0517
MR SPECTROSCOPY ARTIFACT REMOVAL WITH U-NET CONVOLUTIONAL NEURAL NETWORK1CREATIS, CNRS UMR 5220, Lyon, France
Synopsis
In in vivo MR spectroscopy, a variety of artifacts may affect spectral quality and are not easy to detect and remove by non-experts. A U-NET architecture is proposed to remove artifacts from MRS spectra with deep learning. The principle is demonstrated on synthetic simulated data mimicking in vivo conditions.
Introduction
Deep learning algorithms are more and more explored as a novel tool to solve difficult signal processing problem for in vivo MR spectroscopy. They have recently been proposed for MRS data quantification[1] or correction of MR spectra containing spurious echo ghost signals[2]. In in vivo MR spectroscopy, a variety of artifacts may affect spectral quality [3] and are not easy to detect and remove by non-experts. Here we propose a U-Net[4] architecture to remove some of the major signal artifacts encountered in in vivo MR spectroscopy. The whole procedure is based on synthetic MRS signal generation framework.METHOD
To give satisfactory results, supervised deep learning requires a large training dataset with known ground truth. Such a training dataset of in vivo MRS signal cannot be built as it requires costly acquisition on human subjects and ground truth artifact-free spectra are not available for in vivo signals. This was the motivation to set up a synthetic data generation framework. The resulting dataset, if it succeeds to reproduce the distribution of realistic in vivo signals, has the advantage of being generated free of cost and on a massive scale. The procedure to generate the dataset has been described in Fig1. This data generation, in addition to usual combination of metabolite, macromolecule (PRESS signals, TE=30ms, simulated using GAMMA[4]/ GAVA[5]) and additional Gaussian noise , includes different artifacts described Figure2. All these different signal components are parameterized (cf Eq 1 Figure2) and varied within large range during the training.
This process is repeated as many time as needed (2. 10^5 for training) , to create a large dataset of synthetic signals whose ground truth parameters are known. The eddy current effect, first and zero order phase and lipid contamination are corrected here. UNET convolutional neural network have been adopted and different architecture based on UNET tested. The U-Net is a convolutional Neural Network (CNN) that was originally developed for biomedical image segmentation. The network consists of a contracting path and an expansive path, which gives it the u-shaped architecture. The contracting path is a typical CNN that consists of repeated layers of convolutions, each followed by a rectified linear unit (ReLU) and a max-pooling operation. During the contraction, the spatial information is reduced while feature information is increased. The expansive pathway combines the feature and spatial information through a sequence of up-convolutions and concatenations with high-resolution features from the contracting path. Performances of artefact removal capacity is measured with root mean squared error (RMSE) between the normalized predicted and normalized artifact free spectra for the test data set.We show in this study, that a cascade of three U-Nets has better performances than a single U-Net for removal of artifacts in MRS signals. Caffe deep learning framework is used with the following solver parameters: type:"Adam", base_lr: 1e-3, lr_policy: "multistep", gamma: 0.1, stepvalue: 30000, 60000, and max_iter: 90000. Training, validation and test sample sizes are 200,000, 10,000, 10,000, respectively.
Results
Figure 2 shows representative results on data used as a test data set. Histograms of RMSE for a single U-Net vs three cascaded U-Net network are displayed Figure3, in average, the RMSE of the proposed 3 cascaded U-Net is significantly lower than with a single U-Net.Discussion
Our work has focused on some important artifacts such as lipid contamination, phase errors, and eddy current effect. The proposed network deals with raw time series and does not use 2D kernel as previously proposed for spurious echo ghosts signal removal[2]. This artefact removal procedure is intended to be tested either prior to standard quantification algorithm or to ease machine learning quantification task. Here an intense lipid component is supposed to be artifact coming from subcutaneous lipids and so our procedure can integrate a spatial knowledge or knowledge from other MRI modalities (e.g diffusion MRI) to handle tumor cases.Acknowledgements
This work was supported by the LABEX PRIMES (ANR-11-LABX-0063) of Université de Lyon, within the program "Investissements d'Avenir" (ANR-11-IDEX-0007).References
[1] Hatami, N., Sdika, M., & Ratiney, H. (2018). Magnetic Resonance Spectroscopy Quantification using Deep Learning. MICCAI 2018
[2] Kyathanahally, S. P., Döring, A., & Kreis, R. (2018). Deep learning approaches for detection and removal of ghosting artifacts in MR spectroscopy. Magnetic resonance in medicine, 80(3), 851-863.
[3] Kreis, R. (2004). Issues of spectral quality in clinical 1H‐magnetic resonance spectroscopy and a gallery of artifacts. NMR in Biomedicine, 17(6), 361-381.
[4] Ronneberger O., Fischer P., Brox T. (2015) U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N., Hornegger J., Wells W., Frangi A. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham
[5] Smith, S. A., Levante, T. O., Meier, B. H., & Ernst, R. R. (1994). Computer simulations in magnetic resonance. An object-oriented programming approach. Journal of Magnetic Resonance, Series A, 106(1), 75-105.
[6] Soher, B. J., Young, K., Bernstein, A., Aygula, Z., & Maudsley, A. A. (2007). GAVA: spectral simulation for in vivo MRS applications. Journal of magnetic resonance, 185(2), 291-299.
Figures
