient with an IDH mutant grade III Oligodendroglioma (figure 5). Additional testing was performed on three other datasets (one patient and two healthy volunteers).
LCModel: The same basis set (figure 1) was used with the default spline flexibility.
Visualization & Comparison: Figure 4 depicts the parameter maps before and after the block-ICM step. While figure 5 presents the fit and error maps from ProFit & LCModel for four metabolites of interest. For reference, figure 5 contains anatomical scans (A-B).

Results, Conclusions, and Discussions: Results & Discussion
Examining figure 4, we see that the block iterative conditioning step has a visual effect on the global (non-metabolite specific) parameter maps, with the zero-order and first-order phase correction maps displaying less variance in the whole-brain parameter maps and less neighboring-voxel variation. Although the change in the gaussian line-broadening map appears insignificant.
In figure 5, we see that this imposed smoothness on the confounding parameters is reflected by the patient’s metabolite distributions. The ProFit concentration maps are smoother and less noisy when compared to the LCModel concentration maps, while still clearly denoting known white matter-gray matter distributions and the tumor area (arrows) on the patient. When comparing the minimal variance (CRLB) maps, we see that the metabolites are fitted within reasonable bounds for both methods, albeit ProFit does trend higher on the bordering areas and for the Glx distribution. Which may reflect the difficulty for ProFit in fitting lower quality spectra (the borders). Alternatively, the higher variance in the bordering regions may indicate differences in CRLB calculations or the need to establish an anisotropic region-specific definition of spatial covariance. As the true covariance between voxels may differ in these regions than what was estimated with the variogram, which assumes stationarity and isotropic random variables.
Conclusion
In this preliminary study, we show that ProFit-1D yields smoother metabolite maps, similar minimum variance distribution, and a more complete whole brain fitting. Improvements corroborated via the additional test datasets not pictured here. However, for future study, the inclusion of more data sets is still required to determine the robustness of the software. Either through more in-vivo datasets or a synthetic MRSI strategy for ground truth validation to determine the accuracy of ProFit vs LCModel in MRSI.

Acknowledgements (Optional): Acknowledgements
This work was performed at the Advanced Imaging Research Center at the University of Texas Southwestern Medical Center Dallas. This work was supported by the Cancer Prevention and Research Institute of Texas (CPRIT) grant/ RR180056. ProFit-1D was further developed from the initial efforts of Tamas Borbath1,7.

References
1. Laino ME, Young R, Beal K, Haque S, Mazaheri Y, Corrias G, et al. Magnetic resonance spectroscopic imaging in gliomas: clinical diagnosis and radiotherapy planning. BJR Open 2020;2;20190026.
2. Borbath T, Murali-Manohar S, Dorst J, Andrew W, Henning A. ProFit-1D – A 1D fitting software and open-source validation data sets. Magnetic Resonance in Medicine. 2021;00;1-20.
3. Oeltzchner G., Zollner H J, Hui S C.N., et al. Osprey: Open-source processing, reconstruction & estimation of magnetic resonance spectroscopy data. Journal of Neuroscience Methods. 2020;343.
4. Stefan, D., Cesare, F. D., Andrasescu, A., Popa, E., et al. Quantitation of magnetic resonance spectroscopy signals: The jMRUI software package. Measurement Science and Technology, 2009;20(10);104035.
5. Maudsley A. A., Darkazanli A., Alger J. R., Hall L. O., Schuff N., Studholme C., et al. Comprehensive processing, display and analysis for in vivo MR spectroscopic imaging. NMR Biomed. 2006;19;492–503.
6. Provencher S.W. Estimation of Metabolite Concentrations from Localized in Vivo Proton NMR Spectra. Magnetic Resonance in Medicine. 1993;30(6);672-679
7. Kelm M.B., Kaster F.O., Henning A., Webner M., Bachert P., et al. Using spatial prior knowledge in the spectral fitting of MRS images. NMR in Biomedicine. 2012;25;1-13.
8. Borbath T, Murali-Manohar S, Dorst J, Andrew W, Henning A. ProFit-v3: Accuracy and precision evaluation of a new spectral fitting software. International Society for Magnetic Resonance in Medicine. 2021;2004
9. Chan KL, Borbath T, Sherlock S, Maher EA, Patel TR, Henning A. ProFit-1D for quantifying J-difference edited data at 3T. Proceedings of the 2024 annual scientific meeting of ISMRM, May 4th – 9th 2024,Singapore.
10. Cressie N., Hawkins D.M. Robust Estimation of the Variogram: I. Mathematical Geology. 1980;12(2);115-125
11. Kelm M., Menze B.H., Nix O., Zechmann C.M., Hamprecht F.A. Estimating Kinetic Parameter Maps From Dynamic Contrast-Enhanced MRI Using Spatial Prior Knowledge. IEEE Transactions On Medical Imaging, 2009;28(10);1534-1547.
12. Chan KL, Ziegs T, Henning A. Improved signal-to-noise performance of MultiNet GRAPPA 1H FID MRSI reconstruction with semi-synthetic calibration data. Magnetic Resonance in Medicine. 2022;88(4);1500-1515
13. Chan KL, Henning A. MultiNet CAIPIRINHA: accelerated H MRSI with 1-step neural network reconstruction based on augmented MRSI training data. Proceedings of the 30th annual scientific meeting of ISMRM, May 7th – 12th 2022,London, UK
14. Ziegs T, Wright AM, Henning A. Test-retest reproducibility of human brain multi-slice 1H FID-MRSI data at 9.4T after optimization of lipid regularization, macromolecular model, and spline baseline stiffness. Magnetic Resonance in Medicine. 2023;89(1);11-28.
15. Soher B, Semanchuk P, Todd D, Ji X, Deelchand D, Joers J, Oz G and Young K. Vespa: Integrated applications for RF pulse design, spectral simulation and MRS data analysis. Magnetic Resonance in Medicine. 2023;1-16.
16. Simpson R, Devenyi GA, Jezzard P, Hennessy TJ, Near J. Advanced processing and simulation of MRS data using the FID appliance (FID-A)-An open source, MATLAB-based toolkit. Magnetic Resonance in Medicine. 2017;77(1);23-33.