A novel method for Magnetic Resonance Spectroscopy lipid signal suppression using Semi-classical signal analysis and Bidirectional Long short-term memory
Item Type Conference Paper
Authors Gomez, Maria de los Angeles;Serrai, Hacéne;Bhaduri, Sourav;Laleg-Kirati, Taous-Meriem
Citation de los Angeles Gomez, M., Serrai, H., Bhaduri, S., & Laleg- Kirati, T.-M. (2022). A novel method for Magnetic Resonance Spectroscopy lipid signal suppression using Semi-classical signal analysis and Bidirectional Long short-term memory. 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). https://doi.org/10.1109/
embc48229.2022.9871645 Eprint version Post-print
DOI 10.1109/EMBC48229.2022.9871645
Publisher IEEE
Rights This is an accepted manuscript version of a paper before final publisher editing and formatting. Archived with thanks to IEEE.
Download date 2023-10-31 05:11:40
Link to Item http://hdl.handle.net/10754/681139
A novel method for Magnetic Resonance Spectroscopy lipid signal suppression using Semi-classical signal analysis and Bidirectional Long
short-term memory
Mar´ıa de los ´Angeles G´omez1, Hac`ene Serrai2 and Sourav Bhaduri3, Taous-Meriem Laleg-Kirati1,4
Abstract— Magnetic resonance spectroscopy (MRS) is a non- invasive method that enables the analysis and quantification of brain metabolites, which provide useful information about the neuro-biological substrates of brain function. Lactate plays a pivotal role in the diagnosis of various brain diseases. However, accurate lactate quantification is generally difficult to achieve due to the presence of large lipid peaks resonating at a similar spectral position. To overcome this problem several techniques have been proposed. However, most of them suffer from lactate signal loss or poor lipid peak removal. In this paper, a novel method for lipid suppression for MRS signal is proposed. The method combines a semi-classical signal analysis method and a bidirectional long short term memory technique. The method is validated using simulated data that mimics real MRS data.
I. INTRODUCTION
Magnetic resonance spectroscopy (MRS) is a common noninvasive method that allows the detection and quantification of brain tissue metabolites, that provide additional relevant information for several disease families such as brain tumors, metabolic disorders, and systemic diseases[1]. Proton (1H) MRS free induction decay (FID) is the most widely available MRS signal used for the quantitative measurement of various metabolites (e.g:
N-Acetyl Aspartate (NAA), Creatine, Lactate) in the brain, which is valuable to assess the metabolic activity for early diagnosis, planning and monitoring the treatment of various brain diseases [1]. Particularly, the lactate quantification is important as it is a by-product of anaerobic metabolism, therefore an indicator of cellular oxygenation, glycolysis and energy status in living tissues. An accurate quantification of lactate is important; its accumulation indicates altered metabolisms, as found in cancer and also is associated with an increased incidence of necrotic tissue. It plays a role also in the diagnosis and monitoring of many brain diseases such as tumors, stroke, and several mitochondrial disorders [2]. Nonetheless, the measurement of lactate concentration using 1H MRS is impeded by the overlap with large lipid resonances. This spectral overlap has been overcome
*This work was supported by King Abdullah University of Science and Technology (KAUST)
1Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) division, King Abdullah University of Science and Technol- ogy (KAUST), Thuwal, KSA e-mail: [email protected], Taous- [email protected]
2Centre for Clinical Imaging Research, Carle Foundation, Urbana, USA.
3Centre for Preclinical Imaging, University of Liverpool, Liverpool, United Kingdom.
4Taous Meriem Laleg-Kirati is also affiliated to the National Institute for Research in Digital Science and Technology, Paris-Saclay, France.
through spectral editing, including outer-volume suppression (OVS)[3], outer volume crushing[4] and inversion recovery [5] which can provide effective reduction of lipid artifacts.
However, these techniques may suffer from associated loss of brain metabolite signals to varying degrees and insufficient lipid suppression which may lead to bias in the lactate quantification [6]. To overcome these issues, post- acquisition data processing techniques have been proposed.
For instance, techniques based on wavelet transforms [7], [8],[9] have been proposed. However, the method in [7]
has not been designed to suppress the lipid peak when it overlaps with a metabolite and requires prior knowledge on the signal and the methods in [8],[9] are sensitive to motion artifacts.
The applicability of deep learning techniques have seen considerable improvements in biomedical signal processing recently. For instance, a conventional Convolutional Neural Network (CNN)- Long Short-Term Memory (LSTM) for ECG wave segmentation has been developed in [10], a merged LSTM model for an emotion analysis in EEG signals has been proposed in [11], a CNN-BiLSTM for ECG wave segmentation was preformed in [12]. All the work above have in common the use of recurrent neural networks(RNN), more specifically the LSTM that for being a type of RNN stands out for processing sequential data, which also solves the problem of vanishing-exploding gradients in long time series sequences. BiLSTM is an extension of traditional LSTMs which can improve model performance, by using two LSTM cells: one in forward direction and other in the backward direction of the signal. The BiLSTM helps computing both hidden sequences (future and history) by using prior and future data as an informational source [13].
In this paper, a novel method for MRS lipid suppression is proposed. The method combines a pulse-shaped sig- nal processing method called Semi-classical signal analysis (SCSA)[14] and a bidirectional long short term memory technique. The SCSA is based on the spectral problem of the Schr¨odinger operator and plays two roles: filtering the noise and providing new signal features.
The paper is organized as follows. In section 2, the proposed method is described along with the data set that has been used for testing the method. In section 3, the obtained results are presented and discussed. Sections 4 concludes the paper.
II. MATERIALS AND METHODS
The proposed method (Fig. 1) combines mathematical modeling and deep learning approach to simultaneously clean the signal and isolate the lactate peaks.
Fig. 1: General methodology for lipid peak suppression.
a) The input signal is the MRS spectra containing lac- tate doublet contaminated with lipids. b) The signal pre- processing steps: SCSA filtering, Baseline correction and Standardization to ensure the same scale for all the signals.
c) Lipid suppression with BiLSTM. d) The output signal is the MRS spectra with edited lactate doublet.
A. Data
A set of 20001H MRS signals, containing in addition to the main brain metabolites (NAA, Creatine (Cr) and Choline (cho)), the lactate doublet and a large lipid resonance, were generated. In this data, the spectral location of the lipid peak was varied between 0.9 to 1.3 part per million (ppm), and its signal intensity scaled by a factor 2 to 5 times higher than the lactate peak. The spectral location and signal intensity of the NAA, Cho, Cr peaks, were also varied to simulate different tumor grades and healthy data. The Signal to Noise Ratio (SNR), was changed from 2 to 15 dB. The sampling frequency was set to 2 kHz to generate each FID and the length of all signals was set to 1056 complex data points, which was Fourier transformed, and the magnitude of the spectrum was used for this methodology. Details about the simulator used to obtain the data are given in [15] and [16].
B. Signal pre-processing
A pre-processing stage was applied in order to increase the SNR, and to achieve a baseline correction due to the presence of lipid peaks. For that, the SCSA was used to reduce noise in the signal. It interprets the MR spectrum as a potential of a Schr¨odinger operator and decomposes it into a set of linear combinations of squared eigenfunctions than can represent the signaly as follows [17], [14]:
yh(f) = 4h
Nh
X
n=1
κnhψ2nh(f), (1)
where, h is a positive constant, −κ2nh and ψnh(f), for n=1,...,Nh refer to the negative eigenvalues with their re- spective L2-normalized eigenfunctions (κ1h > ... > κnh >
0), respectively, of the semi-classical Schr¨odinger operator H(y) =−h2dfd22−ysuch that its spectral problem is defined as [14], [18]:
H(y)ψ=λψ (2)
Baseline correction was achieved using an iterative adap- tive re-weighted penalized least squares [19] and a standard- ization with z-score in order to ensure that all the signals are bounded within the same range.
C. Lipid suppression
For the lipid peak suppression, different recurrent neural networks (Simple RNN, GRU, LSTM, BiLSTM) were tested and a hyper-parameters optimization by a Bayesian search tuning was performed to find the best model and the best set of hyper-parameters that best fits the data. The data set was divided into a training set(49%) and validation(21%) set to find the best model that fits the data and testing(30%) set to evaluate the performance of the model with blind data. The metric used to train, validate, and test, the model was the mean squared error (3), between the ground truth (yi) that in this case is the clean signal without the lipid peak and ˆ
yi that is the output of the model. Moreover, to evaluate the conservation of the lactate doublet (consisting of proximal peak (Lac 1) and a distal peak (Lac 2), the Amplitude Ratio (AR) (4) and the Signal Intensity Ratio (SIR) (5,6) are measured. Amplitude is calculated using the area under the curve by a numerical integration in the Lipid+Lactate frequency range of 0.9 - 1.5 ppm and the signal intensity is calculated as the highest of Lac 1 or Lac 2 peaks in the MRS spectra (max value).
M SE= 1 n
n
X
i=1
(yi−yˆi)2 (3)
AR=
RGT−R
BiLST M
RGT
, (4)
SIRlac1=
max(GTlac1)−max(BiLST Mlac1) max(GT lac1)
, (5)
SIRlac2=
max(GTlac2)−max(BiLST Mlac2) max(GT lac2)
, (6)
whereGT is the ground truth (clean signal containing lactate without contamination from the lipid peak), and BiLST M is the output of the model. Obtaining that, the best model is the BiLSTM neural network (table I) and the results of the Bayesian optimization shown in table II. From table I, we conclude that the best architecture is a network composed of two layers with a BiLSTM cell with 256 units, to avoid the overfitting a drop out of 0.5 of the fraction of the input units and batch normalization layer were added, ending with a dense layer with a linear function as activation to produce an output signal free from lipid peaks. The model was trained with a batch size of 300 data for 5000 epochs. The results of the training and validation are shown in Fig. 2.
TABLE I: MSE for different architectures.
SimpleRNN GRU LSTM BiLSTM
Training 0.9 0.83 0.56 0.08
Validation 0.11 0.69 0.27 0.02
Testing 0.10 0.65 0.27 0.02
Fig. 2: Curve of training and validation error
III. RESULTS AND DISCUSSION
The results of the first pre-processing stage are shown in Figure 3a, where the SCSA was used to denoise the signals [20]. To check on the usefulness of SCSA in reducing noise and therefore improving lipid suppression with the proposed neural network, a set of lactate measurements of signal intensity (a.u.) (table III) and the amplitude (a.u.) (table IV), ratio between the ground truth and the model output to the ground truth without and with SCSA was conducted.
The results indicate that SCSA improves the lipid peak suppression performance, as it allows for SNR enhancement which permits for accurate lactate quantification.
For lipid suppression, an RNN is proposed as it can extract temporal dependencies in sequential data as are the FIDs.
To select the best RNN model, different models were tested (table I). To evaluate the performance of the models, the MSE was calculated, obtaining the lowest MSE by the BiLSTM.
This is because the BiLSTM is composed of LSTM cells that are capable of learning long term sequences thanks to its gates with the capability to add or remove information by regulating the flow of information into and out of the cell in long sequences and thus leads to proper removal of the lipid peak information. Furthermore, the bidirectional feature adds accuracy to the forecasting output, ensuring that the amplitude of the lactate peak is preserved. The obtained results show an efficient removal of the lipid with conservation of the lactate doublet (Fig 3b).
The amplitude ratio of the lactate peak was calculated using Equation 4, obtaining a mean difference of (0.01 ± 0.00), indicating the loss of the lactate concentration is only 1%. This method preserves better lactate as compared to the
TABLE II: Hyperparameters with Bayesian tunning.
Hyperparameter Search space Step Optimum
Number of layers 1-3 1 2
Units 32-256 32 256
Dropout 0.0-0.5 0.1 0.5
Batch normalization True-False - True Learning rate 1e-05-0.01 Log sampling 4e-04
(a) De-noised and reconstructed MRS signal using SCSA with different SNR levels, SNR=5dB (Red), SNR=10dB (Blue), and SNR=15dB (Green).
(b) Lipid suppression with the BiLSTM. Denoised and reconstructed MRS signal using SCSA (Blue). Ground truth signal (Yellow), and the BiLSTM output (Red).
Fig. 3: Results of the filtering and the lipid peak suppression.
TABLE III: Ratio results of (5 and 6) of the signal intensity (a.u.) of lactate: noisy and de-noised MRS data with SCSA.
Lac 1 is the proximal peak and Lac 2 is the distal peak of the doublet.
SNR
(dB) Peak GT
BiLST M (without
SCSA)
BiLST M (with SCSA)
SIR (without
SCSA) SIR (with SCSA)
5 Lac 1 4.56 3.65 4.15 0.20 0.09
Lac 2 4.36 3.50 3.99 0.20 0.08
10 Lac 1 4.56 3.64 4.30 0.20 0.06
Lac 2 4.36 3.49 4.12 0.20 0.06
15 Lac 1 4.56 3.75 3.88 0.18 0.15
Lac 2 4.36 3.56 3.69 0.18 0.15
20 Lac 1 4.56 3.75 3.93 0.18 0.14
Lac 2 4.36 3.59 3.76 0.18 0.14
50 Lac 1 4.56 3.73 3.89 0.18 0.15
Lac 2 4.36 3.57 3.73 0.18 0.14
Mean 0.15 0.12
Std ±0.04 ±0.04
other approaches that lose up to 50% of the concentrations of the metabolites like Multiple-quantum-filtered (MQF) and longer echo times in spectroscopic sequences (e.g., T2=130 ms) that also results in the loss of metabolites with short spin-spin relaxation time (T2), due to the long echo time[21].
In addition, the proposed method is robust to different MRS signal variations, such as the presence or not of the lactate doublet, its chemical shift and signal intensity variation, alteration of the signal intensity of the other brain metabolites (eg. NAA, Cho, Cr), and the change of the chemical shift and amplitude of the overlapping lipid peaks.
A test with different levels of noise and signal intensity ratios was performed to assess the robustness of the model at higher noise levels and different signal intensity ratios between the lipid peak and the lactate doublet, defined as
TABLE IV: Ratio results of (4) of the amplitude (a.u.) of the lactate: noisy and de-noised MRS data with SCSA
SNR
(dB) GT
BiLST M (without
SCSA)
BiLST M (with SCSA)
AR (without
SCSA) AR (with SCSA)
5 25.15 24.29 25.16 0.03 0.00
10 25.15 24.19 25.29 0.04 0.01
15 25.15 24.55 24.86 0.02 0.01
20 25.15 24.56 24.90 0.02 0.01
50 25.15 24.61 24.88 0.02 0.01
Mean 0.03 0.01
Std ±0.01 ±0.00
TABLE V: Ratio results of (5 and 6) of the signal intensity (a.u.) of lactate peaks with different levels of noise. Lac 1 is the proximal peak and Lac 2 is the distal peak of the doublet.
SNR
(dB) Peak GT BiLST M SIR
5 Lac 1 4.56 4.15 0.09
Lac 2 4.36 3.99 0.08
10 Lac 1 4.56 4.30 0.06
Lac 2 4.36 4.12 0.06
15 Lac 1 4.56 3.88 0.15
Lac 2 4.36 3.69 0.15
TABLE VI: Ratio results of (5 and 6) of the signal intensity (a.u.) of lactate peaks with different α values. Lac 1 is the proximal peak and Lac 2 is the distal peak of the doublet.
α Peak GT BiLST M SIR
2 Lac 1 4.56 4.37 0.04
Lac 2 4.36 4.12 0.05
3 Lac 1 4.56 4.22 0.07
Lac 2 4.36 4.03 0.08
5 Lac 1 4.56 3.68 0.19
Lac 2 4.36 3.53 0.19
α = max(LipidP eak)
max(LactateDoublet). Results with variation in SNR and α are reported in tables V and VI. Mean error values of 0.10±0.04 and 0.10±0.06 were obtained with different SNR andαvalues, respectively. These results are comparable to those obtained using other proposed techniques[8][9], however, a definitive conclusion cannot be drawn with simulated data only. Testing the method on real data is planned for method investigation and validation. Our code, data and trained model are available at https://github.com/
EMANG-KAUST/MRS_LIPID_SUPPRESSION.
IV. CONCLUSION
In this paper, a novel method that combines the SCSA method and BiLSTM technique for lipid suppression in MRS signals is proposed. This method is efficient in significantly reducing noise and suppressing the lipid peak while pre- serving lactate peak information. The method can accurately extract lactate doublet where SNR is very low and lipid resonances very large hindering lactate signals. For future work, the model will be tested on real data for method validation.
V. ACKNOWLEDGMENT
Research reported in this publication was supported by King Abdullah University of Science and Technology (KAUST) with the Base Research Fund (BAS/1/1627-01-01).
REFERENCES
[1] Buonocore M. H. and Maddock R. J. Magnetic resonance spectroscopy of the brain: a review of physical principles and technical methods.
Reviews in the Neurosciences, 26(6):609–632, 2015.
[2] Nakamura H, Doi M, Suzuki T, Yoshida Y, Hoshikawa M, Uchida M, Tanaka Y, Takagi M, and Nakajima Y. The significance of lactate and lipid peaks for predicting primary neuroepithelial tumor grade with proton mr spectroscopy. Magnetic resonance in medical sciences:
MRMS: an official journal of Japan Society of Magnetic Resonance in Medicine, 3:238–243, 2018.
[3] Anke Henning, Alexander Fuchs, James B. Murdoch, and Peter Boesiger. Slice-selective fid acquisition, localized by outer volume suppression (fidlovs) for 1h-mrsi of the human brain at 7 t with minimal signal loss. NMR in Biomedicine, 22:683–696, 2009.
[4] Vincent O. Boer, Tessa Van De Lindt, Peter R. Luijten, and Den- nis W.J. Klomp. Lipid suppression for brain mri and mrsi by means of a dedicated crusher coil. Magnetic Resonance in Medicine, 73:2062–
2068, 6 2015.
[5] N. I. Avdievich, J. W. Pan, J. M. Baehring, D. D. Spencer, and Hoby P.
Hetherington. Short echo spectroscopic imaging of the human brain at 7t using transceiver arrays.Magnetic Resonance in Medicine, 62:17–
25, 2009.
[6] Liangjie Lin, Michal Povaˇzan, Adam Berrington, Zhong Chen, and Peter B. Barker. Water removal in mr spectroscopic imaging with l2 regularization. Magnetic Resonance in Medicine, 82:1278–1287, 10 2019.
[7] Luca T Mainardi, Daniela Origgi, Pietro Lucia, Giuseppe Scotti, and Sergio Cerutti. A wavelet packets decomposition algorithm for quantification of in vivo 1 h-mrs parameters, 2002.
[8] Hac`ene Serrai, Lydie Nadal-Desbarats, Harish Poptani, Jerry D Glick- son, and Lotfi Senhadji. Lactate editing and lipid suppression by continuous wavelet transform analysis: Application to simulated and 1h mrs brain tumor time-domain data. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 43(5):649–656, 2000.
[9] Hacene Serrai, Lotfi Senhadji, Guoyu Wang, Serge Akoka, and Patrick Stroman. Lactate doublet quantification and lipid signal suppression using a new biexponential decay filter: Application to simulated and 1h mrs brain tumor time-domain data. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 50(3):623–626, 2003.
[10] Aman Malali, Srinidhi Hiriyannaiah, G. M. Siddesh, K. G. Srinivasa, and N. T. Sanjay. Supervised ecg wave segmentation using convolu- tional lstm. ICT Express, 6:166–169, 9 2020.
[11] Anumit Garg, Ashna Kapoor, Anterpreet Kaur Bedi, and Ramesh K.
Sunkaria. Merged lstm model for emotion classification using eeg signals. In 2019 International Conference on Data Science and Engineering (ICDSE), pages 139–143, 2019.
[12] Aboli Londhe and Mithilesh Atulkar. Semantic segmentation of ecg waves using hybrid channel-mix convolutional and bidirectional lstm.
Biomedical Signal Processing and Control, 63:102162, 01 2021.
[13] Sima Siami Namini, Neda Tavakoli, and Akbar Siami Namin. The performance of lstm and bilstm in forecasting time series. pages 3285–
3292, 12 2019.
[14] Laleg-Kirati T.M., Cr´epeau E., and Sorine M. Semi-classical signal analysis.Mathematics of Control, Signals, and Systems, 25(1):37–61, 2013.
[15] Abderrazak Chahid, Sourav Bhaduri, Mohamed Maoui, Eric Achten, Hacene Serrai, and Taous Meriem Laleg-Kirati. Residual water sup- pression using the squared eigenfunctions of the schr¨odinger operator.
IEEE Access, 7:69126–69137, 2019.
[16] Sourav Bhaduri, Patricia Clement, Eric Achten, and Hacene Serrai.
Reduction of acquisition time using partition of the signal decay in spectroscopic imaging technique (rapid-si).PLoS ONE, 13, 11 2018.
[17] Bhaduri S., Chahid A., Achten E., Laleg-Kirati T. M., and Serrai H.
Scsa based matlab pre-processing toolbox for 1h mr spectroscopic water suppression and denoising. Informatics in Medicine Unlocked, 18:100294, 2020.
[18] Laleg-Kirati T.M., Zhang J., Achten E., and Serrai H. Spectral data de-noising using semi-classical signal analysis: application to localized mrs. NMR in Biomedicine, 29:1477–1485, 2016.
[19] Zhang Z., Chen S., and Liang Y. Baseline correction using adaptive iteratively re-weighted penalized least squares. Analyst, 135:1138–
1146, 2010.
[20] Abderrazak Chahid, Sourav Bhaduri, A Chahid, S Bhaduri, M Wali, H Serrai, R Achten, and T M Laleg-Kirati. Semi-classical signal analysis method with soft-thresholding for mrs denoising automation of [18f]alf labeling of peptides view project hemodynamic modelling view project quantum processing of mr spectrum mrs denoising using semi-classical signal analysis with soft-thresholding.
[21] In Young Choi, Ovidiu C. Andronesi, Peter Barker, Wolfgang Bogner, Richard A.E. Edden, Lana G. Kaiser, Phil Lee, Małgorzata Marja´nska, Melissa Terpstra, and Robin A. de Graaf. Spectral editing in 1h mag- netic resonance spectroscopy: Experts’ consensus recommendations, 5 2021.