Jan Hendrik Wuelbern1, Ulrich Katscher1, Karsten Sommer1, Axel Saalbach1, and Jalal B Andre2
1Philips Research Europe, Hamburg, Germany, 2University of Washington, Seattle, WA, United States
Synopsis
Since tissue
conductivity is determined by the numerical second derivative of the phase map,
it is particularly susceptible to motion. This abstract investigates the application
of deep-learning based methods for retrospective correction of motion artifacts
to obtain suitable phase maps as input for conductivity reconstruction.
Different types of motion were investigated in the framework of volunteer
experiments, revealing that the applied motion correction was indeed capable of
improving conductivity reconstruction.
Introduction
Tissue conductivity
might be able to serve as important biomarker, particularly for oncologic
applications. It is typically reconstructed by numerical calculation of the
second derivative of the measured transceive phase map [1], and thus, is very
sensitive to any patient motion during acquisition. This study investigated the
impact of motion (here: intentional motion in the framework of volunteer brain
experiments) on reconstructed conductivity, as well as the possibility to
remove motion artefacts by deep-learning techniques prior to conductivity
reconstruction. It turned out that this method is able to stabilize
conductivity reconstruction for all types of motion investigated.Methods
MR Acquisition
Data was acquired on a
Philips Ingenia 3T scanner (Best, The Netherlands) using a head-neck receive
coil. A T2-weighted multi-slice turbo spin-echo sequence was used with TR = 3 s,
TSE factor = 16, TE = 80 ms, echo
spacing = 9.4 ms. Geometric parameter were chosen as follows: FOV
230 x 182 mm2, 0.55 x 0.65 mm2
voxel size in 28 slices with 4 mm slice thickness. Following written
informed consent obtained according to local Institutional Review Board guidelines,
a healthy volunteer performed deliberate, specified motion during data
acquisition to induce varying severity of motion artifacts within the resulting
images. These included remaining as still as possible (as a baseline
experiment), performing repeated coughing, continuous leg shaking, continuous
head movement, performing a single shift in head position, and performing continuous
eye movement.
Motion Correction
Motion
artifacts in the input data were corrected separately for magnitude and phase.
To this end, two tailored fully convolutional networks were trained on a
dataset of MR image pairs with and without artificially generated motion artifacts. The training dataset was based on 16
T2-weighted (multi-2D spin echo, magnitude data only) whole-brain scans
obtained in 16 patients. Following informed patient consent, all data was
anonymized, and all images were deemed to be artifact-free by an experienced,
blinded neuroradiologist. Synthetic phase images were created using a 3rd-order
2D polynomial with randomly selected coefficients and added to the magnitude
data. For both networks, artifact-corrupted magnitude and phase was provided as
input, but only artifact-free magnitude (network 1) or phase (network 2) data
was used as target. Artifact simulation was realized using a previously
described pipeline [2]. A U-Net with 19 convolutional layers and 5 down-/up-sampling
operations in the contracting/expanding path was implemented using the PyTorch
framework.
After training, both
networks were applied to magnitude and transceive phase images to correct for
motion artifacts. The transceive image was used as input to the EPT
reconstruction as outlined below, while the magnitude images are only required
here to illustrate anatomical structures and artefacts.
EPT Reconstruction
Conductivity $$$\sigma$$$
was reconstructed by numerical differentiation from the uncorrected and
corrected transceive phase $$$\phi$$$, respectively, according to $$$\sigma = \nabla^2 \phi /
(2\mu\omega)$$$ (with vacuum
permeability $$$\mu$$$ and Larmor frequency $$$\omega$$$). The kernel of the numerical differentiation
as well as of the subsequent denoising filter was locally shaped to exclude
tissue boundaries (as identified from the magnitude image) to avoid the
otherwise occurring boundary artefacts [1].Results
Anatomical images
before and after U-net correction for the baseline experiment (no motion) and
continuous leg shaking are presented in Fig. 1. The characteristic ghosting
artifacts due to subject motion are well visible before the correction and
significantly reduced after correction.
Figure
2 presents the derived conductivity maps for the same cases as in Fig. 1.
Structures of the cortex are blurred or obstructed in the motion case without
correction. After correction, much more detail is recovered in the conductivity
map. As expected, no significant difference is visible before and after
correction for the case without motion.
Correlation
coefficients of the conductivity maps of the corrupted and corrected motion
cases calculated with respect to the corrected no motion case (applying
corresponding elastic image registration [3] to correct for inter-scan motion
and accompanying shifting of structures) are presented in Tab. 1. After
network processing, the correlation to the baseline experiment was improved for
all motion cases.Discussion
In our experiments, the
resulting conductivity maps showed an improved correlation to the baseline map
independent of the motion case, demonstrating the versatility of the approach.
Moreover, the correction network did not decrease the quality of the conductivity
map in an acquisition without motion of the subject.Conclusion
Retrospective
motion artifact removal in MR phase maps with deep learning methods
substantially improve the quality of EPT reconstructed brain conductivity maps.
This is noteworthy, as the conductivity is derived from the second order
derivative, making this type of reconstruction particularly sensitive to motion
artifacts. The proposed method hence appears to facilitate enhanced motion
robust conductivity mapping.Acknowledgements
No acknowledgement found.References
[1] U. Katscher et
al., Electric properties tomography: Biochemical, physical and technical background,
evaluation and clinical applications. NMR Biomed. 2017; 30:3729
[2] K. Sommer, et al.,
Correction of motion artifacts using a multi-resolution fully convolutional
neural network, Proc. Intl. Soc. Mag. Reson. Med. 26, #1175, 2018.
[3] S. Kabus and C.
Lorenz, “Fast elastic image registration,” in Medical Image Analysis For The
Clinic - A Grand Challenge, 2010, pp. 81–89