4241
No-Cost, Vendor-Agnostic MR Spectroscopy Fitting Software Can Predict Diagnostic Quality and Utility of Clinical Brain Studies1Department of Imaging Physics, The UT MD Anderson Cancer Center, Houston, TX, United States, 2Department of Diagnostic Radiology, The UT MD Anderson Cancer Center, Houston, TX, United States
Synopsis
MR spectroscopy (MRS) allows for non-invasive assessment of tissue metabolites and is rapidly expanding in clinical use. Current MRS QC recommendations, however, are phantom-based and infrequent. This study investigated the feasibility of using no-cost, vendor-agnostic fitting software to provide quantitative quality assessment of patient brain MRS. We demonstrated that freely-available MRS fitting software can stratify spectral diagnostic quality and utility. This technique can be used to provide the clinician with an estimate of quality for each patient spectrum. Additionally, this approach is amenable to real-time patient MRS quality verification and longitudinal monitoring of protocol performance.
Purpose
MR spectroscopy (MRS) is an established clinical technique for the non-invasive assessment of tissue metabolites with a wide array of applications, such as the assessment of treatment response of glioma patients (1). Current MRS quality control (QC) recommendations are periodic and phantom-based, which does not provide any quantitative metrics applicable to benchmarking patient-specific spectra for interpretation (2,3). Additionally, it is unknown how spectral parameters such as signal-to-noise ratio (SNR) and full width at half maximum (FWHM) affect clinical utility and diagnostic quality. This information is critically linked to clinical decision making, but there is currently little information published on assessing the quality of MRS in patients on clinical systems, though machine learning has been used to predict MRS diagnostic utility (4). This study investigated the feasibility of using no-cost, vendor-agnostic fitting software to provide automated and quantitative quality assessment of patient MRS.Methods
All patient data were acquired in compliance with 45 CFR 46.101(b)(4). Patient brain lesion and contralateral MRS data (single voxel) were acquired from a GE Discovery MR750 (n=16) and Siemens MAGNETOM Prisma (n=16) using a PRESS sequence with similar acquisition parameters including a 144 ms echo time. Spectra were processed with TARQUIN (http://tarquin.sourceforge.net/) using the default basis sets and parameters. For this study, no eddy current correction was performed. We investigated spectral parameters that should be sensitive to the most common MRS failure modes including the Q factor (standard deviation of the fit residual divided by the standard deviation of the noise; Fig. 1A), FWHM (Fig. 1B) and SNR (Fig. 1C) of the largest metabolite peak, and deviation of the fitted baseline from a straight line (BL; Fig. 1D). A Likert-type scale was employed to define diagnostic quality (1-5; poor/fair/good/very good/excellent) and utility (0-1; non-diagnostic/diagnostic) as determined by an experienced neuroradiologist. Significant differences were determined using Wilcoxon rank-sum tests.Results
Values were successfully measured for all parameters on both systems. Inter-patient variability typically dominated over inter-scanner variability and spectral parameters were similar between systems. 54 of 64 spectra were considered high quality (quality≥3) and the remaining 10 were low quality. 61 of 64 spectra were considered diagnostic with the remaining 3 considered non-diagnostic. SNR was determined to be the best parameter to stratify diagnostic quality (Fig. 1A) and FWHM was found to be the best parameter to stratify diagnostic utility (Fig. 1B). FWHM, SNR, and BL were significantly different between high- and low-quality spectra with p values of 0.039, <0.001, and 0.001, respectively. Q was not significantly different (p=0.616). While the number of non-diagnostic spectra was low, FWHM and BL were both significantly different between diagnostic and non-diagnostic spectra (p<0.001 for both).Discussion
Spectral quality was evaluated for 32 clinical studies (64 spectra) on two scanner models. FWHM, SNR, and BL were found to be highly sensitive to spectral quality. One spectrum with reduced quality due to lipid infiltration from the scalp was not identified with these parameters. Diagnostic utility was only impaired for spectra with large FWHM and BL, which is indicative of poor water suppression, poor shim, and/or large susceptibility gradients. Interestingly, spectra that did not meet the quality standards put forth by the ISMRM MRS Consensus Group (5) were regularly considered diagnostic (and occasionally high-quality) by the experienced radiologist, suggesting that these standards may need to be re-evaluated. In addition to increasing the sample size, future studies will extend this technique to multi-voxel MRS. We are also moving towards clinical implementation of automated MRS QC that will allow for patient spectral quality verification and rapid identification of scanner malfunctions.Conclusions
Freely-available MRS fitting software successfully stratified spectral diagnostic quality and utility for both systems. This technique could be used to provide the clinician with an estimate of quality for each patient spectrum. Additionally, this technique could be further developed for real-time patient MRS quality verification and longitudinal monitoring of protocol performance.Acknowledgements
The authors thank Wei Wei for assistance with data analysis.References
1. van Dijken BRJ, van Laar PJ, Holtman GA, van der Hoorn A. Diagnostic accuracy of magnetic resonance imaging techniques for treatment response evaluation in patients with high-grade glioma, a systematic review and meta-analysis. Eur Radiol 2017;27:4129–4144 doi: 10.1007/s00330-017-4789-9.
2. Drost DJ, Riddle WR, Clarke GD. Proton magnetic resonance spectroscopy in the brain: Report of AAPM MR Task Group #9. Med Phys 2002;29:2177–2197 doi: 10.1118/1.1501822.
3. Jackson EF, Bronskill M, Drost DJ, et al. Acceptance testing and quality assurance procedures for magnetic resonance imaging facilities. American Association of Physicists in Medicine. College Park, MD, USA; 2010.
4. Kyathanahally SP, Mocioiu V, Pedrosa de Barros N, et al. Quality of clinical brain tumor MR spectra judged by humans and machine learning tools. Magn Reson Med 2018;79:2500–2510 doi: 10.1002/mrm.26948.
5. Öz G, Alger JR, Barker PB, et al. Clinical Proton MR Spectroscopy in Central Nervous System Disorders. Radiology 2014;270:658–679 doi: 10.1148/radiol.13130531.
Figures
