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

WALINET: A water and lipid identification convolutional neural network for nuisance signal removal in 1 H $$ {}^1\mathrm{H} $$ MR spectroscopic imaging

PURPOSE

Proton magnetic resonance spectroscopic imaging (  1 H $$ {}^1\mathrm{H} $$ -MRSI) provides noninvasive spectral-spatial mapping of metabolism. However, long-standing problems in whole-brain  1 H $$ {}^1\mathrm{H} $$ -MRSI are spectral overlap of metabolite peaks with large lipid signal from scalp, and overwhelming water signal that distorts spectra. Fast and effective methods are needed for high-resolution  1 H $$ {}^1\mathrm{H} $$ -MRSI to accurately remove lipid and water signals while preserving the metabolite signal. The potential of supervised neural networks for this task remains unexplored, despite their success for other MRSI processing.

METHODS

We introduce a deep learning method based on a modified Y-NET network for water and lipid removal in whole-brain  1 H $$ {}^1\mathrm{H} $$ -MRSI. The WALINET (WAter and LIpid neural NETwork) was compared with conventional methods such as the state-of-the-art lipid L2 regularization and Hankel-Lanczos singular value decomposition (HLSVD) water suppression. Methods were evaluated on simulated models and in vivo whole-brain MRSI using NMRSE, SNR, CRLB, and FWHM metrics.

RESULTS

WALINET is significantly faster and needs 8s for high-resolution whole-brain MRSI, compared with 42min for conventional HLSVD+L2. WALINET suppresses lipid and water in the brain by 25-45 and 34-53-fold, respectively. WALINET has better performance than HLSVD+L2, providing: (1) more lipid removal with 41% lower NRMSE; (2) better metabolite signal preservation with 71% lower NRMSE in simulated data; 155% higher SNR and 50% lower CRLB in in vivo data. Metabolic maps obtained by WALINET in healthy subjects and patients show better gray-/white-matter contrast with more visible structural details.

CONCLUSIONS

WALINET has superior performance for nuisance signal removal and metabolite quantification on whole-brain  1 H $$ {}^1\mathrm{H} $$ -MRSI compared with conventional state-of-the-art techniques. This represents a new application of deep learning for MRSI processing, with potential for automated high-throughput workflow.

Detecting Tumor-Associated Intracranial Hemorrhage Using Proton Magnetic Resonance Spectroscopy

Intracranial hemorrhage associated with primary or metastatic brain tumors is a critical condition that requires urgent intervention, often through open surgery. Nevertheless, surgical interventions may not always be feasible due to two main reasons: (1) extensive hemorrhage can obscure the underlying tumor mass, limiting radiological assessment; and (2) intracranial hemorrhage may occasionally present as the first symptom of a brain tumor without prior knowledge of its existence. The current review of case studies suggests that advanced radiological imaging techniques can improve diagnostic power for tumoral hemorrhage. Adding proton magnetic resonance spectroscopy (1H-MRS), which profiles biochemical composition of mass lesions could be valuable: it provides unique information about tumor states distinct from hemorrhagic lesions bypassing the structural obliteration caused by the hemorrhage. Recent advances in 1H-MRS techniques may enhance the modality's reliability in clinical practice. This perspective proposes that 1H-MRS can be utilized in clinical settings to enhance diagnostic power in identifying tumors underlying intracranial hemorrhage.

Recurrent neural network-aided processing of incomplete free induction decays in 1H-MRS of the brain

In the case of limited sampling windows or truncation of free induction decays (FIDs) for artifact removal in proton magnetic resonance spectroscopy (1H-MRS) and spectroscopic imaging (1H-MRSI), metabolite quantification needs to be performed on incomplete FIDs. Given that FIDs are naturally time-domain sequential data, we investigated the potential of recurrent neural network (RNN)-types of neural networks (NNs) in the processing of incomplete human brain FIDs with or without FID restoration prior to quantitative analysis at 3.0T. First, we employed an RNN encoder-decoder and developed it to restore incomplete FIDs (rRNN) with different amounts of sampled data. The quantification of metabolites from the rRNN-restored FIDs was achieved by using LCModel. Second, we modified the RNN encoder-decoder and developed it to convert incomplete brain FIDs into incomplete metabolite-only FIDs without restoration, followed by linear regression using a metabolite basis set for quantitative analysis (cRNN). In consideration of the practical benefit of the FID restoration with respect to pure zero-filling, development and analysis of the NNs were focused particularly on the incomplete FIDs with only the first 64 data points retained. All NNs were trained on simulated data and tested mainly on in vivo data acquired from healthy volunteers (n = 27). Strong correlations were obtained between the NN-derived and ground truth metabolite content (LCModel-derived content on fully sampled FIDs) for myo-inositol, total choline, and total creatine (normalized to total N-acetylaspartate) on the in vivo data using both rRNN (R = 0.83-0.94; p ≤ 0.05) and cRNN (R = 0.86-0.91; p ≤ 0.05). RNN-types of NNs have potential in the quantification of the major brain metabolites from the FIDs with substantially reduced sampled data points. For the metabolites with low to medium SNR, the performance of the NNs needs to be further improved, for which development of more elaborate and advanced simulation techniques would be of help, but remains challenging.

Comparison of convolutional-neural-networks-based method and LCModel on the quantification of in vivo magnetic resonance spectroscopy

BACKGROUND

Quantification of metabolites concentrations in institutional unit (IU) is important for inter-subject and long-term comparisons in the applications of magnetic resonance spectroscopy (MRS). Recently, deep learning (DL) algorithms have found a variety of applications on the process of MRS data. A quantification strategy compatible to DL base MRS spectral processing method is, therefore, useful.

MATERIALS AND METHODS

This study aims to investigate whether metabolite concentrations quantified using a convolutional neural network (CNN) based method, coupled with a scaling procedure that normalizes spectral signals for CNN input and linear regression, can effectively reflect variations in metabolite concentrations in IU across different brain regions with varying signal-to-noise ratios (SNR) and linewidths (LW). An error index based on standard error (SE) is proposed to indicate the confidence levels associated with metabolite predictions. In vivo MRS spectra were acquired from three brain regions of 43 subjects using a 3T system.

RESULTS

The metabolite concentrations in IU of five major metabolites, quantified using CNN and LCModel, exhibit similar ranges with Pearson's correlation coefficients ranging from 0.24 to 0.78. The SE of the metabolites shows a positive correlation with Cramer-Rao lower bound (CRLB) (r=0.46) and  absolute CRLB (r=0.81), calculated by multiplying CRLBs with the quantified metabolite content.

CONCLUSION

In conclusion, the CNN based method with the proposed scaling procedures can be employed to quantify in vivo MRS spectra and derive metabolites concentrations in IU. The SE can be used as error index, indicating predicted uncertainties for metabolites and sharing information similar to the absolute CRLB.

Results of the 2023 ISBI challenge to reduce GABA-edited MRS acquisition time

PURPOSE

Use a conference challenge format to compare machine learning-based gamma-aminobutyric acid (GABA)-edited magnetic resonance spectroscopy (MRS) reconstruction models using one-quarter of the transients typically acquired during a complete scan.

METHODS

There were three tracks: Track 1: simulated data, Track 2: identical acquisition parameters with in vivo data, and Track 3: different acquisition parameters with in vivo data. The mean squared error, signal-to-noise ratio, linewidth, and a proposed shape score metric were used to quantify model performance. Challenge organizers provided open access to a baseline model, simulated noise-free data, guides for adding synthetic noise, and in vivo data.

RESULTS

Three submissions were compared. A covariance matrix convolutional neural network model was most successful for Track 1. A vision transformer model operating on a spectrogram data representation was most successful for Tracks 2 and 3. Deep learning (DL) reconstructions with 80 transients achieved equivalent or better SNR, linewidth and fit error compared to conventional 320 transient reconstructions. However, some DL models optimized linewidth and SNR without actually improving overall spectral quality, indicating a need for more robust metrics.

CONCLUSION

DL-based reconstruction pipelines have the promise to reduce the number of transients required for GABA-edited MRS.

Denoising single MR spectra by deep learning: Miracle or mirage?

PURPOSE

The inherently poor SNR of MRS measurements presents a significant hurdle to its clinical application. Denoising by machine or deep learning (DL) was proposed as a remedy. It is investigated whether such denoising leads to lower estimate uncertainties or whether it essentially reduces noise in signal-free areas only.

METHODS

Noise removal based on supervised DL with U-nets was implemented using simulated 1 H MR spectra of human brain in two approaches: (1) via time-frequency domain spectrograms and (2) using 1D spectra as input. Quality of denoising was evaluated in three ways: (1) by an adapted fit quality score, (2) by traditional model fitting, and (3) by quantification via neural networks.

RESULTS

Visually appealing spectra were obtained; hinting that denoising is well-suited for MRS. However, an adapted denoising score showed that noise removal is inhomogeneous and more efficient for signal-free areas. This was confirmed by quantitative analysis of traditional fit results as well as DL quantitation following DL denoising. DL denoising, although apparently successful as judged by mean squared errors, led to substantially biased estimates in both implementations.

CONCLUSION

The implemented DL-based denoising techniques may be useful for display purposes, but do not help quantitative evaluations, confirming expectations based on estimation theory: Cramér Rao lower bounds defined by the original data and the appropriate fitting model cannot be circumvented in an unbiased way for single data sets, unless additional prior knowledge can be incurred in the form of parameter restrictions/relations or applicable substates.

Quantification of short echo time MRS signals with improved version of QUantitation based on quantum ESTimation algorithm

Magnetic resonance spectroscopy offers information about metabolite changes in the organism, which can be used in diagnosis. While short echo time proton spectra exhibit more distinguishable metabolites compared with proton spectra acquired with long echo times, their quantification (and providing estimates of metabolite concentrations) is more challenging. They are hampered by a background signal, which originates mainly from macromolecules (MM) and mobile lipids. An improved version of the quantification algorithm QUantitation based on quantum ESTimation (QUEST), with MM prior knowledge (QUEST-MM), dedicated to proton signals and invoking appropriate prior knowledge on MM, is proposed and tested. From a single acquisition, it enables better metabolite quantification, automatic estimation of the background, and additional automatic quantification of MM components, thus improving its applicability in the clinic. The proposed algorithm may facilitate studies that involve patients with pathological MM in the brain. QUEST-MM and three QUEST-based strategies for quantifying short echo time signals are compared in terms of bias-variance trade-off and Cramér-Rao lower bound estimates. The performances of the methods are evaluated through extensive Monte Carlo studies. In particular, the histograms of the metabolite and MM amplitude distributions demonstrate the performances of the estimators. They showed that QUEST-MM works better than QUEST (Subtract approach) and is a good alternative to QUEST when measured MM signal is unavailable or unsuitable. Quantification with QUEST-MM is shown for 1 H in vivo rat brain signals obtained with the SPECIAL pulse sequence at 9.4 T, and human brain signals obtained, respectively, with STEAM at 4 T and PRESS at 3 T. QUEST-MM is implemented in jMRUI and will be available for public use from version 7.1.

Sodium Hydrosulfide Treatment During Porcine Kidney Ex Vivo Perfusion and Transplantation

Background

In rodents, hydrogen sulfide (H2S) reduces ischemia-reperfusion injury and improves renal graft function after transplantation. Here, we hypothesized that the benefits of H2S are conserved in pigs, a more clinically relevant model.

Methods

Adult porcine kidneys retrieved immediately or after 60 min of warm ischemia (WI) were exposed to 100 µM sodium hydrosulfide (NaHS) (1) during the hypothermic ex vivo perfusion only, (2) during WI only, and (3) during both WI and ex vivo perfusion. Kidney perfusion was evaluated with dynamic contrast-enhanced MRI. MRI spectroscopy was further employed to assess energy metabolites including ATP. Renal biopsies were collected at various time points for histopathological analysis.

Results

Perfusion for 4 h pig kidneys with Belzer MPS UW + NaHS resulted in similar renal perfusion and ATP levels than perfusion with UW alone. Similarly, no difference was observed when NaHS was administered in the renal artery before ischemia. After autotransplantation, no improvement in histologic lesions or cortical/medullary kidney perfusion was observed upon H2S administration. In addition, AMP and ATP levels were identical in both groups.

Conclusions

In conclusion, treatment of porcine kidney grafts using NaHS did not result in a significant reduction of ischemia-reperfusion injury or improvement of kidney metabolism. Future studies will need to define the benefits of H2S in human, possibly using other molecules as H2S donors.

Application of a 1H Brain MRS Benchmark Dataset to Deep Learning for Out-of-Voxel Artifacts

Neural networks are potentially valuable for many of the challenges associated with MRS data. The purpose of this manuscript is to describe the AGNOSTIC dataset, which contains 259,200 synthetic 1H MRS examples for training and testing neural networks. AGNOSTIC was created using 270 basis sets that were simulated across 18 field strengths and 15 echo times. The synthetic examples were produced to resemble in vivo brain data with combinations of metabolite, macromolecule, residual water signals, and noise. To demonstrate the utility, we apply AGNOSTIC to train two Convolutional Neural Networks (CNNs) to address out-of-voxel (OOV) echoes. A Detection Network was trained to identify the point-wise presence of OOV echoes, providing proof of concept for real-time detection. A Prediction Network was trained to reconstruct OOV echoes, allowing subtraction during post-processing. Complex OOV signals were mixed into 85% of synthetic examples to train two separate CNNs for the detection and prediction of OOV signals. AGNOSTIC is available through Dryad and all Python 3 code is available through GitHub. The Detection network was shown to perform well, identifying 95% of OOV echoes. Traditional modeling of these detected OOV signals was evaluated and may prove to be an effective method during linear-combination modeling. The Prediction Network greatly reduces OOV echoes within FIDs and achieved a median log10 normed-MSE of -1.79, an improvement of almost two orders of magnitude.

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.

Quantification of MR spectra by deep learning in an idealized setting: Investigation of forms of input, network architectures, optimization by ensembles of networks, and training bias

PURPOSE

The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution.

METHODS

Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF).

RESULTS

Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value.

CONCLUSION

MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance-like in the case of MF.

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

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 of31 $$ {}^{31} $$ 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-dimensional31 $$ {}^{31} $$ 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.

Subnormothermic Ex Vivo Porcine Kidney Perfusion Improves Energy Metabolism: Analysis Using 31P Magnetic Resonance Spectroscopic Imaging

The ideal preservation temperature for donation after circulatory death kidney grafts is unknown. We investigated whether subnormothermic (22 °C) ex vivo kidney machine perfusion could improve kidney metabolism and reduce ischemia-reperfusion injury.

Convolutional neural networks to predict brain tumor grades and Alzheimer’s disease with MR spectroscopic imaging data

Purpose

To evaluate the value of convolutional neural network (CNN) in the diagnosis of human brain tumor or Alzheimer’s disease by MR spectroscopic imaging (MRSI) and to compare its Matthews correlation coefficient (MCC) score against that of other machine learning methods and previous evaluation of the same data. We address two challenges: 1) limited number of cases in MRSI datasets and 2) interpretability of results in the form of relevant spectral regions.

Methods

A shallow CNN with only one hidden layer and an ad-hoc loss function was constructed involving two branches for processing spectral and image features of a brain voxel respectively. Each branch consists of a single convolutional hidden layer. The output of the two convolutional layers is merged and fed to a classification layer that outputs class predictions for the given brain voxel.

Results

Our CNN method separated glioma grades 3 and 4 and identified Alzheimer’s disease patients using MRSI and complementary MRI data with high MCC score (Area Under the Curve were 0.87 and 0.91 respectively). The results demonstrated superior effectiveness over other popular methods as Partial Least Squares or Support Vector Machines. Also, our method automatically identified the spectral regions most important in the diagnosis process and we show that these are in good agreement with existing biomarkers from the literature.

Conclusion

Shallow CNNs models integrating image and spectral features improved quantitative and exploration and diagnosis of brain diseases for research and clinical purposes. Software is available at https://bitbucket.org/TeslaH2O/cnn_mrsi.

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.

Deep Learning and Its Application to Function Approximation for MR in Medicine: An Overview

This article presents an overview of deep learning (DL) and its applications to function approximation for MR in medicine. The aim of this article is to help readers develop various applications of DL. DL has made a large impact on the literature of many medical sciences, including MR. However, its technical details are not easily understandable for non-experts of machine learning (ML).The first part of this article presents an overview of DL and its related technologies, such as artificial intelligence (AI) and ML. AI is explained as a function that can receive many inputs and produce many outputs. ML is a process of fitting the function to training data. DL is a kind of ML, which uses a composite of many functions to approximate the function of interest. This composite function is called a deep neural network (DNN), and the functions composited into a DNN are called layers. This first part also covers the underlying technologies required for DL, such as loss functions, optimization, initialization, linear layers, non-linearities, normalization, recurrent neural networks, regularization, data augmentation, residual connections, autoencoders, generative adversarial networks, model and data sizes, and complex-valued neural networks.The second part of this article presents an overview of the applications of DL in MR and explains how functions represented as DNNs are applied to various applications, such as RF pulse, pulse sequence, reconstruction, motion correction, spectroscopy, parameter mapping, image synthesis, and segmentation.

Deep learning can accelerate and quantify simulated localized correlated spectroscopy

Nuclear magnetic resonance spectroscopy (MRS) allows for the determination of atomic structures and concentrations of different chemicals in a biochemical sample of interest. MRS is used in vivo clinically to aid in the diagnosis of several pathologies that affect metabolic pathways in the body. Typically, this experiment produces a one dimensional (1D) 1H spectrum containing several peaks that are well associated with biochemicals, or metabolites. However, since many of these peaks overlap, distinguishing chemicals with similar atomic structures becomes much more challenging. One technique capable of overcoming this issue is the localized correlated spectroscopy (L-COSY) experiment, which acquires a second spectral dimension and spreads overlapping signal across this second dimension. Unfortunately, the acquisition of a two dimensional (2D) spectroscopy experiment is extremely time consuming. Furthermore, quantitation of a 2D spectrum is more complex. Recently, artificial intelligence has emerged in the field of medicine as a powerful force capable of diagnosing disease, aiding in treatment, and even predicting treatment outcome. In this study, we utilize deep learning to: (1) accelerate the L-COSY experiment and (2) quantify L-COSY spectra. All training and testing samples were produced using simulated metabolite spectra for chemicals found in the human body. We demonstrate that our deep learning model greatly outperforms compressed sensing based reconstruction of L-COSY spectra at higher acceleration factors. Specifically, at four-fold acceleration, our method has less than 5% normalized mean squared error, whereas compressed sensing yields 20% normalized mean squared error. We also show that at low SNR (25% noise compared to maximum signal), our deep learning model has less than 8% normalized mean squared error for quantitation of L-COSY spectra. These pilot simulation results appear promising and may help improve the efficiency and accuracy of L-COSY experiments in the future.

Separation of Metabolites and Macromolecules for Short-TE 1H-MRSI Using Learned Component-Specific Representations

Short-echo-time (TE) proton magnetic resonance spectroscopic imaging (MRSI) allows for simultaneously mapping a number of molecules in the brain, and has been recognized as an important tool for studying in vivo biochemistry in various neuroscience and disease applications. However, separation of the metabolite and macromolecule (MM) signals present in the short-TE data with significant spectral overlaps remains a major technical challenge. This work introduces a new approach to solve this problem by integrating imaging physics and representation learning. Specifically, a mixed unsupervised and supervised learning-based strategy was developed to learn the metabolite and MM-specific low-dimensional representations using deep autoencoders. A constrained reconstruction formulation is proposed to integrate the MRSI spatiospectral encoding model and the learned representations as effective constraints for signal separation. An efficient algorithm was developed to solve the resulting optimization problem with provable convergence. Simulation and experimental results have been obtained to demonstrate the component-specific representation power of the learned models and the capability of the proposed method in separating metabolite and MM signals for practical short-TE [Formula: see text]-MRSI data.

Unsupervised anomaly detection using generative adversarial networks in 1H-MRS of the brain

The applicability of generative adversarial networks (GANs) capable of unsupervised anomaly detection (AnoGAN) was investigated in the management of quality of ¹H-MRS human brain spectra at 3.0 T. The AnoGAN was trained in an unsupervised manner solely on simulated normal brain spectra and used for filtering out abnormal spectra with a broad range of abnormalities, which were simulated by including abnormal ranges of SNR, linewidth and metabolite concentrations and spectral artifacts such as ghost, residual water, and lipid. The AnoGAN was able to filter out those spectra with SNR less than ~11-12 dB with an accuracy of ~80% or higher (assuming a normal SNR range to be 15-18 dB). It also detected with an accuracy of ~80% or higher those spectra, in which NAA levels were reduced by ~25-30% or more from the lower bound and elevated by ~20-30% or more from the upper bound of the normal concentration range (7.5-17 mmol/L), while the concentrations of the rest of the metabolites were all within the normal ranges. Despite the fact that those spectra contaminated with ghost, residual water or lipid have never been involved in the training or optimization of the AnoGAN, they were correctly classified as abnormal regardless of the types of the artifacts, depending solely on their intensity. Although the current version of our AnoGAN requires further technical improvement particularly for the detection of linewidth-associated abnormality and validation on in vivo data, our unsupervised deep learning-based approach could be an option in addition to those previously reported supervised deep learning-based approaches in the binary classification of spectral quality with an extended abnormal spectra regime.

Comparison of different linear-combination modeling algorithms for short-TE proton spectra

Short-TE proton MRS is used to study metabolism in the human brain. Common analysis methods model the data as a linear combination of metabolite basis spectra. This large-scale multi-site study compares the levels of the four major metabolite complexes in short-TE spectra estimated by three linear-combination modeling (LCM) algorithms. 277 medial parietal lobe short-TE PRESS spectra (TE = 35 ms) from a recent 3 T multi-site study were preprocessed with the Osprey software. The resulting spectra were modeled with Osprey, Tarquin and LCModel, using the same three vendor-specific basis sets (GE, Philips and Siemens) for each algorithm. Levels of total N-acetylaspartate (tNAA), total choline (tCho), myo-inositol (mI) and glutamate + glutamine (Glx) were quantified with respect to total creatine (tCr). Group means and coefficient of variations of metabolite estimates agreed well for tNAA and tCho across vendors and algorithms, but substantially less so for Glx and mI, with mI systematically estimated as lower by Tarquin. The cohort mean coefficient of determination for all pairs of LCM algorithms across all datasets and metabolites was R 2 ¯ = 0.39, indicating generally only moderate agreement of individual metabolite estimates between algorithms. There was a significant correlation between local baseline amplitude and metabolite estimates (cohort mean R 2 ¯ = 0.10). While mean estimates of major metabolite complexes broadly agree between linear-combination modeling algorithms at group level, correlations between algorithms are only weak-to-moderate, despite standardized preprocessing, a large sample of young, healthy and cooperative subjects, and high spectral quality. These findings raise concerns about the comparability of MRS studies, which typically use one LCM software and much smaller sample sizes.

Frequency and phase correction of J-difference edited MR spectra using deep learning

PURPOSE

To investigate whether a deep learning-based (DL) approach can be used for frequency-and-phase correction (FPC) of MEGA-edited MRS data.

METHODS

Two neural networks (1 for frequency, 1 for phase) consisting of fully connected layers were trained and validated using simulated MEGA-edited MRS data. This DL-FPC was subsequently tested and compared to a conventional approach (spectral registration [SR]) and to a model-based SR implementation (mSR) using in vivo MEGA-edited MRS datasets. Additional artificial offsets were added to these datasets to further investigate performance.

RESULTS

The validation showed that DL-based FPC was capable of correcting within 0.03 Hz of frequency and 0.4°of phase offset for unseen simulated data. DL-based FPC performed similarly to SR for the unmanipulated in vivo test datasets. When additional offsets were added to these datasets, the networks still performed well. However, although SR accurately corrected for smaller offsets, it often failed for larger offsets. The mSR algorithm performed well for larger offsets, which was because the model was generated from the in vivo datasets. In addition, the computation times were much shorter using DL-based FPC or mSR compared to SR for heavily distorted spectra.

CONCLUSION

These results represent a proof of principle for the use of DL for preprocessing MRS data.