3719

Rapid multi-slice whole-brain B₁⁺-mapping at 7T using deep learning

Felix Krueger¹, Christoph Stefan Aigner¹, Max Lutz¹, Layla Tabea Riemann¹, Katja Degenhardt¹, Bernd Ittermann¹, Tobias Schaeffter^1,2, Kerstin Hammernik^3,4, and Sebasian Schmitter^1,5,6
¹Physikalisch-Technische Bundesanstalt, Berlin and Braunschweig, Germany, ²Division of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom, ³Technical University of Munich, Munich, Germany, ⁴Imperial College London, London, United Kingdom, ⁵Medical Physics in Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁶Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, United States

Synopsis

Keywords: RF Pulse Design & Fields, Parallel Transmit & Multiband

In this work we utilize deep learning to estimate multi-slice whole-brain B₁⁺-maps in sub-seconds from initial localizer scans at 7T. The investigated neural networks use the receive profiles of the individual coil elements of an 8Tx/8Rx transceiver head coil as input information. The networks are trained on seven volunteers and tested in 2 unseen subjects for transversal/coronal/sagittal slices by comparing the prediction with the acquired B₁⁺-maps. Subsequently, the feasibility of using the DL-based B₁⁺-maps in a subject-specific calibration pipeline is demonstrated.

Purpose

Subject-tailored parallel transmit (pTx) pulses for ultra-high fields (UHF) applications are typically calculated based on subject-specific data of all transmit (Tx) channels, which requires additional adjustment scans¹. In the human brain, transmit magnetic field (B₁⁺) maps based on pre-saturation acquisitions have been obtained in less than a minute with 8 Tx channels². Yet, a higher resolution or an increase to 16 or 32 Tx channels^3,4 will lead to calibration times of several minutes. Particularly for routine clinical scans, a fast, push-button, in-situ Tx field calibration embedded in the scanner's adjustment routine is highly desired. Deep learning (DL) based B₁⁺-mapping techniques have shown promising results in addressing this need and reducing calibration times^5,6.
This study investigates the feasibility of utilizing a recently presented DL-based method applied to single slices in the heart⁶ for rapid, whole-brain relative B₁⁺-mapping at 7T. The B₁⁺-maps are derived in sub-seconds from an initial multi-slice localizer scan obtained in a CP⁺-like mode, typically acquired for planning at the beginning of each scan session. Hence, the neural networks (NNs) use the receive (Rx) profiles (B₁^-) of the individual coil elements of an 8Tx/8Rx transceiver head-coil as input information. The NNs are trained on seven volunteers and tested in two unseen subjects by comparing DL-predictions with acquired B₁⁺-maps for different slice orientations.

Methods

Multi-slice B₁⁺-maps (8-10 slices, depending on anatomy) were measured in 9 subjects on a 7T whole-body MRI system (Magnetom 7T, Siemens, Erlangen, Germany) with a custom-built 8Tx/8Rx head-coil array⁷ utilizing a hybrid approach⁸. First a small flip-angle 2D multi-slice GRE image (T_R=100ms, T_E=2.9ms, resolution=2x2mm², FOV=256x192mm², slice thickness=4mm, nominal FA=10°, 15 slices) is obtained with all Tx channels transmitting using a default shim (i.e., the localizer). Then, the acquisition is repeated 8 times where only a single Tx channel is active each. The latter scans are then merged with a 3D actual flip-angle image (T_R=75ms, T_E=1.9ms, resolution=2x2x4mm³, FOV=256x192mm², slice thickness=4mm, nominal FA=60°) for all Tx coils active resulting in the ground truth (GT) multi-slice B₁⁺-maps.
The Rx-channel-wise localizer is split into real and imaginary parts and stacked with the magnitude image along the channel dimension, forming the NN's input resulting in an input size of 128x96x17 per slice. Similarly, the GT Tx-channel-wise B₁⁺-maps are split up into real and imaginary data and arranged to a size of 128x96x16. Normalizing the input and output gives the originally absolute data a relative nature. However, biases of relative methods (e.g., T₁/proton density)⁹ are avoided by this procedure.
The investigated NNs are modified UNets^10,4 with a symmetric loss¹¹ function. The first NN (NN_tra) was trained on 66 transversal slices from seven volunteers and tested on 17 transversal slices from two unseen subjects. The second (NN_cor) was trained on 86 coronal and tested on 17 coronal slices, and the third (NN_all) was trained on 219 transversal/coronal/sagittal slices and tested on 51 transversal/sagittal/coronal slices. The training parameters are as follows: ADAM optimizer, learning rate=1*10^-4, 2000 epochs, batch size=2. DL is performed on a 24 GB NVIDIA Titan RTX. The predicted (PR) maps are qualitatively compared to the GT evaluated by the root mean squared error (RMSE) and structural similarity index measure (SSIM). Static (phase-only) and dynamic (4kT points) pTx optimization was subsequently performed on the DL-based B₁⁺-maps for different optimization targets.

Results and Discussion

Fig.1 shows the magnitude and phase of the combined, multi-slice B₁⁺-maps for the GT and PR by NN_tra trained on multi-slice data in transversal orientation for one unseen test subject. The maps reflect the default shim setting b_def (no additional RF phases) resembling a CP⁺-mode. The DL-based B₁⁺-maps follow the GT for magnitude and phase (RMSE=0.053±0.010%). The match between PR compared to the GT B₁⁺-maps is further highlighted when looking at the channel-wise data for a central slice (Fig.2). Fig.3 depicts the B₁⁺-shimming results for the same example slice after calculating a homogeneous shim vector b_hom and an efficiency setting b_eff. The latter is highly useful, e.g., for subcutaneous fat saturation. The optimizations are based on the PR B₁⁺-maps and the resulting setting is applied to the PR and GT B₁⁺-maps. The GT and PR maps always reflect all features in the shimmed cases and for the default shim b_def . The feasibility of using DL-based B₁⁺-maps in a subject-specific calibration pipeline is demonstrated by a quantitative evaluation for the shim vectors b_hom and b_eff (Fig.3 (B)/(C)). This approach even provides good results for whole-brain dynamic pTx (4kt-points) excitation (Fig.3 (D)). Fig.4 shows the combined, multi-slice B₁⁺-maps for four coronal slices for the GT and predicted by NN_tra, NN_cor, and NN_all. While NN_tra trained on transversal slices fails to predict coronal slices accurately, NN_cor and NN_all successfully predict the combined maps. An evaluation of the PR and the GT combined B₁⁺-maps is shown for coronal slices for all NNs (Fig.5 (A)/(B)) and for all slice orientations predicted by NN_all (Fig.5 (C)/(D)).

Conclusion

Our results indicate the feasibility of estimating complex, channel-wise, multi-slice B₁⁺-maps in the human head at 7T from a quick GRE localizer scan in multiple orientations. Although we relied on relative B₁⁺-maps due to the normalization of the training data, an expansion to absolute data may be feasible.

Acknowledgements

We gratefully acknowledge funding from the German Research Foundation (GRK2260-BIOQIC).

References

1) Padormo F, Beqiri A, Hajnal JV, Malik SJ. Parallel transmission for ultrahigh-field imaging. NMR Biomed. 2016;29:1145-1161.

2) Herrler J, Liebig P, Gumbrecht R, Ritter D, Schmitter S, Maier A, Schmidt M, Uder M, Doerfler A, Nagel AM. Fast online-customized (FOCUS) parallel transmission pulses: A combination of universal pulses and individual optimization. Magn Reson Med. 2021; 85:3140–3153.

3) Orzada S, Solbach K, Gratz M, Brunheim S, Fiedler TM, Johst S, et al. A 32-channel parallel transmit system add-on for 7T MRI. Plos one. 2019; 14(9):e0222452.

4) Schmitter S, Wu X, Auerbach EJ, Adriany G, Pfeuffer J, Hamm M, Ugurbil K, Van de Moortele PF. Seven-tesla time-of-flight angiography using a 16-channel parallel transmit system with power-constrained 3-dimensional spoke radiofrequency pulse design. Invest Radiol. 2014; 49(5):314–325.

5) Eberhardt B, Poser BA, Shah NJ, Felder J. B1 field map synthesis with generative deep learning used in the design of parallel-transmit RF pulses for ultra-high field MRI. Z. Med. Phys. 2022; 1–12.

6) Krueger F, Aigner C, Hammernik K, Dietrich S, Lutz M, Menger J, Schaeffter T, Schmitter S. Rapid estimation of 2D relative B1+-maps from localizers in the human heart at 7T using deep learning. Magn Reson Med. 2022; 1-14.

7) Seifert F, Pfeiffer H, Mekle R, Waxmann P, Ittermann B. 7T 8-channel pTx head coil with high B1+ efficiency optimized for MRS. In: Proceedings of the 24th Annu. Meet. ISMRM, Singapore, Singapore. Abstract 3545, 2016.

8) Van de Moortele PF, Snyder C, DelaBarre L, Adriany G, Vaughan T, Ugurbil K. Calibration Tools for RF Shim at Very High Field with Multiple Element RF Coils: from Ultra Fast Local Relative Phase to Absolute Magnitude B1+ Mapping. In: Proceedings of the 15th Annual Meeting of ISMRM, Berlin, Germany: Abstract 1676; 2007.

9) Dietrich S, Aigner CS, Kolbitsch C, Mayer J, Ludwig J, Schmidt S, Schaeffter T, Schmitter S. 3D Free-breathing multichannel absolute B1+ Mapping in the human body at 7T. Magn Reson Med. 2021;85:2552–2567.

10) Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI. 2015;9351:234-241.

11) Terpstra ML, Maspero M, Sbrizzi A, van den Berg CAT. ⊥-Loss: a Symmetric Loss Function for Magnetic Resonance Imaging Reconstruction and Image Registration With Deep Learning. Med. Image Anal. 2022;80:102509.

Figures

Figure 1: Multi-slice, combined B₁⁺-maps for an unseen test subject. The NN_tra-predicted (B₁⁺_PR) and the ground truth (B₁⁺_GT) data are shown for the magnitude of sum (MOS) over the 8 Tx channels and the phase of the summed-up data (POS). The blue boxes mark the slice used in Figure 2.

Figure 2: B₁⁺-magnitude and phase maps for the prediction of NN_tra, B₁⁺_PR compared to the ground truth B₁⁺_GT for the unseen slice 4. The absolute error ΔB₁⁺ between the prediction and GT shows a higher residual error for the first and eighth Tx channel compared to channels 2-7. Overall, the prediction qualitatively matches the GT for both the magnitude and the phase.

Figure 3(A): Different B₁⁺-shimming results for subject slice 4. The default shim setting b_def, the homogeneous shim b_hom, the efficiency shim b_eff, applied to the predicted (B₁⁺_PR) and the ground truth (B₁⁺_GT) data are shown for the magnitude of the sum (MOS), and the phase of the sum (POS). (B)/(C) Coefficient of variation (CV)/ transmit efficiency per slice in the brain ROI for the default shim setting b_def and b_hom/ b_eff, optimized on the generated B₁⁺-maps and applied to the PR and the GT data. (D) Dynamic pTx (4kT points) calculted for the PR transversal data applied to PR and GT.

Figure 4: Multi-slice, combined B₁⁺-maps for four unseen coronal slices. The predicted (B₁⁺_PR) for the three introduced neural networks (NN_tra: trained on transversal slices, NN_corr: trained on coronal slices, NN_all: trained on all slice orientations) and the ground truth (B1+GT) data are shown for the magnitude of sum (MOS) over the 8 Tx channels and the phase of the summed-up data (POS).

Figure 5: Boxplots summarizing the RMSE (A) and the SSIM (B) between the predicted and the ground truth B₁⁺ data for unseen coronal slices predicted by the NN_tra, NN_cor, and NN_all. NN_all benefits from the slice variation in the training data resulting in a higher mean SSIM value for coronal slices. When relying on NN_all trained with all slice orientations, the image quality remains consistent for all slice orientations of the unseen test subjects (C)/(D).

Proc. Intl. Soc. Mag. Reson. Med. 31 (2023)

3719

DOI: https://doi.org/10.58530/2023/3719