Resolving Aliasing Artifacts from Undersampled Frequency Data

56 concepts59 papers1 proposal

In fields like Magnetic Resonance Imaging (MRI), acquisition speed is fundamentally limited by physics. To capture an image, scanners must sequentially sample data points in the frequency domain, known as k-space. Gathering every single data point required for a perfect reconstruction takes time—time that translates to patient discomfort, high costs, and motion blur in moving organs like the heart.

To accelerate this process, researchers use undersampling: they deliberately skip large fractions of k-space data. However, skipping data violates the Nyquist-Shannon sampling theorem. In the spatial domain, this violation does not manifest as random noise. Instead, it creates highly structured aliasing artifacts—ghostly, overlapping duplicates of anatomy that wrap around the image plane and obscure critical diagnostic details.

Resolving these artifacts requires sophisticated reconstruction paradigms that can untangle overlapping spatial structures using incomplete frequency measurements.


The Geometry of Aliasing

To understand why deep learning is so effective at de-aliasing, we must first understand the structured nature of the artifacts. When we uniformly drop lines in k-space, we increase the effective spacing between our frequency samples. According to the properties of the Fourier transform, sparsifying the frequency domain causes the spatial domain to compress and repeat.

As Ahmad et al. point out, conventional uniform undersampling patterns produce coherent, structured artifacts like "ghosting." Because these errors look exactly like actual anatomical features, traditional linear reconstruction algorithms cannot distinguish between a real organ boundary and an aliased duplicate.

Fully Sampled K-Space  ---> Smooth, Clean Spatial Image
Uniformly Undersampled ---> Repeated, Overlapping Spatial "Ghosts"

However, this structured nature is also a leverage point. Because the aliasing follows strict mathematical rules, the overlapping information isn't lost; it is just misplaced. Defazio demonstrates this by showing that shifting the starting point of equispaced sampling by a single pixel—an "offset-1" approach—exploits the conjugate symmetry of the Fourier transform. This simple shift alters how the aliasing overlaps, reducing spatial ambiguity and making it significantly easier for neural networks to separate the true anatomy from the artifacts.


Dual-Domain Untangling

If a neural network tries to fix an aliased image purely in the spatial domain, it faces a steep challenge: the overlapping structures have destroyed local pixel context. Conversely, working purely in the frequency domain makes it difficult to enforce geometric constraints like edge sharpness.

The modern solution is dual-domain reconstruction. Networks alternate between processing data in the spatial domain and the frequency domain.

  ┌────────────────────────────────────────────────────────┐
  │                                                        │
Input K-Space ──> Spatial Domain ──> Frequency Domain ──> Data Consistency
                      (CNN/ViT)          (FFT)            (Enforce Knowns)

In the spatial domain, convolutional layers or Transformers identify and smooth out the non-physical "ghost" repetitions. The image is then transformed via a Fast Fourier Transform (FFT) back into k-space. Here, a Data Consistency (DC) layer replaces the network's predicted frequency values with the actual raw data points measured by the scanner.

Zhou et al. implemented this in DuDoRNet, showing that recurrently alternating between these domains prevents the network from generating "hallucinations" or physically impossible structures. By locking down the known k-space values, the network's only job is to interpolate the missing frequencies that correspond to the overlapping spatial artifacts. Qin et al. extended this concept to the spatio-temporal frequency domain ($x-f$ space), proving that alternating domains allows models to exploit global redundancies far more effectively than single-domain networks.


Exploiting Spatio-Temporal and Cross-Modal Redundancy

When imaging a moving structure, such as a beating heart, the aliasing artifacts change over time based on the movement of the tissue. Rather than treating each video frame as an isolated de-aliasing problem, modern frameworks exploit temporal correlations.

Qin et al. introduced a convolutional recurrent neural network (CRNN-MRI) that tracks anatomical structures across time steps. Because the true anatomy moves continuously while the undersampling artifacts fluctuate rapidly across frames, the recurrent connections allow the network to "see through" the transient ghosting patterns by integrating clean structural information from adjacent frames.

Similarly, we can exploit redundancies across different imaging modalities. In a typical clinical session, a patient might undergo multiple scans, such as a fast T1-weighted scan and a slower, high-resolution T2-weighted scan. Zhou et al. leverage the fast-acquired T1 scan as a structural "cheat sheet." Since both scans share the same underlying anatomy, the network uses the sharp boundaries of the T1 image to guide the de-aliasing of the undersampled T2 data, resolving spatial ambiguities that would be impossible to untangle from the T2 data alone.


Learning to De-Alias Without Ground Truth

Historically, training these deep networks required pairs of undersampled (aliased) images and fully-sampled (clean) ground-truth images. However, collecting fully-sampled data for dynamic or high-resolution scans is often physically impossible due to patient movement or hardware constraints.

To bypass this limitation, researchers developed self-supervised learning frameworks that train directly on the aliased data itself.

Acquired Undersampled K-Space Data
       │
       ├───> Subset A (Input to Network) ───> Predicted Image ───> FFT ───> Predicted K-Space
       │                                                                           │
       └───> Subset B (Loss Evaluation) <──────────────────────────────────────────┘

The mechanism, popularized by works like those of Yaman et al., relies on splitting the available, incomplete k-space data into two disjoint subsets: Subset A and Subset B.

  1. The network takes only Subset A as an input and attempts to reconstruct a full image, which inherently contains some residual aliasing.
  2. This reconstructed image is converted back to k-space.
  3. The network's loss is calculated by comparing its predictions only against the held-out points in Subset B.

Because the artifacts caused by missing data in Subset A do not correlate with the missing data points in Subset B, the network cannot minimize the loss by memorizing or amplifying the aliasing patterns. Instead, it is forced to learn the underlying physical and structural continuity of the anatomy. Xu et al. demonstrated that this self-supervised feature learning matches the performance of models trained on rare fully-sampled datasets, making robust de-aliasing clinically viable for a wider range of sequences.


Co-Designing Sampling and Reconstruction

For decades, the patterns used to skip k-space data were fixed or random. However, different neural network architectures have different blind spots; a pattern that creates easy-to-fix artifacts for a CNN might create insurmountable ghosting for a Transformer.

This realization led to the co-design of sampling and reconstruction. Instead of treating the undersampling pattern as a fixed hardware constraint, frameworks like AutoSamp, developed by Alkan et al., use variational information maximization to optimize the k-space sampling mask jointly with the reconstruction network.

By parameterizing the sampling mask and making it differentiable, the reconstruction network can "tell" the sampling mechanism which frequency lines are absolutely critical for resolving spatial overlap. The system learns an optimal sampling strategy that deliberately distributes the resulting aliasing artifacts in a way that the companion neural network is uniquely specialized to untangle, pushing acceleration factors far beyond traditional physical limits.

Key Foundational Work

MRF denoising with compressed sensing and adaptive filtering (2014) — contributed 1 concept elements to this direction.
High-resolution, real-time exercise stress cine (2014) — contributed 1 concept elements to this direction.
Compressed Quantitative MRI: Bloch Response Recovery through Iterated Projection (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_undersampling_induced_aliasing_9b4c1959,
  title = {undersampling-induced aliasing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/undersampling-induced-aliasing/rdr_9b4c19592c634291b80467555fb2fdbc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_undersampling_induced_aliasing_9b4c1959_v1,
  title = {undersampling-induced aliasing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/undersampling-induced-aliasing/rdr_9b4c19592c634291b80467555fb2fdbc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake