Deep Learning for Retrospective Motion Correction in MRI

This direction is split into other directions: motion-induced image degradation. Its last version stays readable and citable here.

Version 1 of 1214 papers224 conceptsfrozen 2026-10-04

Medical imaging involves a fundamental trade-off between spatial resolution and acquisition time. High-fidelity modalities like Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) require minutes to capture a single 3D volume. During this window, any patient movement—whether voluntary (a child shifting) or involuntary (breathing, heartbeat, or swallowing)—corrupts the data.

In MRI, data is collected in "k-space" (the spatial frequency domain). Movement during acquisition causes inconsistencies in this domain, which manifest as blurring, ghosting, and geometric distortions in the reconstructed image. Traditionally, these artifacts required discarding the scan and starting over, a costly and inefficient process. A new wave of research is moving toward retrospective motion correction, using deep learning to "untangle" motion from corrupted data after the scan is complete.

Joint Reconstruction and Registration

A significant shift in the field is the move from sequential processing to "multi-tasking" frameworks. Historically, researchers would first try to estimate the motion (registration) and then fix the image (reconstruction). However, errors in the first step inevitably ruin the second.

Angelica I Aviles-Rivero and Veronica Corona proposed a variational framework in Multi-tasking to Correct: Motion-Compensated MRI via Joint Reconstruction and Registration that treats these as a single optimization problem. By solving for the clean image and the deformation map simultaneously, the model ensures that the estimated motion is physically realistic and consistent with the raw k-space data. This "joint" approach prevents error propagation and produces sharper images even from highly undersampled data.

Leveraging Generative Priors

If a scan is heavily corrupted, some information is effectively lost. To recover it, researchers are using generative models that have "learned" what a clean human organ should look like. These models act as a powerful statistical prior, filling in the gaps caused by motion artifacts.

For example, Brett Levac and Jonathan I. Tamir developed diffusion-based generative models for neonatal MRI in Accelerated, Robust Lower-Field Neonatal MRI with Generative Models. Because infants are highly prone to movement, these models are trained to recognize the underlying anatomy of the infant brain, allowing them to suppress noise and ghosting while maintaining clinical accuracy. Similarly, Jingyu Guo introduced MAMOC: MRI Motion Correction via Masked Autoencoding, which uses self-supervised learning on massive "clean" datasets to learn how to reconstruct high-fidelity brain images from corrupted inputs.

Test-Time Adaptation

One of the hardest challenges in motion correction is that every patient moves differently. A model trained on "simulated" motion might fail when faced with a unique real-world movement pattern.

To solve this, Tobit Klug developed MotionTTT: 2D Test-Time-Training Motion Estimation for 3D Motion Corrected MRI. Instead of relying solely on pre-training, this framework optimizes motion parameters for the specific patient during the reconstruction process. By minimizing a "data consistency loss"—essentially checking if the corrected image still matches the original raw data—the model adapts to the specific rigid motion of that individual scan without needing a ground-truth "clean" version of that patient.

Acceleration as a Defense Mechanism

While retrospective correction fixes artifacts, another branch of research aims to prevent them by making scans too fast for motion to occur. This is particularly vital for quantitative imaging, where researchers measure specific tissue properties like oxygenation or water diffusion.

Hongyu Li and colleagues introduced SuperDTI, a framework that accelerates Diffusion Tensor Imaging by up to 15x. As detailed in Deep Learning for Highly Accelerated Diffusion Tensor Imaging, by bypassing traditional noise-sensitive fitting processes, they can generate high-quality fiber tractography from a fraction of the usual data. Similarly, Peng Lai demonstrated in Sub-8-minute cardiac four dimensional flow MRI that spatiotemporal parallel imaging can reduce cardiac scan times to under 8 minutes, effectively "outrunning" the artifacts that typically plague free-breathing protocols.

Automated Quality Control

As these deep learning tools move into the clinic, the final hurdle is knowing when a scan is too corrupted to be saved. In large-scale studies like the UK Biobank, manual review is impossible. Wenjia Bai has pioneered automated pipelines, as seen in Large-scale Quality Control of Cardiac Imaging in Population Studies, which use machine learning to flag motion-corrupted scans instantly. This ensures that only high-quality data reaches the diagnostic stage, while also identifying specific patient populations—such as those with respiratory disease—who may need specialized, motion-robust imaging protocols.

Papers in this version

Showing 60 of 214 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_deep_learning_for_retrospective_motion_c_67b6bb6b,
  title = {Deep Learning for Retrospective Motion Correction in MRI},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-learning-for-retrospective-motion-correction-in-mri/rdr_67b6bb6bece54701bd61c7c144331d70},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_deep_learning_for_retrospective_motion_c_67b6bb6b_v1,
  title = {Deep Learning for Retrospective Motion Correction in MRI},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-learning-for-retrospective-motion-correction-in-mri/rdr_67b6bb6bece54701bd61c7c144331d70/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}