Synopsis

Subject-specific local SAR assessment is one of the outstanding challenges of ultra-high field MRI. In this work, we present experimental, in-vivo results on four healthy subjects of a novel deep learning approach which provides online subject-specific local SAR distributions based on measured B₁⁺ maps. Results are validated by creating a subject-specific model of each subject and calculate a reference local SAR distribution off-line by FDTD simulations. The results show that a Convolutional Neural Network (CNN) trained with synthetic MR data enables accurate subject-specific local SAR prediction during an MRI examination.

PURPOSE

During an MRI examination, subject-specific local SAR cannot be measured, thus it is usually determined by off-line electromagnetic simulations using generic body models and the results are used as an estimate of the actual peak local SAR value^1,2,3. However, this approach requires additional safety factors to compensate for the deviations between the generic simulated model and the real, subject-specific, scenario.

Recently, a new approach based on deep learning was presented⁴. It consists of data-driven approach where a convolutional neural network (CNN) is trained to learn a “surrogate SAR model” to map the relation between subject-specific available MR data and the corresponding local SAR. In this work, we present the first in-vivo results of this on-line subject-specific local SAR assessment.

MATERIAL AND METHODS

Similarly to⁴, a Conditional Generative Adversarial Networks (cGAN)⁵ was trained to map the relation between the complex B­₁⁺map, i.e. B_­1⁺ magnitude and relative phase, and the 10g-average SAR (SAR_10g) distribution. The training was performed using simulated data by penalizing the underestimation error more⁴. Then, the trained network was used to predict SAR_10g distributions based on measured B₁⁺maps. For the validation, another independent pipeline was implemented to obtain simulated SAR_10g distributions by constructing dedicated models.

Synthetic Training set

A synthetic image dataset was generated by means of our database containing simulated B₁⁺ distributions and Q-Matrices¹ of 23 subject-specific models⁶ with a 8-fractionated dipole array for prostate imaging at 7T⁷(Figure 1.a). These simulated results were processed to produce 23x250=5750 different local SAR distributions (250 random phase settings for each model).

The extent of the averaging 10g cubes in the pelvic region is about 2cm, thus to capture all required information, each training sample consists of a volume of 5 consecutive, complex transverse B­₁⁺maps 0.5cm distance apart from one another, and the corresponding ground-truth SAR_10g distribution for the center slice (Figure 1.b).

To produce training samples with complex B­₁⁺maps similar to those that can be realistically acquired, the relative transmit B­₁⁺ phase distributions were obtained with respect to the first channel. Furthermore, since the DREAM⁸ method only presents a good accuracy for STEAM flip angles from 10° to 70°, the B₁⁺map regions outside this range were removed (Figure 1.b).

In-Vivo Validation

Four healthy volunteers not included in our database (BMI: 21/23/24/27 kg/m²) were scanned at 7T (Achieva, Philips Healthcare, Best, The Netherlands) with a 8-channel fractionated dipole array for prostate imaging⁶. Written consent was obtained according to local IRB regulations.

Image Acquisition: The following sequences were performed (FOV=430x430x100mm, Voxel size=2.8x2.8x5.0mm):

• B1default: DREAM B1-mapping with default phase settings (TR/TE=4.0/0.8ms,STEAM flip angle=40°);
• B1shimmed: DREAM B1-mapping with prostate shimmed phase settings $$$p$$$ (TR/TE=4.0/0.8ms,STEAM flip angle=40°);
• FFE-dynamic: 3D Fast Field Echo with subsequently each transmit channel active alone (TR/TE=5.00/1.95ms,flip angle=1°);
• DIXON: M2D Fast Field Echo (TR/TE=10.00/2.70ms,Voxel size=1.3x1.3x5.0mm).

Complex $$$B­_1^+$$$Maps: The measured images were combined as follows to obtain complex B₁⁺maps:

$$B1_{default}^{Measured}=|B1default|\exp\left(i\arg\left(\frac{\sum_{n=1}^{N_{ch}}\left(|FFE_{dyn}\left(n\right)|\exp\left(i\Phi\left(FFE_{dyn}\left(n\right)\right)\right)\right)}{\exp\left(i\Phi\left(FFE_{dyn}\left(1\right)\right)\right)}\right)\right)$$

$$B1_{shimmed}^{Measured}=|B1shimmed|\exp\left(i\arg\left(\frac{\sum_{n=1}^{N_{ch}}\left(|FFE_{dyn}\left(n\right)|\exp\left(i\Phi\left(FFE_{dyn}\left(n\right)+p\left(n\right)\right)\right)\right)}{\exp\left(i\Phi\left(FFE_{dyn}\left(1\right)\right)\right)}\right)\right)$$

Predicted $$$SAR_{10g}$$$ Distribution: The obtained complex B­₁⁺maps were the input for the trained cGAN to generate the predicted SAR_10g distributions (Figure 1.c).

$$SAR_{10g}^{Predicted}=cGAN\left(B1^{Measured}\right)$$

Simulated $$$SAR_{10g}$$$ Distribution: Following the pipeline presented in⁶, the DIXON images were used to build subject-specific 3-tissue models (fat/muscle/skin) of the volunteers and EM simulations (Sim4Life,ZurichMedTech,Switzerland) were performed (Figure 1.c).

For each model, the drive vector $$$s$$$ to minimize the L2-norm between the simulated and measured complex B₁⁺maps was evaluated numerically and used to calculate the reference SAR_10g distribution.

$$SAR_{10g}^{Simulated}=s^{\dagger}Q_{10g}s$$

RESULTS AND DISCUSSION

The cGAN⁵ was implemented in TensorFlow and trained for about 4 hours on a GPU (NVIDIA Tesla P100-PCIe-16GB).

In Figures 2-5, the measured and the simulated complex B₁⁺maps, and the predicted and simulated SAR_10g distributions are reported for each volunteer. Above each SAR_10g distribution the peak value (pSAR_10g) is reported. All validation tests showed a good qualitative and quantitative match between predicted and simulated SAR_10g distributions. In agreement with our previous study⁴, moderate pSAR_10g overestimation errors were observed (between 2% and 21%). Only one pSAR_10g underestimation error occurred (2%).

Note that complex B₁⁺maps inherently include all relevant and exam specific information such as the patient anatomy, antenna position and reflected/lost power.

CONCLUSION

Acknowledgements

References

¹I. Graesslin, H. Homann, S. Biederer, P. Börnert, K. Nehrke, P. Vernickel, G. Mens, P. Harvey, U. Katscher, A specific absorption rate prediction concept for parallel transmission MR. Magn Reson Med. 2012 Nov;68(5):1664-74.

²G. Eichfelder and M. Gebhardt, Local Specific Absorption Rate Control for Parallel Transmission by Virtual Observation Points, Magn. Reson. Med. 2011, vol. 66, no. 5, pp. 1468-1476.

³A. Sbrizzi, H. Hoogduin, J.J. Lagendijk, P. Luijten, G.L. Sleijpen and C.A.T van den Berg. Fast design of local N‐gram‐specific absorption rate–optimized radiofrequency pulses for parallel transmit systems. Magnetic resonance in medicine, 67(3), 824-834.(2012).

⁴E.F. Meliadò, A.J.E. Raaijmakers, M.H.F. Savenije, A. Sbrizzi, M. Maspero, P.R. Luijten, and C.A.T van den Berg. On-line Subject-Specific Local SAR Assessment by Deep Learning. Proceedings of the ISMRM 26th Annual Meeting, Paris, 16-21 June 2018.

⁵P. Isola, J. Zhu, T. Zhou, A.A. Efros, Image-to-Image Translation with Conditional Adversarial Networks, arXiv:1611.07004v1 [cs.CV] 21 Nov 2016.

⁶E.F. Meliadò, C.A.T van den Berg, P.R. Luijten, A.J.E. Raaijmakers. Intersubject specific absorption rate variability analysis through construction of 23 realistic body models for prostate imaging at 7T. Magn Reson Med. 2018;00:1–14. https://doi.org/10.1002/mrm.27518.

⁷A.J.E. Raaijmakers, M. Italiaander, I.J. Voogt, P.R. Luijten, J.M. Hoogduin, D.W.J. Klomp, C.A.T. van den Berg. The fractionated dipole antenna: A new antenna for body imaging at 7 Tesla. Magn Reson Med. 2015 May 2 : 10.1002/mrm.25596. Published online 2015 May 2. doi: 10.1002/mrm.25596.

⁸K. Nehrke and P. Börnert. DREAM—a novel approach for robust, ultrafast, multislice B1 mapping, Magnetic Resonance in Medicine 68:1517–1526 (2012).