Presentation for MR motion artifact correction using outlier-rejecting bootstrap aggregation (IEEE TMI, 2021)

1. Unpaired MR Motion Artifact Deep Learning
Using Outlier-Rejecting Bootstrap Aggregation
Gyutaek Oh, Jeong Eun Lee, and Jong Chul Ye

2. I. Introduction
• Scan time of MRI is long  motion artifact
• Existing deep learning methods for motion artifact correction: supervised learning
• Based on simulated motion artifact data
• Real motion artifact data obtained in controlled experiments
 Difficult to use in real situations, obtain matched clean and artifact images
Unpaired Deep Learning Methods for Real Motion Artifact Correction

3. II. Theory
Motion Artifact
• Patient’s motion  k-space phase error at the specific phase encoding line
• k-space outlier along the phase encoding direction can be assumed sparse
Clean image k-space of clean image
Fourier transform Motion
k-space of artifact image
Inverse Fourier transform
Artifact image
: motion-corrupted k-space
: motion-free k-space
: phase error

4. II. Theory
Bootstrap Aggregation
• Random sampling along the phase encoding direction  remove sparse outliers
• Reconstruction of  reduce the contribution of motion artifacts
• Bootstrap aggregation of several reconstructed images
 much closer to the artifact-free image
: k-space subsampling
: reconstructed image
: reconstruction network
: weighting factor

5. II. Theory
5
Clean image k-space of clean image Downsampled k-space of clean image Downsampled clean image
Artifact image k-space of artifact image Downsampled k-space of artifact image Downsampled artifact image
≠ ≈

6. III. Method
Training phase
• Network is trained to reconstruct the downsampled clean image to fully sampled clean image

7. III. Method
Test phase
• Several downsampled artifact images are reconstructed
• Aggregate reconstructed images  motion corrected image

8. III. Method
1. Experiments using simulated data
• Use simulated motion artifact data
• Brain, knee: random motion
• Liver: periodic motion due to the breathing
2. Experiments using in vivo data
• Use in vivo motion artifact data
• Liver: Gd-EOB-DTPA-enhanced MR, arterial phase
[1]
[1] Tamada, Daiki, et al., Magnetic Resonance in Medical Sciences 19.1 (2020): 64.

9. 1. MARC[1]
• Supervised learning method based on simulated data
• Training using simulated data  testing using simulated data
• Training using simulated data  testing using in vivo data
2. Cycle-MedGAN[2]
• Unpaired learning method based on in vivo data
• Training using simulated data  testing using simulated data
• Training using in vivo data  testing using in vivo data
III. Method
[1] Tamada, Daiki, et al., Magnetic Resonance in Medical Sciences 19.1 (2020): 64.
[2] Tamada, Daiki, et al., Magnetic Resonance in Medical Sciences 19.1 (2020): 64.

10. IV. Experimental Results
Artifact Ours MARC Cycle-MedGAN Ground Truth
Simulation (2D motion)– brain

11. IV. Experimental Results
Artifact Ours MARC Cycle-MedGAN Ground Truth
Simulation (3D motion)– brain

12. IV. Experimental Results
Artifact Ours MARC Cycle-MedGAN Ground Truth
Simulation (2D motion)– knee

13. IV. Experimental Results
Artifact Ours MARC Cycle-MeGAN Ground Truth
Simulation (2D motion)– liver

14. IV. Experimental Results
Artifact Ours MARC Cycle-MedGAN
In vivo – liver

15. IV. Experimental Results
Quantitative evaluation (simulation data)

16. IV. Experimental Results
Clinical evaluation (in vivo data)
Artifact Ours MARC Cycle-MedGAN

17. V. Conclusion
• Unpaired MRI motion artifact correction algorithm using the bootstrap subsampling
aggregation
• Convert motion artifact correction problem to k-space outlier-rejecting bootstrap subsampling
and aggregation approach for MR reconstruction
• Our method outperforms other existing methods in terms of qualitative and clinical evaluation
• The proposed method may be an important platform for MRI motion artifact correction when
paired clean data do not exist