💻 computer science

Motion Artifact-Aware Self-Supervised Representation Learning for 3D Brain MRI Motion Artifact Reduction

The paper introduces SSRL-MAR, a self-supervised, unpaired deep learning framework that effectively reduces motion artifacts in 3D brain MRI by leveraging a three-stage training strategy to achieve near-oracle performance without requiring paired clean-corrupted data or explicit motion labels.

Original authors: Mojtaba Safari, Shansong Wang, Zach Eidex, Matthew Goette, Tonghe Wang, Zhen Tian, Xiaofeng Yang

Published 2026-08-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mojtaba Safari, Shansong Wang, Zach Eidex, Matthew Goette, Tonghe Wang, Zhen Tian, Xiaofeng Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to take a perfect, high-definition photo of a tiny, intricate castle made of glass. If your hand shakes even a little while you snap the picture, the image comes out blurry, with ghostly double images floating around the towers. Now, imagine that instead of a camera, you are using a giant, super-sensitive scanner to look inside a human brain. The brain is that glass castle, and the scanner is the camera. But here's the catch: people can't hold perfectly still for the minutes it takes to scan a brain. They might twitch, cough, or just drift off to sleep. This movement creates "motion artifacts"—blurry streaks, ghostly duplicates, and warped shapes that ruin the picture.

For scientists and doctors, these blurry pictures are a nightmare. They need crystal-clear images to measure the size of brain parts, track diseases, or understand how the brain works. Usually, to fix a blurry photo, you'd need to know exactly what the original, perfect photo looked like so you could compare the two and fix the mistakes. But in the real world of medical scans, you can't go back in time to take a "perfect" picture of the same person's brain while they are moving. You only have the messy, blurry one. This paper tackles the problem of how to clean up these blurry brain scans without ever having seen the "perfect" version to begin with.

The researchers behind this study, led by Mojtaba Safari and Xiaofeng Yang, have built a clever new tool called SSRL-MAR. Think of it as a digital detective that learns to fix blurry brain scans by playing a game of "spot the difference" and "make a mess," all without needing a reference standard.

Here is how their system works, broken down into three fun steps:

Step 1: The "Spot the Difference" Training
First, the computer needs to learn what "motion" actually looks like. Imagine you have a stack of clean, sharp photos of different people's brains and a stack of blurry, messy ones. The computer doesn't know which is which, but it is told to find the blurry ones. It looks at tiny 3D chunks of the images and tries to figure out: "Does this chunk look like it was shaken?" It learns to separate the shape of the brain (the castle) from the blur (the shaking hand). The paper shows that the computer gets really good at this, learning to ignore the specific details of the person's brain and focusing only on the weird, wavy patterns caused by movement.

Step 2: The "Mess-Maker" Simulator
Once the computer knows what motion looks like, it gets a new job: it becomes a prankster. It takes a clean brain scan and tries to add the exact kind of blur it learned in Step 1. It's like a video game character learning to throw mud so well that it can recreate a muddy scene from scratch. The paper notes that this "degrader" network is very good at making fake blurry scans that look almost exactly like real ones. This is crucial because it allows the computer to create its own training data. It can take a clean scan, make it messy, and then try to fix it, effectively teaching itself how to clean up the mess it just made.

Step 3: The "Restorer"
Finally, the computer plays the hero. It takes the messy scans it created in Step 2 and tries to turn them back into clean images. Because it knows exactly what kind of mess it made, it can figure out how to undo it. This whole process happens in a loop: learn the blur, make the blur, fix the blur. The paper calls this a "closed-loop" system because it doesn't need anyone to tell it the right answer; it figures it out by practicing on its own.

What Did They Find?
The team tested their new tool on two types of data. First, they used "in-silico" data, which means they took real brain scans and used a computer to artificially add motion to them. In these tests, their method improved the image quality significantly, boosting a score called PSNR (which measures how close the image is to perfect) to 23.81 dB and making the images look 91.55% similar to the original structure (SSIM).

Then, they tried it on real, messy scans from actual patients (the MR-ART dataset). This is the tricky part because they didn't have the "perfect" version to compare against. Even without that, their tool successfully reduced the blur and ghosting. When they compared their method to other smart computer programs, SSRL-MAR came out on top. It was even better than a "source-only" supervised model (a program trained only on fake data) by up to 2.0 dB in PSNR.

The paper also looked at how well the tool preserved the brain's actual shape. They measured the volume of important parts like the corpus callosum (the bridge between the brain's two halves) and the ventricles (fluid-filled spaces). In cases of mild motion, the error in measuring these parts dropped by more than 50% after the tool cleaned up the image. This suggests the tool doesn't just smooth out the picture; it actually keeps the brain's anatomy accurate.

What They Don't Claim
It's important to note what this paper doesn't say. The authors are careful to point out that while their tool is great, it isn't magic.

It's not a perfect fix for extreme motion: If the motion is too severe (like a huge shake), the tool can't recover the lost information. The paper shows that as motion gets worse, the image quality still drops, just not as badly as it would without the tool.
It doesn't invent new things: The researchers were worried the computer might "hallucinate" (make up) brain parts that weren't there. But their tests showed the opposite: the tool is conservative. It tends to smooth things out rather than inventing fake details, which is safer for medical use.
It's not a cure-all for every scan type yet: The study focused specifically on T1-weighted brain scans. The paper suggests that applying this to other types of scans (like T2 or FLAIR) or other body parts would need more testing.

The Bottom Line
The paper suggests that SSRL-MAR is a robust and scalable way to clean up brain MRI scans without needing those impossible-to-get "perfect vs. messy" pairs of images. By teaching a computer to understand motion, create fake motion, and then fix it, they created a system that works almost as well as a "perfect" supervised model (which would require real paired data that doesn't exist in practice). The authors conclude that this approach could help doctors and researchers get clearer, more reliable measurements from large groups of patients, even if those patients couldn't hold perfectly still during their scan. It's a step toward making brain imaging more reliable, even when the human body refuses to stay still.

Technical Summary: Motion Artifact-Aware Self-Supervised Representation Learning for 3D Brain MRI

Problem Statement
Patient motion during Magnetic Resonance Imaging (MRI) acquisition remains a critical source of image degradation, causing signal loss, blurring, and geometric distortions that compromise quantitative neuroimaging analysis. While retrospective motion correction is essential for large-scale studies, existing deep learning solutions face significant limitations. Supervised methods rely on paired motion-corrupted and motion-free data, which are rarely available in clinical or population-scale archives. Conversely, unsupervised approaches often suffer from training instability, hallucination of anatomical details, or dependence on raw k-space data that is frequently unavailable in retrospective cohorts. There is a critical need for a robust, scalable framework that operates directly in the image domain without requiring paired supervision or explicit motion labels.

Methodology: SSRL-MAR
The authors propose SSRL-MAR (Self-Supervised Representation Learning for Motion Artifact Reduction), a three-stage, unpaired framework designed to learn, simulate, and correct motion artifacts in 3D brain MRI. The method operates entirely in the image domain through a closed-loop self-supervised pipeline:

Motion Artifact Representation Learning (Stage 1):
A contrastive encoder is trained on unpaired 3D patches (96³ voxels) from clean and motion-corrupted volumes. Using an anisotropic InfoNCE loss, the network learns to disentangle anatomical content from motion patterns. By contrasting clean patches against corrupted ones (where negative samples come from anatomically distinct subjects), the encoder is forced to become invariant to anatomy while remaining sensitive to motion-induced distortions. This results in a latent space where motion embeddings are separated from clean embeddings.
Motion Artifact-Aware Synthesis (Stage 2):
A motion-conditioned degrader network (based on a U-Net backbone) synthesizes realistic motion artifacts from clean scans. This network utilizes the motion representations extracted in Stage 1 to condition the degradation process. It approximates the conditional distribution $p(x_m | x_c, m)$ , generating pseudo-corrupted volumes that mimic the statistical properties of real motion artifacts without requiring ground-truth motion parameters. The training employs a composite loss including content consistency, adversarial loss (PatchGAN), and the contrastive loss to ensure the synthesized artifacts align with the learned motion manifold.
Motion Artifact-Aware Correction (Stage 3):
A motion-aware generator (also U-Net based, incorporating Swin Transformer blocks) is trained to restore clean volumes from the synthesized corrupted pairs. This stage utilizes the learned motion representations as priors to guide the restoration. The generator is optimized using a composite loss function that enforces both intensity consistency and gradient consistency (preserving edges and fine anatomical details) against the original clean inputs used to generate the pseudo-corrupted data.

The training strategy is sequential: the encoder is pre-trained and frozen, followed by the degrader, and finally the generator. For in-vivo data, the framework supports unsupervised domain adaptation, where the encoder and subsequent networks are fine-tuned on unlabeled, motion-corrupted real-world scans to adapt to the target domain without paired references.

Key Contributions

Unpaired Framework: SSRL-MAR is the first framework to unify contrastive motion representation learning with learned degradation modeling and motion-conditioned correction for 3D brain MRI, eliminating the need for paired training data or raw k-space.
Anatomy-Invariant Representation: The proposed contrastive encoder learns motion embeddings that are explicitly invariant to subject-specific anatomy, enabling the synthesis of realistic artifacts conditioned on motion rather than anatomical structure.
Novel Degrader Network: A motion-conditioned degrader generates realistic artifacts directly in the image domain, creating a self-supervised training signal that bridges the gap between simulation and reality.
Comprehensive Validation: The method is validated across in-silico (IXI, HCP) and in-vivo (MR-ART) datasets, including rigorous comparisons against supervised baselines (both "source-only" and "oracle") and state-of-the-art unsupervised methods (CycleGAN, AutoDPS, UDDN).

Experimental Results

Quantitative Performance: On in-silico datasets, SSRL-MAR achieved a PSNR of 23.81 dB, SSIM of 91.55%, and NMSE of 0.79%. On the in-vivo MR-ART dataset, the method reduced motion distortion significantly. After unsupervised domain adaptation, it improved PSNR by up to 2.0 dB over a "source-only" supervised model trained on the same simulated data.
Comparison to Supervised Models: While an "oracle" supervised model (fine-tuned on paired real data) achieved slightly higher scores (0.25–0.47 dB PSNR gain), such data is unavailable in practice. SSRL-MAR outperformed the practical "source-only" supervised baseline, demonstrating that self-supervised learning generalizes better to real motion than direct supervision on simulated data alone.
Anatomical Fidelity: Volumetric analysis using FreeSurfer showed that SSRL-MAR reduced volumetric errors in critical structures (e.g., corpus callosum, ventricular system) by more than 50% at mild motion levels. Cortical surface analysis confirmed the preservation of gyri-sulci continuity and topological correctness, with no evidence of hallucinated structures.
Robustness: The method demonstrated robust performance across varying motion severities (from mild to severe) and maintained 3D anatomical coherence across orthogonal views, outperforming 2D slice-based methods which introduced slice-wise artifacts.

Significance and Claims
The paper claims that SSRL-MAR provides a robust and scalable image-domain solution for 3D brain MRI motion correction. Its primary significance lies in enabling reliable structural quantification in large-scale neuroimaging studies without the prohibitive requirement of prospectively acquired paired data or scanner-specific calibration. By integrating contrastive representation learning with a learned degradation model, the framework achieves a "physics-inspired, analysis-by-synthesis" pathway that effectively corrects motion while preserving anatomical fidelity. The authors assert that this approach overcomes the generalization limits of supervised models trained on simulated data and the instability/hallucination risks of other unsupervised generative models, offering a practical tool for retrospective clinical and research neuroimaging pipelines.

More like this