Rigid Motion Compensated Compressed Sensing MRI with Untrained Neural Networks

Mohammad Zalbagi Darestani, Stefan Ruschke, Reinhard Heckel

Keywords: Compressed Sensing, Deep Learning, Unsupervised Learning, MRI Motion Correction

Abstract:

Deep neural networks trained end-to-end for accelerated magnetic resonance imaging give excellent performance. However, these networks are trained and evaluated under a setup where the object to be imaged is static.
However, in practice, patients often move during measurement acquisition which leads to motion artifacts in the reconstructed images.
In this work, we first demonstrate that if we train state-of-the-art neural networks to reconstruct an image for accelerated MRI under motion well, significantly larger training sets are required for good performance.
Secondly, we demonstrate that as an alternative, one can resort to utilizing untrained neural networks for this task. We propose a modified untrained network which does not rely on any training set and performs single-instance motion-compensated compressed sensing MRI. Our approach outperforms untrained and trained optimization-based baselines such as $\ell_1$-norm minimization and score-based generative models.

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7606