Magnetic Resonance Imaging (MRI) provides unparalleled soft-tissue contrast, but it suffers from a fundamental vulnerability: acquisition latency. Unlike a digital camera that takes a snapshot in milliseconds, an MRI scanner must sequentially sample raw frequency data—known as k-space—over several minutes.
If a patient moves, breathes, or has a beating heart during this window, the mapping between spatial coordinates and frequency components is broken. The result is motion degradation: a mix of blurring and periodic "ghost" duplicates that can render a scan clinically unusable. To resolve this, modern medical imaging leverages deep learning to either outrun motion through accelerated acquisition or untangle it after the fact using physics-informed neural networks.
Why Motion Corrupts K-Space
To understand why motion artifacts are so destructive, we must look at how MRI constructs an image. The scanner does not measure pixels directly; it measures the spatial Fourier transform of the object. Each line of k-space contains global frequency information about the entire image.
If a patient moves while a specific line of k-space is being acquired, the phase information for that entire line is shifted. When the inverse Fourier transform is applied to reconstruct the final image, this localized error in the frequency domain distributes globally across the spatial domain. This is why respiratory or involuntary movement does not just blur a single moving edge—it creates ghost artifacts that repeat across the entire field of view, obscuring static anatomical structures.
Accelerating Acquisition to Outrun Motion
The most direct way to prevent motion artifacts is to reduce the time a patient spends inside the scanner. By undersampling k-space—collecting only a fraction of the required frequency lines—scans can be accelerated by an order of magnitude, finishing before significant motion occurs. However, naive undersampling introduces severe aliasing artifacts.
Deep learning models solve this by learning to reconstruct clean images from sparse data. For instance, Anuroop Sriram et al. developed end-to-end variational networks (End-to-End Variational Networks for Accelerated MRI Reconstruction) that directly model the physics of the scanner hardware, allowing for high-fidelity reconstructions from highly accelerated scans.
Rather than treating acceleration and sampling as separate problems, Tomer Weiss et al. proposed a joint learning framework (Joint learning of cartesian undersampling and reconstruction for accelerated MRI) that simultaneously optimizes which k-space lines to sample and how to reconstruct them. By capturing only the most informative frequency data, these methods minimize the scanner's window of vulnerability to patient movement.
Disentangling Content from Artifacts
When motion is completely unavoidable—such as in fetal, pediatric, or cardiac imaging—post-processing methods must isolate and remove the artifacts from the completed scan. A major challenge here is the lack of "paired data": it is practically impossible to simultaneously acquire a motion-corrupted and a perfectly still scan of the same patient to use as a training target.
One solution is to simulate the physics of motion to generate synthetic training data. Kamlesh Pawar et al. demonstrated this approach (Suppressing motion artefacts in MRI using an Inception‐ResNet network with motion simulation augmentation), training a network on simulated motion-corrupted data to successfully remove real-world artifacts without altering standard scanner workflows.
To bypass simulation entirely, Siyuan Liu et al. introduced an unsupervised framework called DUNCAN (Learning MRI artefact removal with unpaired data). By utilizing cycle-consistent adversarial networks, the authors disentangled the underlying anatomical content from the artifact noise. This allows the network to strip away motion distortions while preserving genuine clinical structures, even when trained purely on independent sets of clean and corrupted scans.
Physics-Informed Optimization and Test-Time Adaptation
While pure image-to-image translation networks can produce visually appealing results, they risk hallucinating anatomical details because they lack a physical anchor to the actual scan. To guarantee clinical fidelity, recent paradigms integrate the forward physics of MRI directly into the neural optimization loop via data consistency.
Instead of relying on large, pre-trained datasets that may not generalize to unusual pathologies, Tobit Klug et al. introduced MotionTTT (MotionTTT: 2D Test-Time-Training Motion Estimation for 3D Motion Corrected MRI). This method estimates 3D rigid motion parameters directly at test-time by optimizing the parameters to minimize a data consistency loss through a frozen, motion-free network.
In a similar vein, Qing Wu et al. proposed Moner (Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation), which leverages Implicit Neural Representations (INRs) to model continuous motion fields in undersampled radial MRI without any external training data.
[Raw K-Space Data] ──> [Motion Estimation Model] ──> [Estimated Motion Parameters]
│ │
└──────────────────> [Data Consistency Loss] <──────────┘
│
(Updates Reconstruction)
▼
[Artifact-Free Image]
The core intuition behind these physics-informed approaches is simple: if you take the reconstructed, motion-corrected image and mathematically simulate the MRI scanning process, the resulting synthetic k-space data must match the actual raw data collected by the scanner. By enforcing this constraint during inference, the network is prevented from hallucinating features, ensuring that the final, artifact-free image remains entirely faithful to the patient's true anatomy.