[Paper Review] Deep Learning for Retrospective Motion Correction in MRI: A Comprehensive Review

Veronika Spieker, Hannah Eichhorn+5 more|arXiv (Cornell University)|May 11, 2023

Advanced MRI Techniques and Applications8 citations

TL;DR

A comprehensive survey of learning-based retrospective motion correction (MoCo) methods in MRI, detailing data usage, architectures, training, evaluation, and future directions.

ABSTRACT

Motion represents one of the major challenges in magnetic resonance imaging (MRI). Since the MR signal is acquired in frequency space, any motion of the imaged object leads to complex artefacts in the reconstructed image in addition to other MR imaging artefacts. Deep learning has been frequently proposed for motion correction at several stages of the reconstruction process. The wide range of MR acquisition sequences, anatomies and pathologies of interest, and motion patterns (rigid vs. deformable and random vs. regular) makes a comprehensive solution unlikely. To facilitate the transfer of ideas between different applications, this review provides a detailed overview of proposed methods for learning-based motion correction in MRI together with their common challenges and potentials. This review identifies differences and synergies in underlying data usage, architectures, training and evaluation strategies. We critically discuss general trends and outline future directions, with the aim to enhance interaction between different application areas and research fields.

Motivation & Objective

• Summarize motion artefacts in MRI and existing conventional MoCo approaches.
• Systematically review learning-based MoCo methods across image- and k-space-based strategies.
• Analyze data availability, motion simulation, training objectives, and evaluation metrics used in the field.
• Identify differences, synergies, challenges, and future directions to foster cross-field collaboration.

Proposed method

• Classify MoCo methods into image-based and k-space-based architectures.
• Discuss hybrid classical-learning and pure learning-based reconstruction frameworks.
• Describe training objectives including voxel-wise, image-domain, and k-space losses, plus adversarial and self-supervised approaches.
• Outline motion simulation strategies for brain, cardiac, and abdominal MRI to enable supervised training.
• Summarize evaluation metrics including full-reference and reference-free image quality measures, and motion detection/estimation metrics.

Experimental results

Research questions

• RQ1 What are the main learning-based strategies for retrospective MoCo in MRI (image-based vs. k-space-based)?
• RQ2 How do data usage, training objectives, and evaluation practices differ across MoCo methods and anatomies?
• RQ3 What are the common challenges and potential synergies across reconstruction and motion-correction methods?
• RQ4 How is motion simulated and what impact does it have on training and evaluation?
• RQ5 What directions are likely to advance interaction between ML and MRI communities in this area?

Key findings

• There is a broad spectrum of architectures: image-based, k-space-based, hybrid model-based with learning components, and pure learning-based end-to-end reconstructions.
• Many methods rely on supervised training with paired motion-corrupted and motion-free data, but motion simulation is commonly used due to limited paired datasets.
• Training objectives span L1/L2, SSIM, perceptual and adversarial losses, with self-supervised and cycle-consistent (CycleGAN) approaches used to handle unpaired data.
• Evaluation combines full-reference and reference-free image quality metrics (e.g., SSIM, PSNR, MSE, SNR, CNR, Tenengrad) and may include motion detection/estimation metrics and downstream task performance.
• Retrospective MoCo can be integrated with reconstruction (unrolled networks, DC blocks, motion estimation) and with downstream tasks such as segmentation, emphasizing cross-task benefits.
• The review highlights differences and synergies in data usage, architectures, training, and evaluation, calling for stronger interaction between ML and MRI communities.