Siyuan Liu1, Kim-Han Thung1, Weili Lin1, Pew-Thian Yap1, and the UNC/UMN Baby Connectome Project Consortium2
1Biomedical Research Imaging Center, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States, 2University of North Carolina at Chapel Hill, Chapel Hill, NC, United States
Synopsis
Retrospective artifact
removal using supervised learning requires explicit generation of
artifact-corrupted images and is impractical since generating the wide variety of potential
artifacts can be challenging. Using unsupervised learning, we show how
artifacts can be disentangled with remarkable efficacy from artifact-corrupted
images to recover the artifact-free counterparts, without requiring explicit
artifact generation.
Purpose
Structural magnetic
resonance imaging (sMRI) is highly susceptible to motion artifacts, which can
be difficult to avoid especially for pediatric subjects1. Deep
learning retrospective artifact correction (RAC)2 can be employed to
improve image quality. However, acquiring a large amount of paired data, i.e., images
with and without artifacts, for supervised training is impractical since generating
data for a wide range of artifacts is challenging. In this abstract, we
demonstrate that RAC can be carried out effectively via a cycle-consistency
adversarial network that is trained with unpaired artifact-free and artifact-corrupted
images.Method
A. Data
Preparation: From T1- and T2-weighted MR volumes of pediatric subjects from
birth to age six, 20 artifact-free and 20 artifact-corrupted volumes were selected
for training and 5 artifact-corrupted volumes were selected for testing. 1620, 1600 and 425 axial slices were extracted, respectively, from
the 20 artifact-free, 20 artifact-corrupted, and 5 artifact-corrupted
T1-weighted volumes. 1520, 1550 and 425 axial slices were extracted,
respectively, from the 20 artifact-free, 20 artifact-corrupted, and 5 artifact-corrupted
T2-weighted volumes.
B. Network
Architecture: We consider two image domains: artifact-free and
artifact-corrupted. To learn these domains from the data, similar to CycleGAN3,
we employ two auto-encoders to learn a cycle translation (Fig. 1) that translates
forward and backward the images in the two domains. During training, two patchGAN4
discriminators are used to distinguish between translated and real images in
each domain. Each auto-encoder consists of two encoders to disentangle
content and artifact information (Fig. 2). For complete disentanglement of
content and artifact, we propose a content-swapping mechanism where the translated
image in each domain is constructed with the content information of the
opposite domain. We also enforce that when the input image is artifact-free, the output
is unaltered with no removal of image details.
C. Loss Functions: In addition to pixel and
perceptual cycle-consistency losses, we propose a multi-scale content
consistency loss based on pixel, low- and high-level content features between images.
For adversarial learning, we employ two least-squares adversarial losses. For
quality-maintaining learning, we use a pixel-wise consistency loss to enforce
identity translation mappings.Results
The corrected results for different
levels of artifacts T1- and T2-weighted images are shown in Fig. 3 and 4,
respectively. It can be observed that the artifacts are removed without
significantly without introducing new artifacts. The results for the quality
maintaining of T1- and T2-weighted images are summarized in Fig. 5, from which
we can observe that the corrected images are highly consistent with the input
artifact-free images.Conclusions
We have
demonstrated that our disentangled cycle-consistency adversarial network
achieves remarkable efficacy in artifact correction and yields high-quality
corrected images. Thanks to the
unsupervised nature of our method, the artifacts do not need to be explicitly
specified.Summary of Main Findings
Our network
reduces artifacts in MR images without specifying the nature of the artifacts.
This potentially allows image imperfections such as noise, streaking, and ghosting
to be removed without explicit generating them for supervised training.Acknowledgements
This work was supported in part by NIH grants
(EB006733 and 1U01MH110274). References
1. Zhu, J.,
Gullapalli, R. P., MR artifacts, safety, and quality control. RadioGraphics 26,
275-297 (2006).
2. Zaitsev, M.,
Maclaren, J. & Herbst, M. Motion artifacts in MRI: A complex problem with
many partial solutions. Journal of Magnetic Resonance Imaging 42, 887–901
(2015).
3. Zhu, J.-Y.,
Park, T., Isola, P., Efros, A. A. Unpaired image-to-image translation using
cycle-consistent adversarial networks. IEEE International Conference on
Computer Vision (ICCV) (2017).
4. Isola,
P., Zhu, J.-Y., Zhou, T., Efros, A. A. Image-to-image translation with
conditional adversarial networks. IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), (2017).