HyperFitS -- Hypernetwork Fitting Spectra for metabolic quantification of 1H MR spectroscopic imaging
Authors: Paul J. Weiser, Gulnur Ungan, Amirmohammad Shamaei, Georg Langs, Wolfgang Bogner, Malte Hoffmann, Antoine Klauser, Ovidiu C. Andronesi
Organizations: Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, USA · Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA · Computational Imaging Research Lab – Department of Biomedical Imaging and Image-Guided Therapy, Medical University of Vienna, Vienna, Austria · Electrical and Software Engineering, University of Calgary, Calgary, Canada · Hotchkiss Brain Institute, University of Calgary, Calgary, Canada · High Field MR Center – Department of Biomedical Imaging and Image-Guided Therapy, Medical University of Vienna, Vienna, Austria · Advanced Clinical Imaging Technology, Siemens Healthineers International AG, Lausanne, Switzerland
Purpose: Proton magnetic resonance spectroscopic imaging (1H MRSI) enables the mapping of whole-brain metabolites concentrations in-vivo. However, a long-standing problem for its clinical applicability is the metabolic quantification, which can require extensive time for spectral fitting. Recently, deep learning methods have been able to provide whole-brain metabolic quantification in only a few seconds. However, neural network implementations often lack configurability and require retraining to change predefined parameter settings. Methods: We introduce HyperFitS, a hypernetwork for spectral fitting for metabolite quantification in whole-brain 1H MRSI that flexibly adapts to a broad range of baseline corrections and water suppression factors. Metabolite maps of human subjects acquired at 3T and 7T with isotropic resolutions of 10 mm, 3.4 mm and 2 mm by water-suppressed and water-unsuppressed MRSI were quantified with HyperFitS and compared to conventional LCModel fitting. Results: Metabolic maps show a substantial agreement between the new and gold-standard methods, with significantly faster fitting times by HyperFitS. Quantitative results further highlight the impact of baseline parametrization on metabolic quantification, which can alter results by up to 30%. Conclusion: HyperFitS shows strong agreement with state-of-the-art conventional methods, while reducing processing times from hours to a few seconds. Compared to prior deep learning based spectral fitting methods, HyperFitS enables a wide range of configurability and can adapt to data quality acquired with multiple protocols and field strengths without retraining.
Figures & tables
Figure 1: HyperFitS: Top: The HyperFitS strategy showing a hypernetwork taking baseline and water suppression correction parameters and predicting weights for a quantification network with a physics-based encoder. Bottom: The physical model combining metabolite and macromolecule basis set, baseline spline functions, and water suppression correction factor used to predict the spectral fit.
Figure 2: A qualitative comparison showing metabolic maps of NAA+NAAG, Cr+PCr, GPC+PCh, inositol, and Glu+Gln, at 7T quantified with HyperFitS and LCModel. Individual representative spectra are displayed below. Processing times for a single MRSI volume of HyperFitS and LCModel are 10sec vs ∼ 1h for 3.4mm, and 90sec vs ∼ 8h for 2mm
Figure 3: Metabolic maps of NAA+NAAG, Cr+PCr, GPC+PCh, Inositol, and Glu+Gln quantified with HyperFitS and LCModel from 3D MRSI acquired at 3T in a glioma patient. Maps are compared for different baseline modeling. Individual representative spectra are displayed below.
Figure 4: Top: A representative spectrum fitted by varying the hyperparameters for baseline flexibility (vertical) and spline node distance (horizontal) in increments of 101 of the total range. Three cases are shown enlarged on the right. The original spectrum is white, fit is red, baseline is yellow, and spline nodes are blue. Bottom: Contour plots of metabolite concentrations. A subject acquired at 7T with 3.4mm resolution was evaluated for 100 different baseline configurations. Baseline flexibility and spline node distance were changed over the hyperparameter range with increments of 101 of the total range. Values show the difference to the center value (0.5 flexibility, 0.15 ppm node distance) in percent.
Figure 5: Metabolite concentration boxplots for metabolites NAA+NAAG, Cr+PCr, PCh+GPC, Glu+Gln, Ins, Tau, GSH, and GABA. Evaluations are performed for 7T (top) and 3T (bottom). The 7T was evaluated with water suppression correction by HyperFitS. The 3T was processed with different baseline configurations by HyperFitS. Comparison to LCModel is included.
Figure 6: Uncertainty maps for the metabolites NAA, Glu and Ins at 7T (left) and 3T (right). LCModel uncertainties are computed as CRLBs, HyperFitS uncertainties are shown with Aleatoric and Epistemic uncertainty.
Figure 7: Correlation matrix of in-vivo metabolite concentrations fitted by HyperFitS at a.) 7T and b.) 3T. c.) A spectrum acquired at 7T with its fit and individual metabolite components.
7 Tesla
3 Tesla
15.0/4.0
10.0/3.0
7.0/4.0
24.0/4.0
24.0/3.0
19.0/4.0
16.0/4.0
8.0/4.0
8.0/4.0
18.0/4.0
16.0/7.0
15.0/4.0
13.0/6.0
8.0/4.0
10.0/4.0
7.0/6.0
9.0/6.0
12.0/4.0
Supplementary Table 1: Percent of significantly (more than 0.3 or less than -0.3) correlated or anti-correlated metabolites for increasing baseline flexibility and node distance. Baseline flexibility parameters from top to bottom: 0.0, 0.5, 1.0. Node distance from left to right: 0.1 ppm, 0.15 ppm, 0.2 ppm. 7T left, 3T right. The first value shows the correlated percentage, the second shows the anti-correlation percentage.
Supplementary Figure 1: Top: Correlation matrix of Monte Carlo analysis for subjects acquired at 3T. Each subject was evaluated for 9 different baseline configurations. Bottom: Boxplots showing correlation values of metabolite pairs that are less than -0.2 or larger than 0.2 in all baseline configurations. Glu & Ins and Glu & Gln are plotted separately as diamond/star. The color selection is based on the median values.
Supplementary Figure 2: Top: Correlation matrix of Monte Carlo analysis for subjects acquired at 7T. Each subject was evaluated for 9 different baseline configurations. Bottom: Boxplots showing correlation values of metabolite pairs that are less than -0.2 or larger than 0.2 in all baseline configurations. Glu & Ins and Glu & Gln are plotted separately as diamond/star. The color selection is based on the median values.
School of Computer Science and Informatics, Cardiff University, Cardiff, CF24 4AG, United Kingdom · Faculty of Science and Engineering, Swansea University, Swansea, SA2 8PP, United Kingdom
School of Biomedical Engineering, Tsinghua University, Beijing, China. · Oxford Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom. · Department of Radiology, West China Second University Hospital, Sichuan University, Chengdu, China. +4