The Sim-to-Real Gap in MRS Quantification: A Systematic Deep Learning Validation for GABA

Zien Ma; S. M. Shermer; Oktay Karakuş; Frank C. Langbein

eess.SPFeb 23, 2026

The Sim-to-Real Gap in MRS Quantification: A Systematic Deep Learning Validation for GABA

Authors: Zien Ma, S. M. Shermer, Oktay Karakuş, Frank C. Langbein

Organizations: 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

Abstract

Magnetic resonance spectroscopy (MRS) is used to quantify metabolites in vivo and estimate biomarkers for conditions ranging from neurological disorders to cancers. Quantifying low-concentration metabolites such as GABA ( $γ$ -aminobutyric acid) is challenging due to low signal-to-noise ratio (SNR) and spectral overlap. We investigate and validate deep learning for quantifying complex, low-SNR, overlapping signals from MEGA-PRESS spectra, devise a convolutional neural network (CNN) and a Y-shaped autoencoder (YAE), and select the best models via Bayesian optimisation on 10,000 simulated spectra from slice-profile-aware MEGA-PRESS simulations. The selected models are trained on 100,000 simulated spectra. We validate their performance on 144 spectra from 112 experimental phantoms containing five metabolites of interest (GABA, Glu, Gln, NAA, Cr) with known ground truth concentrations across solution and gel series acquired at 3 T under varied bandwidths and implementations. These models are further assessed against the widely used LCModel quantification tool. On simulations, both models achieve near-perfect agreement (small MAEs; regression slopes $\approx 1.00$ , $R^2 \approx 1.00$ ). On experimental phantom data, errors initially increased substantially. However, modelling variable linewidths in the training data significantly reduced this gap. The best augmented deep learning models achieved a mean MAE for GABA over all phantom spectra of 0.151 (YAE) and 0.160 (FCNN) in max-normalised relative concentrations, outperforming the conventional baseline LCModel (0.220). A sim-to-real gap remains, but physics-informed data augmentation substantially reduced it. Phantom ground truth is needed to judge whether a method will perform reliably on real data.

Figures & tables

Panel of Figure 1 — Figure 1: The sequence diagram (a) shows the RF and gradient pulses with actual pulse shapes and timings used in the simulations. The initial excitation pulse is modelled as an ideal (instantaneous) slice-selective $90^{\circ}$ pulse and the corresponding slice selection gradient $G_{z}$ is therefore omitted. The excitation pulse excites a slice of thickness $3\text{\,}\mathrm{c}\mathrm{m}$ perpendicular to the $z$ -axis. The refocusing pulses refocus the magnetisation of $3\text{\,}\mathrm{c}\mathrm{m}$ thick slabs perpendicular to the $x$ and $y$ axis, respectively, to define the localised voxel. What differentiates the MEGA-PRESS sequence from the standard PRESS sequence is the presence of two $20\text{\,}\mathrm{m}\mathrm{s}$ frequency-selective Gaussian editing pulses (yellow and green) at $1.9\text{\,}\mathrm{p}\mathrm{p}\mathrm{m}$ for the ON acquisition. For the OFF spectra the editing pulses could in principle be omitted, but the simulation follows the experimental implementation where editing pulses at $7.5\text{\,}\mathrm{p}\mathrm{p}\mathrm{m}$ , which have no effect on the metabolites of interest, are applied instead. The readout of the signal starts at $68\text{\,}\mathrm{m}\mathrm{s}$ as indicated. Experimentally, it can last over $1\text{\,}\mathrm{s}$ , depending on the dwell time and number of samples acquired. For a bandwidth of $2000\text{\,}\mathrm{H}\mathrm{z}$ , the dwell time is $0.5\text{\,}\mathrm{m}\mathrm{s}$ and acquiring $N=2048$ samples, a typical signal length, would therefore require $2048\times$ 0.5\text{,}\mathrm{m}\mathrm{s} $=$ 1.024\text{,}\mathrm{s}$$ . The readout block in the diagram is truncated at $80\text{\,}\mathrm{m}\mathrm{s}$ for clarity, to show the RF pulses and timings. To account for imperfect slice profiles of the refocusing pulses, the spectra are simulated on a spatial grid (b) and the average over all positions is calculated.

Series	Medium	# Phantoms	NAA	Cr	Glu	Gln	GABA
E1	Solution	$13$	$15.0$	$0/8.0$	$0.0$	$0.0$	$0.00$ , $0.52$ , $1.04$ , $1.56$ , $2.07$ , $2.59$ , $3.10$ , $4.12$ , $6.15$ , $8.15$ , $10.12$ , $11.68$
E2	Solution	$1$	$15.0$	$0.0$	$0.0$	$0.0$	$0.0$
		$15$	$15.0$	$8.0$	$0.0$	$0.0$	$0.00$ , $0.50$ , $1.00$ , $1.50$ , $2.00$ , $2.50$ , $3.00$ , $3.99$ , $4.98$ , $5.96$ , $6.95$ , $7.93$ , $8.90$ , $9.88$ , $11.81$
E3	Solution	$15$	$15.0$	$8.0$	$12.0$	$3.0$	$0.00$ , $1.00$ , $2.00$ , $3.00$ , $3.99$ , $4.98$ , $5.97$ , $6.95$ , $7.93$ , $8.91$ , $9.88$ , $10.85$ , $11.81$ , $12.77$ , $13.73$
E4	Gel	$8\times 4$	$15.0$	$8.0$	$12.0$	$3.0$	$0.00$ , $1.00$ , $2.00$ , $3.00$ , $4.00$ , $6.00$ , $8.00$ , $10.00$
E6	Solution	$10$	$15.0$	$8.0$	$12.0$	$3.0$	$0.00$ , $1.03$ , $2.05$ , $3.06$ , $4.07$ , $6.09$ , $8.09$ , $10.08$ , $12.05$ , $14.98$

Table 1: Composition of the experimental data:

144

spectra from

112

phantoms (phantom count is the sum of the first number in each row, ignoring the repetition factor when given as e.g.

1\times 2

8\times 4

). For E4,

8\times 4

denotes 8 gel phantoms each scanned in four acquisition rounds (E4a–E4d). Concentrations are in

\text{mM}=\text{mmol}/\text{L}

. E1 contains a phantom with NAA only and a phantom with NAA and Cr only, followed by phantoms with increasing amounts of GABA (Cr indicated by

0/8.0

where applicable). E4 phantoms were acquired four times with different bandwidths (E4a/c:

1250\text{\,}\mathrm{H}\mathrm{z}

, E4b/d:

2000\text{\,}\mathrm{H}\mathrm{z}

) and reacquired one week later (E4c/d). E9 consists of

7

phantoms with varying concentrations of all metabolites; one phantom was scanned three times (

1\times 3

) and the rest twice (

1\times 2

), so that E9a contains one more spectrum than E9b.

Layer	Description
input	(channels, $N_{f}$ )
conv1	FCONV ( $f_{1}$ , $(1,k_{1})$ , $s_{1}$ , $d_{1}$ )
conv2	FCONV ( $f_{1}$ , $(1,k_{2})$ , $s_{1}$ , $d_{1}$ )
reduce	Repeat until one output channel:
	FCONV( $f_{1}$ , ( $\min(\text{channels},3)$ , $k_{3}$ ), $1$ , $d_{1}$ )
conv3	FCONV ( $f_{1}$ , $(1,k_{4})$ , $1$ , $d_{1}$ )

Table 2: Parameterised convolutional multi-class regressor (CNN) architecture with parameters.

N_{f}

is the number of frequencies in the inputs, and channels refers to the number of input spectra. The parameters

f_{1}

and

f_{2}

determine the number of convolutional filters,

k_{1}

k_{2}

k_{3}

and

k_{4}

determine the kernel size of the convolutions,

s_{1}

and

s_{2}

determine the use of strides or max-pooling,

d_{1}

determines the use of dropout or batch normalisation,

d_{2}

the dropout rate in the dense layer, and

e

the number of neurons in the dense layer. The activation function at the output is either sigmoid or softmax, and the output consists of

N_{m}

metabolite concentrations (fixed to

5

Figure 2: Illustration of the Y-shaped autoencoder (YAE) architecture. The input consists of one or more noisy MEGA-PRESS spectra (OFF, ON, DIFF using real, imaginary or magnitude representations). The encoder maps the input to a compressed latent representation. The decoder branch reconstructs denoised versions of the input spectra from this latent space. The quantifier branch predicts the metabolite concentrations from the same latent representation. Key components are highlighted: ① hidden layer activation function, ② dropout layer, ③ decoder output activation function, and ④ quantifier output activation function.

Encoder
Layer	Description
input	$(\text{channels},N_{f})$
dense_e	for each channel c:
	for $l=1,\dots,L_{e}-1$ :
	Dense $\left(\tfrac{N_{f}}{2^{l-1}},a_{e}\right)$
	Dropout $\left(d_{e}\right)$

Table 3: Parameterised architecture for the YAE’s encoder, decoder, and quantifier modules. The specific values explored for the parameters are given in Section 5 . Channels refer to the number of channels used as input spectra.

N_{f}

: number of frequency samples in the input spectrum.

N_{q}

: number of neurons in the first dense layer of the quantifier.

N_{m}

: number of output metabolite concentrations (fixed to

5

L_{e},L_{d},L_{q}

: number of dense layers in the encoder, decoder, and quantifier, respectively.

a_{e},a_{d}

: activation function for the encoder and decoder.

a_{q},a_{m}

: activation function for the hidden layers and output layer of the quantifier.

d_{e}

: dropout rate for the encoder.

Group	Parameter	Values	Best Model
Dataset	Normalisation	sum, max	sum
	Acquisitions	(DIFF, OFF), (DIFF, ON), (OFF, ON), (DIFF,OFF,ON)	(OFF,ON)
	Datatypes	(magnitude), (real), (imaginary, real)	(real)
Model	Kernel sizes	small: $k_{1}=7,k_{2}=5,k_{3}=3,k_{4}=3$
		medium: $k_{1}=9,k_{2}=7,k_{3}=5,k_{4}=3$	medium
		large: $k_{1}=11,k_{2}=9,k_{3}=7,k_{4}=3$

Table 4: Grid search model selection parameters for CNN model, varying the dataset options, exploring different convolutional kernel sizes (small, medium, large), activation functions and batch sizes, leaving the remaining model parameters fixed.

Figure 3: Validation and training concentration MAE for the top

25

432

CNN configurations from the grid search model selection (Table 4 ). Each pair of horizontal bars corresponds to one configuration: blue, validation MAE; red, training MAE. Error bars show standard deviation across five-fold cross-validation. Configurations are ordered by validation MAE (best at top). Dataset:

10{,}000

simulated spectra, sum normalisation, basis set linewidth

2\text{\,}\mathrm{H}\mathrm{z}

100

epochs.

Figure 4: Performance of

50

repetitions of the Bayesian optimisation for the CNN model selection over the Bayesian optimisation iterations. The grey shaded area (right axis) shows the percentage of runs that selected the same best configuration as the full grid search (“converged”).

Group	Parameter	Values
Dataset	Normalisation	sum
	Acquisitions	(DIFF, OFF, ON), (DIFF, ON), (DIFF, OFF), (OFF, ON)
	Datatypes	(real), (imaginary), (magnitude)
Model	$L_{e}$	$5,6,7,8,9$
	$L_{d}$	$5,6,7,8,9$
	$a_{e}$	tanh, ReLU (magnitude)

Table 5: Model selection parameters for the YAE encoder-decoder path.

Group	Parameter	Values
Dataset	Normalisation	sum
	Acquisitions	(DIFF, OFF, ON), (DIFF, ON), (DIFF, OFF), (OFF, ON)
	Datatypes	(real)
Model	$N_{q}$	$2048,1024,512,384,256,192,128$
	$L_{q}$	$2,3,4,5$
	$a_{q}$	linear, ReLU, sigmoid, tanh

Table 6: Model selection parameters for quantifier path

Group	Parameter	Refined Values (Stage 3)
Dataset	Normalisation	sum
	Acquisitions	(DIFF, OFF, ON), (DIFF, ON), (OFF, ON)
	Datatypes	(real), (imaginary, real)
Model	Encoder Depth ( $L_{e}$ )	$5,6$
	Decoder Depth ( $L_{d}$ )	$6,7,8$
	Encoder Activation ( $a_{e}$ )	tanh, ReLU

Table 7: Refined Search Space for the final joint optimisation (Stage 3), derived from the insights of Stages 1 & 2.

Figure 5: Validation and training concentration MAE for YAE configurations from the final joint optimisation (Stage 3). Each pair of horizontal bars corresponds to one configuration: blue, validation MAE; red, training MAE. Error bars show standard deviation across five-fold cross-validation. Configurations are ordered by validation MAE (best at top). Dataset: $10{,}000$ simulated spectra, sum normalisation, $200$ epochs (Table 7 ).

Group	Parameter	Selected Value
Dataset	Acquisitions	edit_off , edit_on
	Datatypes	imaginary , real
	Normalisation	sum
Encoder	Depth ( $L_{e}$ )	5
	Activation ( $a_{e}$ )	tanh
	Dropout Rate ( $d_{e}$ )	0.2

Table 8: The final, optimal YAE configuration selected from the joint optimisation Stage 3. This configuration achieved the lowest validation concentration error and demonstrated high stability across five-fold cross-validation.

Figure 6: Performance of optimal-configuration models in estimating metabolite concentrations for $2{,}000$ simulated validation spectra. Scatter plots of predicted vs actual (ground-truth) concentrations for each metabolite, with identity line; each point is one spectrum; each panel corresponds to one model and one metabolite (Cr, GABA, Gln, Glu).

Table 9: Performance indices for optimal-configuration models in estimating metabolite concentrations for

2{,}000

simulated validation spectra

Figure 7: Distribution of absolute quantification errors (max-normalised concentrations) for MRSNet-YAE, MRSNet-CNN, FCNN, and LCModel-OFF across $144$ spectra from $112$ phantoms, grouped by experimental series (E1–E14) and metabolite (Cr, GABA, Gln, Glu). Each box summarises the errors for one series–metabolite–model combination; outliers (points beyond $1.5$ IQR) are shown. NAA is the reference metabolite and is not quantified separately.

Metabolite	$p$ -value (OFF $<$ DIFF)	Lower error in experiments	Mean MAE difference	$N_{\text{exp}}$
GABA	$0.045$	$11/14$	$-1.35\times 10^{-2}$	$14$
Gln	$0.809$	$4/14$	$+3.62\times 10^{-3}$	$14$
Glu	$0.0019$	$12/14$	$-7.49\times 10^{-2}$	$14$

Table 10: Experiment-level comparison between LCModel-OFF and LCModel-DIFF using absolute errors on phantom data. Cr is omitted as it can be quantified from OFF; the comparison focuses on GABA, Gln and Glu, for which OFF vs DIFF quantification differs. For each experiment and metabolite, absolute errors were first computed for individual spectra, and the mean error across spectra was used as a single experiment-level summary statistic to avoid pseudo-replication (i.e., all values in this table are based on mean MAE). Paired one-tailed Wilcoxon signed-rank tests were applied across experiments to assess whether LCModel-OFF yields systematically lower errors (Section 3.7 ). LCModel-OFF achieved significantly lower errors for GABA (

p=0.045

, lower error in

11

out of

14

experiments) and Glu (

p=0.0019

, lower error in

12

out of

14

experiments), while no significant difference was observed for Gln. These results indicate that LCModel-OFF provides equal or superior quantification performance compared to LCModel-DIFF at the experiment level, supporting its use as a conservative and stable baseline for subsequent model comparisons.

Series	Cr				GABA
	MRSNet-YAE	MRSNet-CNN	LCModel-OFF	FCNN	MRSNet-YAE	MRSNet-CNN	LCModel-OFF	FCNN
E1	$0.069\pm.009$	$0.067\pm.009$	$\mathbf{0.009}\pm.004$	$0.015\pm.004$	$\mathbf{0.079}\pm.017$	$0.086\pm.019$	$0.096\pm.029$	$0.081\pm.020$
E2	$0.108\pm.015$	$0.229\pm.023$	$\mathbf{0.043}\pm.007$	$0.092\pm.014$	$\mathbf{0.171}\pm.029$	$0.188\pm.026$	$0.261\pm.056$	$0.185\pm.036$
E3	$0.291\pm.015$	$0.188\pm.009$	$\mathbf{0.133}\pm.022$	$0.137\pm.011$	$0.167\pm.020$	$\mathbf{0.070}\pm.016$	$0.247\pm.031$	$0.132\pm.019$
E4a	$0.305\pm.027$	$0.081\pm.009$	$0.038\pm.012$	$\mathbf{0.026}\pm.009$	$\mathbf{0.059}\pm.015$	$0.236\pm.065$	$0.133\pm.046$	$0.074\pm.015$
E4b	$0.317\pm.039$	$0.115\pm.016$	$0.045\pm.009$	$\mathbf{0.032}\pm.011$	$\mathbf{0.076}\pm.019$	$0.141\pm.049$	$0.145\pm.045$	$0.115\pm.022$

Table 11: Phantom performance for optimal-configuration models compared with the LCModel baseline. Experimental series E1–E14 and phantom compositions are summarised in Table 1 . Values are experiment-level mean MAE

\pm

SEM (standard error of the mean; max-normalised concentrations).

Metabolite	MRSNet-YAE	MRSNet-CNN	FCNN	LCModel-OFF
Cr	$0.165\pm 0.009$ [ $.146$ , $.183$ ]	$0.136\pm 0.010$ [ $.116$ , $.155$ ]	$0.091\pm 0.009$ [ $.074$ , $.109$ ]	$\mathbf{0.067}\pm 0.006$ [ $.056$ , $.078$ ]
GABA	$\mathbf{0.161}\pm 0.011$ [ $.139$ , $.183$ ]	$0.203\pm 0.013$ [ $.178$ , $.229$ ]	$0.167\pm 0.012$ [ $.144$ , $.190$ ]	$0.220\pm 0.015$ [ $.191$ , $.249$ ]
Gln	$0.238\pm 0.019$ [ $.201$ , $.276$ ]	$0.153\pm 0.012$ [ $.129$ , $.177$ ]	$\mathbf{0.080}\pm 0.009$ [ $.062$ , $.098$ ]	$0.116\pm 0.008$ [ $.101$ , $.132$ ]
Glu	$\mathbf{0.219}\pm 0.014$ [ $.191$ , $.246$ ]	$0.230\pm 0.014$ [ $.201$ , $.258$ ]	$0.237\pm 0.015$ [ $.208$ , $.266$ ]	$0.304\pm 0.016$ [ $.272$ , $.336$ ]
Overall	$0.160\pm 0.007$ [ $.147$ , $.173$ ]	$0.147\pm 0.006$ [ $.134$ , $.159$ ]	$\mathbf{0.117}\pm 0.006$ [ $.105$ , $.129$ ]	$0.144\pm 0.006$ [ $.131$ , $.156$ ]

Table 12: Overall MAE over all

144

phantom spectra (max-normalised concentrations). Primary target: GABA. Computed from raw per-spectrum absolute errors (DL models from model-dist; LCModel-OFF from LCM-results analysis on the same

144

spectra). Values are mean

\pm

SEM with

95\%

CI (normal approximation) in brackets. The “Overall” row is the mean over all

144

spectra of the per-spectrum MAE (mean over the five metabolites Cr, GABA, Gln, Glu, NAA).

Metabolite	MRSNet-YAE (Aug)	MRSNet-CNN (Aug)	FCNN (Aug)	LCModel-OFF
Cr	$0.103\pm 0.008$ [ $.088$ , $.118$ ]	$0.129\pm 0.009$ [ $.111$ , $.147$ ]	$0.102\pm 0.009$ [ $.084$ , $.120$ ]	$\mathbf{0.067}\pm 0.006$ [ $.056$ , $.078$ ]
GABA	$\mathbf{0.151}\pm 0.010$ [ $.132$ , $.170$ ]	$0.161\pm 0.011$ [ $.140$ , $.182$ ]	$0.160\pm 0.011$ [ $.139$ , $.181$ ]	$0.220\pm 0.015$ [ $.191$ , $.249$ ]
Gln	$0.194\pm 0.020$ [ $.155$ , $.234$ ]	$0.151\pm 0.013$ [ $.126$ , $.176$ ]	$\mathbf{0.099}\pm 0.011$ [ $.079$ , $.120$ ]	$0.116\pm 0.008$ [ $.101$ , $.132$ ]
Glu	$\mathbf{0.184}\pm 0.015$ [ $.156$ , $.213$ ]	$0.212\pm 0.014$ [ $.185$ , $.239$ ]	$0.209\pm 0.014$ [ $.183$ , $.236$ ]	$0.304\pm 0.016$ [ $.272$ , $.336$ ]
Overall	$0.131\pm 0.008$ [ $.115$ , $.146$ ]	$0.133\pm 0.005$ [ $.122$ , $.143$ ]	$\mathbf{0.116}\pm 0.005$ [ $.106$ , $.127$ ]	$0.144\pm 0.006$ [ $.131$ , $.156$ ]

Table 13: MAE over all

144

phantom spectra for models trained with variable linewidth augmentation (

1

–

10

Hz). Values are max-normalised relative concentrations (mean

\pm

SEM with

95\%

CI in brackets). Bold indicates the lowest error per metabolite. The “Overall” row is the mean over all

144

spectra of the per-spectrum MAE (mean over the five metabolites Cr, GABA, Gln, Glu, NAA).

Panel of Figure 8 — Figure 8: Comparison of simulated (FID-A) vs experimental (Skyra) basis spectra. Experimental basis spectra derived from $100\text{\,}\mathrm{m}\mathrm{M}$ metabolite solutions acquired on the same scanner.

Panel of Figure 9 — Figure 9: Representative comparison of simulated vs experimental ON spectra for a solution series (E1) and a gel phantom series (E4a), together with their residuals. Overall spectral magnitudes and peak positions agree well, particularly for the solution series, while the gel phantoms exhibit broader lineshapes and more pronounced baseline and phase differences that are closer to in vivo conditions.

Explore similar work

Data-driven Synthesis of Magnetic Resonance Spectroscopy Data using a Variational Autoencoder
Feb 28, 2026Dennis M. J. van de Sande, Julian P. Merkofer, Sina Amirrajab +4Synthetic Data Generation Synthetic Data Evaluation
Department of Biomedical Engineering Eindhoven University of Technology Eindhoven, The Netherlands · Department of Electrical Engineering Eindhoven University of Technology Eindhoven, The Netherlands · Department of Radiology & Nuclear Medicine Maastricht University Medical Center Maastricht, The Netherlands
HyperFitS -- Hypernetwork Fitting Spectra for metabolic quantification of ${}^1$ H MR spectroscopic imaging
Apr 3, 2026Paul J. Weiser, Gulnur Ungan, Amirmohammad Shamaei +5Magnetic Resonance Imaging Hypernetworks
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 +4
Physics-Informed Sylvester Normalizing Flows for Bayesian Inference in Magnetic Resonance Spectroscopy
May 6, 2025Julian P. Merkofer, Dennis M. J. van de Sande, Alex A. Bhogal +1Normalizing Flows Bayesian Inference
Eindhoven University of Technology, Eindhoven, The Netherlands · University Medical Center Utrecht, Utrecht, The Netherlands

Feb 28, 2026J/K move · Enter open · S save

Data-driven Synthesis of Magnetic Resonance Spectroscopy Data using a Variational Autoencoder

Dennis M. J. van de Sande, Julian P. Merkofer, Sina Amirrajab, Mitko Veta+3

Department of Biomedical Engineering Eindhoven University of Technology Eindhoven, The Netherlands · Department of Electrical Engineering Eindhoven University of Technology Eindhoven, The Netherlands · Department of Radiology & Nuclear Medicine Maastricht University Medical Center Maastricht, The Netherlands

Synthetic Data Generation Synthetic Data Evaluation Variational Autoencoders

Open paper AskPDF arXiv

The development of deep learning methods for magnetic resonance spectroscopy (MRS) is often hindered by limited availability of large, high-quality training datasets. While physics-based simulations are commonly used to mitigate this limitation, accurately modeling all in-vivo signal components remains challenging. In this work, we propose a data-driven framework for synthesizing in-vivo MRS data using a variational autoencoder (VAE) trained exclusively on measured single-voxel spectroscopy data. The model learns a low-dimensional latent representation of complex-valued spectra and enables generation of new samples through latent-space sampling and interpolation. The generative performance of the proposed approach is evaluated using a comprehensive set of complementary analyses, including reconstruction quality, feature-level similarity using low-dimensional embeddings, application-based signal quality metrics, and metabolite quantification agreement. The results demonstrate that the VAE accurately reconstructs dominant spectral patterns and generates synthetic spectra that occupy the same feature space as in-vivo data. In an example application targeting GABA-edited spectroscopy, augmenting limited subsets of transients with synthetic spectra improves signal quality metrics such as signal-to-noise ratio, linewidth, and shape scores. However, the results also reveal limitations of the generative approach, including under-representation of stochastic noise and reduced accuracy in absolute metabolite quantification, particularly for applications sensitive to concentration estimates. These findings highlight both potential and limitations of data-driven MRS synthesis. Beyond the proposed model, this study introduces a structured evaluation framework for generative MRS methods, emphasizing the importance of application-aware validation when synthetic data are used for downstream analysis.

Abstract

Explore similar work

Data-driven Synthesis of Magnetic Resonance Spectroscopy Data using a Variational Autoencoder

HyperFitS -- Hypernetwork Fitting Spectra for metabolic quantification of 1{}^11H MR spectroscopic imaging

Physics-Informed Sylvester Normalizing Flows for Bayesian Inference in Magnetic Resonance Spectroscopy

HyperFitS -- Hypernetwork Fitting Spectra for metabolic quantification of ${}^1$ H MR spectroscopic imaging