Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2023 Jan;89(1):40-53.
doi: 10.1002/mrm.29446. Epub 2022 Sep 25.

In vivo magnetic resonance 31 P-Spectral Analysis With Neural Networks: 31P-SPAWNN

Julien Songeon  1 , Sébastien Courvoisier  1   2 , Lijing Xin  2   3 , Thomas Agius  4 , Oscar Dabrowski  1 , Alban Longchamp  4 , François Lazeyras  1   2 , Antoine Klauser  1   2
Affiliations

In vivo magnetic resonance 31 P-Spectral Analysis With Neural Networks: 31P-SPAWNN

Julien Songeon et al. Magn Reson Med. 2023 Jan.

Abstract

Purpose: We have introduced an artificial intelligence framework, 31P-SPAWNN, in order to fully analyze phosphorus-31 ( 31 $$ {}^{31} $$ P) magnetic resonance spectra. The flexibility and speed of the technique rival traditional least-square fitting methods, with the performance of the two approaches, are compared in this work.

Theory and methods: Convolutional neural network architectures have been proposed for the analysis and quantification of 31 $$ {}^{31} $$ P-spectroscopy. The generation of training and test data using a fully parameterized model is presented herein. In vivo unlocalized free induction decay and three-dimensional 31 $$ {}^{31} $$ P-magnetic resonance spectroscopy imaging data were acquired from healthy volunteers before being quantified using either 31P-SPAWNN or traditional least-square fitting techniques.

Results: The presented experiment has demonstrated both the reliability and accuracy of 31P-SPAWNN for estimating metabolite concentrations and spectral parameters. Simulated test data showed improved quantification using 31P-SPAWNN compared with LCModel. In vivo data analysis revealed higher accuracy at low signal-to-noise ratio using 31P-SPAWNN, yet with equivalent precision. Processing time using 31P-SPAWNN can be further shortened up to two orders of magnitude.

Conclusion: The accuracy, reliability, and computational speed of the method open new perspectives for integrating these applications in a clinical setting.

Keywords: LCModel; convolutional neural network; deep learning; in vivo; phosphorus magnetic resonance spectroscopy.

PubMed Disclaimer

Figures

FIGURE 1
FIGURE 1
Data analysis flowchart illustrating the different steps of the Spectral Analysis With Neural Networks (SPAWNN) pipeline. The method uses three convolutional neural networks (CNN) to estimate the metabolite concentrations, spectrum parameters, and baseline. Data preparation consists of spectrum normalization, with windowing over the range of 10 to −25 ppm. Padding (mirror replication of the first and last 5 points of the array) and stacking of the real and imaginary parts (switch from a complex array of 1×899 points to a real array of 2×909 points by separating the real and complex part of each point) were then applied.
FIGURE 2
FIGURE 2
SPAWNN‐Quantification (SPAWNN‐Q) model architecture for metabolite concentration and parameter estimation. The convolutional neural network takes the spectrum as an input layer and performs six successive steps of convolution, ReLU activation, and pooling. Then, the five final steps are fully connected layers with PReLU activation. δ is equal to 20 for the parameters, and 14 for the metabolite concentration estimations.
FIGURE 3
FIGURE 3
SPAWNN‐Baseline (SPAWNN‐Bl) model architecture for baseline estimation. The U‐Net takes the spectrum as an input layer and performs three down‐sampling steps with convolution and PReLU activation. Then, it performs three up‐scaling steps with convolution and PReLU activation. At each up‐scaling, the new layer is concatenated with the corresponding down‐sampling layer. The output is the estimation of the baseline and has the same dimension as the input.
FIGURE 4
FIGURE 4
Results on simulated data. (A) Examples of SNRmean for a simulated spectrum with no noise (top), SNRmean of 4 (middle), and SNRmean of 2 (bottom). (B) Comparison of the coefficient of determination R2 with bootstrapping between SPAWNN and LCModel for each metabolite. A new dataset of 104 spectra was created to compare the two methods. (C) SPAWNN's coefficient of determination R2 for each metabolite concentration estimation as a function of the SNRmean range.
FIGURE 5
FIGURE 5
Comparison of in vivo 31P‐MRSI spectrum (TR = 1.5 s, TE = 0.5 ms, sum of 8 voxels with 24 weighted averages each) evaluated and reconstructed with SPAWNN (left) and fitted with LCModel (right). The figure displays SPAWNN reconstruction and the LCModel fit (top), the residuals (middle), and the contribution of each metabolite (bottom). No correction was applied before analysis.
FIGURE 6
FIGURE 6
Results on 31P‐MRSI data, with the sum on 8 voxels. (A) Comparison of metabolite concentration ratios for the 10 volunteers with SPAWNN (circles) and LCModel (squares). ATP concentration is computed by averaging the concentration of the three resonances. (B) Bland–Altman plot of the difference between estimated values by SPAWNN and LCModel versus the average of the estimated values across all metabolite ratios and subjects.
FIGURE 7
FIGURE 7
Results from unlocalized FID data of the occipital lobe. Comparison of metabolite concentration estimation by SPAWNN and LCModel. The reference value is the estimate obtained with the average of the full acquisition (600 spectra). The plots show the convergence and the difference in estimation of each model with respect to the highest SNR estimate.
FIGURE 8
FIGURE 8
Examples of SPAWNN evaluation of a 3D 31P‐MRSI acquired on a human brain. The figure shows two spectra (TR = 1.5 s, TE = 0.5 ms, 24 weighted averages) arising from two brain regions. The T1 MP‐RAGE 1H MRI is displayed on the left, indicating the slice position and voxel locations. The voxel has a resolution of 25 mm isotropic. The spectra on the left represent the SPAWNN reconstruction with the measurement data, and the spectra on the right the LCModel fitting for voxels A and B.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

References

    1. Reto B, Dieter M, Ernst M, Peter B. Assessment of absolute metabolite concentrations in human tissue by 31P MRS in vivo part II: muscle, liver, kidney. Magn Reson Med. 1994;32:453‐458. - PubMed
    1. Chmelik M, Schmid A, Gruber S, et al. Three‐dimensional high‐resolution magnetic resonance spectroscopic imaging for absolute quantification of 31P metabolites in human liver. Magn Reson Med. 2008;60:796‐802. - PubMed
    1. Webb Graham A, ed. Methods and applications of phosphorus NMR spectroscopy in vivo. Annual Reports on NMR Spectroscopy. Vol 75. Academic Press; 2012.
    1. Jong‐Hee H, Soo CC. Use of in vivo magnetic resonance spectroscopy for studying metabolic diseases. Exp Mol Med. 2015;47:e139. - PMC - PubMed
    1. Rudin M, ed. In‐Vivo Magnetic Resonance Spectroscopy III: In‐Vivo MR Spectroscopy: Potential and Limitations. Springer Berlin Heidelberg; 1992.

Publication types