Synopsis

Keywords: CEST & MT, Data Analysis

Z-spectrum acquired under ultra-high field (> 3T) features stronger and better isolated CEST peaks than those under 3T. But from imaging aspect, 3T scanners perform better and are clinically accessible. Herein, we built a deep neural network (DNN) for predicting Z-spectrum under higher B₀ from the corresponding measurement at 3T. The network was trained by 10 million Z-spectra calculated from Bloch-equation models. Simulations with various B₀ shifts and noise suggested that 3T Z-spectra could be rapidly and accurately transformed to those under 7T or 9.4T. This network may help improve signal extraction and interpretation of CEST data acquired at 3T.

Introduction

In vivo Z-spectrum contains multiple signal contributions, including direct water saturation (DS), CEST effects, magnetization transfer contrast (MTC), and nuclear Overhauser effects (rNOEs), presenting as overlapped peaks under clinical field (e.g., 3T)¹. High-field static magnetic field facilitates the evaluation of CEST effects due to the well-isolated signal peaks in the Z-spectrum^2-4. However, high-field CEST data are difficult to obtain due to equipment limitations, specific absorption rate (SAR)⁵, and field inhomogeneity.
Some recent deep learning methods have been proposed to predict multi-pool Lorentzian parameters under high-field by learning Z-spectrum features under low B₀ field^6,7. However, only the Lorentzian parameters cannot fully reconstruct a realistic Z-spectrum. Here, we developed a deep neural network (DNN) transformation model to synthesize high-field Z-spectrum from the corresponding realistic low-field data. Bloch-equation simulated datasets under different B₀ fields were generated to verify the feasibility of the proposed method.

Methods

CEST Z-spectra were simulated by using a five-pool Bloch-McConnell model, including pools of water, amide, amine, rNOEs, and MTC⁸. The various combinations of tissue parameters with different concentrations and exchange rates were used to simulate Z-spectra⁹. The source B₀ was 3T, and the target B₀ were 7T and 9.4T, respectively. The paired Z-spectra under both low and high B₀ fields were generated over ±9.5ppm with 55 frequency offsets, B₁ power of 0.7μT. To mimic realistic scenarios, Gaussian white noise with an average of 0 and standard deviation (SD) of 0.35% was added to the 3T data. Additionally, B₀ shifts within ±0.25ppm and B₁ inhomogeneity within ±0.05μT were randomly introduced to the low B₀ field Z-spectra. Finally, a total of 10 million simulated Z-spectra were randomly divided into the training, validation, and test datasets with a ratio of 3:1:1.
Figure 1 illustrates the scheme of the proposed DNN-based Z-spectrum transformation model. The Z-spectrum (55 offsets) under low B₀ field is inputted to the model, which learns the features of these Z-spectral data and provides the corresponding Z-spectrum under high B₀ field. The DNN model used here is a conventional multi-layer perceptron, where the custom-weighted L1-loss (L1_W) is the model optimization objective:
$$\mathrm{L1}_W=\sum_{i=1}^n \alpha \cdot\left|y_i-f\left(x_i\right)\right| [1]$$
where n is the frequency offset number, x is the Z-spectrum value under the source field, y is the Z-spectrum value under the target field, f(.) is the model prediction results, and the customized weight factor ɑ = 2 if the frequency offsets are -3.5, -2.0, 0, 2.0, and 3.5ppm, otherwise ɑ = 1.
The performance of the proposed method is evaluated using the representative linear regression task evaluation indexes, including normalized root mean square error (NRMSE) and goodness of fit (R²) between the transformed Z-spectra and ground truth. To further evaluate the accuracy of the transformations, we also used the Z-spectrum analysis methods to quantify the CEST effects, including magnetization transfer ratio asymmetry (MTR_asym)¹⁰ and Lorentzian difference (LD)¹¹.

Results

Figure 2 shows the representative high B₀ field Z-spectra obtained from the corresponding low B₀ field Z-spectral data by using the proposed DNN-based Z-spectrum transformation model, including the ground truth. The linear regression evaluation results show R²=0.998, NRMSE=0.010 for 7T transformation, and R²=0.998, NRMSE=0.009 for 9.4T.
Results of MTR_asym and LD analyses for evaluating the DNN-based transformed 7T and 9.4 T Z-spectra are shown in Figures 3,4. The contrast maps of the DNN-based transformed Z-spectrum under high B₀ fields closely resemble the ground truth, with minor differences. Also, the statistical analysis illustrates no significant difference between the DNN-based transformed contrast signals and ground truth. Both present the increment of signal contrast compared with 3T data, except for the LD results at 3.5ppm under 7T. This might be due to the high concentrations of amide proton used in simulations, which can cause an isolated 3.5 ppm peak under 3T. The separation of this peak was not obvious under 7T but became significant at a much higher B₀ field (9.4T).

Discussion

The proposed DNN-based Z-spectrum transformation model can extract the features of Z-spectrum under low B₀ field and then generate reliable Z-spectral data under high B₀ field, allowing to provide well-isolated CEST signals. The isolated CEST peaks in Z-spectrum could be further extracted and quantified with enhanced accuracy by using Z-spectral analysis approaches, such as MTR_asym and LD. However, in this work, we only validated the feasibility of the transformation model by using preliminary simulations. In vivo studies need to be further performed. Due to the hard access of high B₀ field Z-spectral data, the proposed DNN-based Z-spectrum model provides great opportunities to obtain a good quality high B₀ field Z-spectral data without interference of low signal-to-noise ratio (SNR) and field inhomogeneities, facilizing the clinical use of high B₀ field CEST data.

Conclusion

We developed a DNN-based transformation model to rapidly and accurately synthesize Z-spectrum under high B₀ field according to the corresponding low B₀ field data. This network may be helpful for improving signal extraction and interpretation of CEST data acquired at 3T, but further validations using phantom or in vivo experiments are required.

Acknowledgements

No acknowledgement found.