Quantification of Metabolites in Magnetic Resonance Spectroscopic Imaging Using Machine Learning

A review of machine learning applications for the proton MR spectroscopy workflow

This literature review presents a comprehensive overview of machine learning (ML) applications in proton MR spectroscopy (MRS). As the use of ML techniques in MRS continues to grow, this review aims to provide the MRS community with a structured overview of the state-of-the-art methods. Specifically, we examine and summarize studies published between 2017 and 2023 from major journals in the MR field. We categorize these studies based on a typical MRS workflow, including data acquisition, processing, analysis, and artificial data generation. Our review reveals that ML in MRS is still in its early stages, with a primary focus on processing and analysis techniques, and less attention given to data acquisition. We also found that many studies use similar model architectures, with little comparison to alternative architectures. Additionally, the generation of artificial data is a crucial topic, with no consistent method for its generation. Furthermore, many studies demonstrate that artificial data suffers from generalization issues when tested on in vivo data. We also conclude that risks related to ML models should be addressed, particularly for clinical applications. Therefore, output uncertainty measures and model biases are critical to investigate. Nonetheless, the rapid development of ML in MRS and the promising results from the reviewed studies justify further research in this field.

Direct inference of Patlak parametric images in whole-body PET/CT imaging using convolutional neural networks

Purpose

This study proposed and investigated the feasibility of estimating Patlak-derived influx rate constant (K_i) from standardized uptake value (SUV) and/or dynamic PET image series.

Methods

Whole-body ¹⁸F-FDG dynamic PET images of 19 subjects consisting of 13 frames or passes were employed for training a residual deep learning model with SUV and/or dynamic series as input and K_i-Patlak (slope) images as output. The training and evaluation were performed using a nine-fold cross-validation scheme. Owing to the availability of SUV images acquired 60 min post-injection (20 min total acquisition time), the data sets used for the training of the models were split into two groups: “With SUV” and “Without SUV.” For “With SUV” group, the model was first trained using only SUV images and then the passes (starting from pass 13, the last pass, to pass 9) were added to the training of the model (one pass each time). For this group, 6 models were developed with input data consisting of SUV, SUV plus pass 13, SUV plus passes 13 and 12, SUV plus passes 13 to 11, SUV plus passes 13 to 10, and SUV plus passes 13 to 9. For the “Without SUV” group, the same trend was followed, but without using the SUV images (5 models were developed with input data of passes 13 to 9). For model performance evaluation, the mean absolute error (MAE), mean error (ME), mean relative absolute error (MRAE%), relative error (RE%), mean squared error (MSE), root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) were calculated between the predicted K_i-Patlak images by the two groups and the reference K_i-Patlak images generated through Patlak analysis using the whole acquired data sets. For specific evaluation of the method, regions of interest (ROIs) were drawn on representative organs, including the lung, liver, brain, and heart and around the identified malignant lesions.

Results

The MRAE%, RE%, PSNR, and SSIM indices across all patients were estimated as 7.45 ± 0.94%, 4.54 ± 2.93%, 46.89 ± 2.93, and 1.00 ± 6.7 × 10⁻⁷, respectively, for models predicted using SUV plus passes 13 to 9 as input. The predicted parameters using passes 13 to 11 as input exhibited almost similar results compared to the predicted models using SUV plus passes 13 to 9 as input. Yet, the bias was continuously reduced by adding passes until pass 11, after which the magnitude of error reduction was negligible. Hence, the predicted model with SUV plus passes 13 to 9 had the lowest quantification bias. Lesions invisible in one or both of SUV and K_i-Patlak images appeared similarly through visual inspection in the predicted images with tolerable bias.

Conclusion

This study concluded the feasibility of direct deep learning-based approach to estimate K_i-Patlak parametric maps without requiring the input function and with a fewer number of passes. This would lead to shorter acquisition times for WB dynamic imaging with acceptable bias and comparable lesion detectability performance.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-022-05867-w.

Prediction of multiple pH compartments by deep learning in magnetic resonance spectroscopy with hyperpolarized 13C-labelled zymonic acid

Background

Hyperpolarization enhances the sensitivity of nuclear magnetic resonance experiments by between four and five orders of magnitude. Several hyperpolarized sensor molecules have been introduced that enable high sensitivity detection of metabolism and physiological parameters. However, hyperpolarized magnetic resonance spectroscopy imaging (MRSI) often suffers from poor signal-to-noise ratio and spectral analysis is complicated by peak overlap. Here, we study measurements of extracellular pH (pH_e) by hyperpolarized zymonic acid, where multiple pH_e compartments, such as those observed in healthy kidney or other heterogeneous tissue, result in a cluster of spectrally overlapping peaks, which is hard to resolve with conventional spectroscopy analysis routines.

Methods

We investigate whether deep learning methods can yield improved pH_e prediction in hyperpolarized zymonic acid spectra of multiple pH_e compartments compared to conventional line fitting. As hyperpolarized ¹³C-MRSI data sets are often small, a convolutional neural network (CNN) and a multilayer perceptron (MLP) were trained with either a synthetic or a mixed (synthetic and augmented) data set of acquisitions from the kidneys of healthy mice.

Results

Comparing the networks’ performances compartment-wise on a synthetic test data set and eight real kidney data shows superior performance of CNN compared to MLP and equal or superior performance compared to conventional line fitting. For correct prediction of real kidney pH_e values, training with a mixed data set containing only 0.5% real data shows a large improvement compared to training with synthetic data only. Using a manual segmentation approach, pH maps of kidney compartments can be improved by neural network predictions for voxels including three pH compartments.

Conclusion

The results of this study indicate that CNNs offer a reliable, accurate, fast and non-interactive method for analysis of hyperpolarized ¹³C MRS and MRSI data, where low amounts of acquired data can be complemented to achieve suitable network training.

Bayesian deep learning-based 1 H-MRS of the brain: Metabolite quantification with uncertainty estimation using Monte Carlo dropout

PURPOSE

To develop a Bayesian convolutional neural network (BCNN) with Monte Carlo dropout sampling for metabolite quantification with simultaneous uncertainty estimation in deep learning-based proton MRS of the brain.

METHODS

Human brain spectra were simulated using basis spectra for 17 metabolites and macromolecules (N = 100 000) at 3.0 Tesla. In addition, actual in vivo spectra (N = 5) were modified by adjusting SNR and linewidth with increasing severity of spectral degradation (N = 50). A BCNN was trained on the simulated spectra to generate a noise-free, line-narrowed, macromolecule signal-removed, metabolite-only spectrum from a typical human brain spectrum. At inference, each input spectrum was Monte Carlo dropout sampled (50 times), and the resulting mean spectrum and variance spectrum were used for metabolite quantification and uncertainty estimation, respectively.

RESULTS

Using the simulated spectra, the mean absolute percent errors of the BCNN-predicted metabolite content were < 10% for Cr, Glu, Gln, mI, NAA, and Tau (< 5% for Glu, NAA, and mI). For all metabolites, the correlations (r's) between the ground-truth error and BCNN-predicted uncertainty ranged 0.72-0.94 (0.83 ± 0.06; p < 0.001). Using the modified in vivo spectra, the extent of variation in the estimated metabolite content against the increasing severity of spectral degradation tended to be smaller with BCNN than with linear combination of model spectra (LCModel). Overall, the variation in metabolite content tended to be more highly correlated with the uncertainty from BCNN than with the Cramér-Rao lower-bounds from LCModel (0.938 ± 0.019 vs. 0.881 ± 0.057 [p = 0.115]).

CONCLUSION

The BCNN with Monte Carlo dropout sampling may be used in deep learning-based MRS for the estimation of uncertainty in the machine-predicted metabolite content, which is important in the clinical application of deep learning-based MRS.

The clinical utility of proton magnetic resonance spectroscopy in traumatic brain injury: recommendations from the ENIGMA MRS working group

Proton (1H) magnetic resonance spectroscopy provides a non-invasive and quantitative measure of brain metabolites. Traumatic brain injury impacts cerebral metabolism and a number of research groups have successfully used this technique as a biomarker of injury and/or outcome in both pediatric and adult TBI populations. However, this technique is underutilized, with studies being performed primarily at centers with access to MR research support. In this paper we present a technical introduction to the acquisition and analysis of in vivo 1H magnetic resonance spectroscopy and review 1H magnetic resonance spectroscopy findings in different injury populations. In addition, we propose a basic 1H magnetic resonance spectroscopy data acquisition scheme (Supplemental Information) that can be added to any imaging protocol, regardless of clinical magnetic resonance platform. We outline a number of considerations for study design as a way of encouraging the use of 1H magnetic resonance spectroscopy in the study of traumatic brain injury, as well as recommendations to improve data harmonization across groups already using this technique.

Deep learning-based target metabolite isolation and big data-driven measurement uncertainty estimation in proton magnetic resonance spectroscopy of the brain

PURPOSE

The aim of this study was to develop a method for metabolite quantification with simultaneous measurement uncertainty estimation in deep learning-based proton magnetic resonance spectroscopy (¹ H-MRS).

METHODS

The reliability of metabolite quantification depends on signal-to-noise ratio (SNR), linewidth, and degree of spectral overlap (DSO), and therefore knowledge about these factors may be utilized in measurement uncertainty estimation in deep learning-based ¹ H-MRS. While SNR and linewidth are typically estimated from a representative singlet, DSO needs to be estimated metabolite-specifically. We developed convolutional neural networks (CNNs) capable of isolating target metabolite signal on simulated rat brain spectra at 9.4T, such that, in addition to metabolite content, the signal-to-background ratio (SBR) as a quantitative metric of DSO can be estimated directly from CNN-output for each metabolite. The CNN-predicted SBR was adjusted according to its pre-defined relationship to the ground-truth SBR by exploiting the big spectral data (N = 80 000), and used for measurement uncertainty estimation together with the SNR and linewidth from the CNN-input spectrum. The proposed method was tested first on the simulated spectra in comparison with LCModel and jMRUI and further on in vivo spectra.

RESULTS

The proposed method outperformed LCModel and jMRUI in both quantitative accuracy and measurement uncertainty estimation. Using in vivo data, the metabolite concentrations from the proposed method were close to the reported ranges with the measurement uncertainty of glutamine, glutamate, myo-inositol, N-acetylaspartate, and Tau less than 10%.

CONCLUSION

The proposed method may be used for metabolite quantification with measurement uncertainty estimation in rat brain at 9.4T by exploiting the spectral isolation capability of the CNNs and the availability of big spectral data.

Constrained Magnetic Resonance Spectroscopic Imaging by Learning Nonlinear Low-Dimensional Models

Magnetic resonance spectroscopic imaging (MRSI) is a powerful molecular imaging modality but has very limited speed, resolution, and SNR tradeoffs. Construction of a low-dimensional model to effectively reduce the dimensionality of the imaging problem has recently shown great promise in improving these tradeoffs. This paper presents a new approach to model and reconstruct the spectroscopic signals by learning a nonlinear low-dimensional representation of the general MR spectra. Specifically, we trained a deep neural network to capture the low-dimensional manifold, where the high-dimensional spectroscopic signals reside. A regularization formulation is proposed to effectively integrate the learned model and physics-based data acquisition model for MRSI reconstruction with the capability to incorporate additional spatiospectral constraints. An efficient numerical algorithm was developed to solve the associated optimization problem involving back-propagating the trained network. Simulation and experimental results were obtained to demonstrate the representation power of the learned model and the ability of the proposed formulation in producing SNR-enhancing reconstruction from the practical MRSI data.

Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain

PURPOSE

To develop a robust method for brain metabolite quantification in proton magnetic resonance spectroscopy (¹ H-MRS) using a convolutional neural network (CNN) that maps in vivo brain spectra that are typically degraded by low SNR, line broadening, and spectral baseline into noise-free, line-narrowed, baseline-removed intact metabolite spectra.

METHODS

A CNN was trained (n = 40 000) and tested (n = 5000) on simulated brain spectra with wide ranges of SNR (6.90-20.74) and linewidth (10-20 Hz). The CNN was further tested on in vivo spectra (n = 40) from five healthy volunteers with substantially different SNR, and the results were compared with those from the LCModel analysis. A Student t test was performed for the comparison.

RESULTS

Using the proposed method the mean-absolute-percent-errors (MAPEs) in the estimated metabolite concentrations were 12.49% ± 4.35% for aspartate, creatine (Cr), γ-aminobutyric acid (GABA), glucose, glutamine, glutamate, glutathione (GSH), myo-Inositol (mI), N-acetylaspartate, phosphocreatine (PCr), phosphorylethanolamine, and taurine over the whole simulated spectra in the test set. The metabolite concentrations estimated from in vivo spectra were close to the reported ranges for the proposed method and the LCModel analysis except mI, GSH, and especially Cr/PCr for the LCModel analysis, and phosphorylcholine to glycerophosphorylcholine ratio (PC/GPC) for both methods. The metabolite concentrations estimated across the in vivo spectra with different SNR were less variable with the proposed method (~10% or less) than with the LCModel analysis.

CONCLUSION

The robust performance of the proposed method against low SNR may allow a subminute ¹ H-MRS of human brain, which is an important technical development for clinical studies.

Convolutional Neural Networks for Direct Inference of Pharmacokinetic Parameters: Application to Stroke Dynamic Contrast-Enhanced MRI

Background and Purpose: The T1-weighted dynamic contrast enhanced (DCE)-MRI is an imaging technique that provides a quantitative measure of pharmacokinetic (PK) parameters characterizing microvasculature of tissues. For the present study, we propose a new machine learning (ML) based approach to directly estimate the PK parameters from the acquired DCE-MRI image-time series that is both more robust and faster than conventional model fitting.

Materials and Methods: We specifically utilize deep convolutional neural networks (CNNs) to learn the mapping between the image-time series and corresponding PK parameters. DCE-MRI datasets acquired from 15 patients with clinically evident mild ischaemic stroke were used in the experiments. Training and testing were carried out based on leave-one-patient-out cross- validation. The parameter estimates obtained by the proposed CNN model were compared against the two tracer kinetic models: (1) Patlak model, (2) Extended Tofts model, where the estimation of model parameters is done via voxelwise linear and nonlinear least squares fitting respectively.

Results: The trained CNN model is able to yield PK parameters which can better discriminate different brain tissues, including stroke regions. The results also demonstrate that the model generalizes well to new cases even if a subject specific arterial input function (AIF) is not available for the new data.

Conclusion: A ML-based model can be used for direct inference of the PK parameters from DCE image series. This method may allow fast and robust parameter inference in population DCE studies. Parameter inference on a 3D volume-time series takes only a few seconds on a GPU machine, which is significantly faster compared to conventional non-linear least squares fitting.