4995

Can a neural network predict B0 maps from uncorrected CEST-spectra?
Felix Glang1, Anagha Deshmane1, Florian Martin1, Kai Herz1, Klaus Scheffler1,2, and Moritz Zaiss1

1High-Field Magnetic Resonance Center, Max Planck Institute for Biological Cybernetics, Tuebingen, Germany, 2Department of Biomedical Magnetic Resonance, Eberhard Karls University Tuebingen, Tuebingen, Germany

Synopsis

Analysis of chemical exchange saturation transfer (CEST) effects suffers from B0 inhomogeneity. Common correction methods involve computationally expensive algorithms or even additional measurements. Here we demonstrate that deep neural networks are able to predict B0 maps from raw Z-spectra by training the networks with measured B0 maps. Moreover, we show that CEST contrast parameters representing amide, amine and NOE resonance peaks can be directly predicted from uncorrected Z-spectra in a fast single step. This provides a shortcut to conventional evaluation procedures and will be useful to guide nonlinear model fitting.

Introduction

The chemical exchange saturation transfer (CEST) spectrum is affected by B0 inhomogeneities. Required correction is usually done using additional field mapping or Z-spectra interpolation. Furthermore, extraction of isolated CEST contrasts at 3 T usually needs sophisticated and potentially time-consuming evaluation steps, e.g. multiple non-linear least squares fits. The purpose of this work is to investigate, if B0 maps and the desired CEST contrasts can be directly extracted from uncorrected 3T Z-spectra employing a deep neural network trained with conventionally acquired B0 maps and fitted CEST contrast parameters.

Methods

Highly resolved Z-spectra (53 frequency offsets) of five healthy subjects were acquired by 3D-snapshot CEST MRI1 at 3T (Siemens PRISMA) after written informed consent. B0 and B1 maps were acquired using WASABI2. 3T-Z-spectra were evaluated using a multi-step procedure of conventional B0 correction, background removal, de-noising and multi-Lorentzian line fitting3. This yields pixel-wise vectors of optimized multi-Lorentzian parameters popt containing spectral position, width and amplitude of the amide (+3.5ppm), amine (+2.0ppm) and NOE (-3.5ppm) resonances. The uncorrected Z‑spectra were then used as inputs for a 3-layer feed-forward neural network (400 neurons in total, random weight initialization) with the measured B0 shift and the popt vector as target values. To avoid overfitting, a combination of an early stopping criterion and a regularization factor was used. Two differently trained networks were examined: The first (NN1) was trained only with a single subject dataset, whereas the second (NN2) was trained using the combined datasets of four subject measurements. To evaluate training variability, NN2 was trained multiple times with randomly initialized layer weights. All trained neural networks were applied to untrained test data of a fifth healthy subject of which the central slice is shown herein.

Results

Our deep learning approach is able to generalize well to an untrained dataset and to generate a prediction that closely matches the measured B0 map (Figure 1). However, the predicted B0 maps appear slightly noisier than the reference map. Training with more datasets improves the performance: The B0 prediction of NN1 shows a vertical structure in the middle of the brain slice with increased values compared to the reference image, which is also visible in the corresponding difference image (Figure 1D). This structure seems to resemble anatomical structures of the ventricles and is less present in the prediction of NN2 (E).

Both networks are able to predict CEST contrasts from uncorrected Z-spectra of an untrained dataset, as the predictions closely match the fitted reference contrasts regarding amplitude and spatial distribution (Figures 2, 3). It can be noticed that the difference images of NN2 (Figure 3, bottom row) show less anatomical structures, thus again the network trained on four datasets shows a better capability of generalization.

Figure 4 shows the variability of the nets due to randomly initialized training. In the first row, the mean predictions of NN2 when trained five times with identical parameters and training data are shown, which appear to predict less noisy B0 maps than only one net. The second row shows the corresponding standard deviation of the three predictions which are below 0.02ppm.

Discussion

A crucial issue when training deep neural networks is to provide a sufficient amount of data. This was ensured by the 3D-snapshot-CEST-sequence, which yielded ~70.000 Z-spectra per volunteer measurement used in NN1, thus ~300.000 for NN2. No spatial information was used for the training. The regularization factor was optimized manually; in combination with an early stopping criterion, this avoids overfitting.

The commonly employed procedures for B0 correction like WASABI2, WASSR4, or spline interpolation of Z-spectra5 involve interpolation or fitting of the direct saturation peak, which is computationally expensive and time consuming (several 10 minutes for 3D data). In contrast, applying a trained neural network to a dataset needs few seconds for 3D data. Moreover, the presented approach is able to directly generate CEST contrast from raw Z-spectra in one fast computation step, without the need for any additional field mapping measurement and correction procedure. Even though stability and reliability of the neural network predictions must be further investigated, those predictions can already be useful as initial values for established nonlinear least squares fitting evaluation.

Conclusion

Neural networks can be used to predict B0 maps from uncorrected Z-spectra. Beyond that, it is even possible to directly generate the desired CEST contrasts from uncorrected data without consideration of additionally measured B0 maps. This approach may allow for fast online calculations running directly on a MRI scanner providing CEST contrast within the clinical workflow.

Acknowledgements

Max Planck Society; German Research Foundation (DFG, grant ZA 814/2-1, support to KH); European Union Horizon 2020 research and innovation programme (Grant Agreement No. 667510, support to MZ, AD).

References

1. Zaiss , M., Ehses, P. and Scheffler K. Snapshot‐CEST: Optimizing spiral‐centric‐reordered gradient echo acquisition for fast and robust 3D CEST MRI at 9.4 T. NMR in Biomedicine. 2018;31:e3879. doi:10.1002/nbm.3879

2. Schuenke, P., Windschuh, J., Roeloffs, V., Ladd, M. E., Bachert, P. and Zaiss, M. (2017), Simultaneous mapping of water shift and B1(WASABI)—Application to field‐Inhomogeneity correction of CEST MRI data. Magn. Reson. Med., 77: 571-580. doi:10.1002/mrm.26133

3. Deshmane, A., Zaiss, M., Lindig, T., Herz, K., Schuppert, M., Gandhi, C., Bender, B., Ernemann, U. and Scheffler, K. MRM in press. doi:10.1002/mrm.27569.

4. Kim, M., Gillen, J., Landman, B. A., Zhou, J. and van Zijl, P. C. (2009), Water saturation shift referencing (WASSR) for chemical exchange saturation transfer (CEST) experiments. Magn. Reson. Med., 61: 1441-1450. doi:10.1002/mrm.21873

5. Terreno, E. , Stancanello, J. , Longo, D. , Castelli, D. D., Milone, L. , Sanders, H. M., B. Kok, M. , Uggeri, F. and Aime, S. (2009), Methods for an improved detection of the MRI‐CEST effect. Contrast Media Mol Imaging, 4: 237-247. doi:10.1002/cmmi.290

Figures

Figure 1: Comparison of B0 map predictions of NN1 and NN2 when applied to the untrained test dataset. (A) Relative B0 shift map obtained from WASABI as reference, (B) B0 map predicted from uncorrected Z-spectra by NN1, which was trained on a single volunteer dataset. (C) B0 map prediction by NN2, which was trained on all four volunteer dataset. (D) and (E) show the respective difference between the WASABI B0 map and the predictions. Color scales are given in units of ppm.

Figure 2: Prediction of CEST contrasts and B0 map made by NN1 (trained on one volunteer datasets) from an untrained test dataset. The first row shows peak amplitudes for amide, amine and NOE resonances, as obtained by the fitting procedure described in the methods section, together with the previously shown WASABI B0 map as references. The second row shows the corresponding predictions of the first neural net (NN1). In the third row, the differences between reference data (first row) and predictions (second row) are displayed.

Figure 3: Prediction of CEST contrasts and B0 map made by NN2 (trained on four volunteer datasets) from an untrained test dataset. The first row shows peak amplitudes for amide, amine and NOE resonances, as obtained by the fitting procedure described in the methods section, together with the previously shown WASABI B0 map as references. The second row shows the corresponding predictions of the first neural net (NN2). In the third row, the differences between reference data (first row) and predictions (second row) are displayed.

Figure 4: Variation of network training due to random initialization of layer weights. In the first row, the mean predictions of NN2 when trained five times with identical parameters and training data are shown. The second row shows the corresponding standard deviation of the five prediction.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4995