Generalizing Reconstruction to Non-Cartesian Trajectories

56 concepts53 papers

In computational imaging modalities like Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), data is not captured as pixels. Instead, hardware sensors sample physical frequencies—known as $k$-space in MRI or sinograms in CT.

Historically, the easiest way to sample this data has been on a Cartesian grid: row by row, column by column. This grid structure allows reconstruction algorithms to leverage the Fast Fourier Transform (FFT), turning raw frequency data into a clean image almost instantaneously.

However, physical systems are constrained by biology and hardware. Patients move, hearts beat, and breathing deforms organs. To capture these dynamic processes without motion blur, scanning must be incredibly fast. This has led to the development of non-Cartesian trajectories—such as radial spokes, spirals, or rosette patterns. These trajectories cover the frequency domain far more efficiently and are inherently robust to motion.

The catch? Modern high-performance reconstruction models—especially those driven by deep learning—are fundamentally "grid-locked." They fail to generalize when shifted from standard Cartesian grids to complex, non-linear trajectories. Bridging this gap is one of the most active frontiers in computational imaging.


The Cartesian Assumption in Deep Learning

Over the past decade, deep learning has revolutionized image reconstruction. By training neural networks on undersampled data, researchers can reconstruct high-fidelity images from a fraction of the usual measurements. Yet, as Thomas Küstner et al. demonstrate in CINENet: deep learning-based 3D cardiac CINE MRI reconstruction with multi-coil complex-valued 4D spatio-temporal convolutions, these models exhibit a severe lack of generalization across different sampling patterns. A network optimized for a specific Cartesian mask typically requires complete retraining to function on non-Cartesian paths.

Why are these architectures so brittle? The problem lies in their core architectural assumptions:

  1. Grid-Based Convolutions: Standard Convolutional Neural Networks (CNNs) rely on square 2D or 3D kernels. These kernels assume that spatial neighborhoods are uniform. In a non-Cartesian trajectory, such as a spiral, the data points are arranged along continuous curves. Forcing this data onto a grid breaks the spatial relationships the network relies on.
  2. Linear Encoding Assumptions: As Zimeng Li et al. point out in Encoding Enhanced Complex CNN for Accurate and Highly Accelerated MRI, many physics-informed models use specialized 1D convolutional kernels designed to mimic linear frequency and phase encoding. When applied to non-linear data acquisition paths, these grid-aligned operations fail to maintain physical integrity.
  3. Structured vs. Diffuse Aliasing: When you skip lines on a Cartesian grid, the resulting image suffers from highly structured "ghosting" artifacts. Networks can easily learn to recognize and remove these predictable patterns. However, as Aaron Defazio and Tullie Murrell show in MRI Banding Removal via Adversarial Training., non-Cartesian trajectories produce entirely different, diffuse, or streak-like artifact signatures. A model trained to eliminate Cartesian ghosting is blind to radial streaks.

This trajectory dependency limits the clinical deployment of advanced models. Self-supervised frameworks like those proposed by Neel Dey et al. in DSFormer: A Dual-domain Self-supervised Transformer for Accelerated Multi-contrast MRI Reconstruction or the generative approaches in CDiffMR: Can We Replace the Gaussian Noise with K-Space Undersampling for Fast MRI? remain constrained by the physical geometry of the initial training data.


The Mathematical and Computational Bottleneck

If a model cannot process non-Cartesian data directly, the obvious workaround is to interpolate the data onto a Cartesian grid before feeding it to the network. This process is known as gridding.

Non-Cartesian Points (Spirals/Radial) 
          │
          ▼  (Requires Interpolation / Density Compensation)
   Gridding Operator
          │
          ▼
    Cartesian Grid ──► Standard FFT ──► Image Space

While conceptually simple, gridding introduces severe mathematical and computational complications:

1. The Cost of Data Consistency

Iterative reconstruction algorithms and unrolled neural networks must constantly move back and forth between the image domain and the frequency domain to ensure the reconstruction matches the raw sensor data. In Cartesian imaging, this loop is incredibly fast because the FFT is computationally cheap.

For non-Cartesian paths, every single iteration requires gridding and regridding. In An Augmented Lagrangian Based Compressed Sensing Reconstruction for Non-Cartesian Magnetic Resonance Imaging without Gridding and Regridding at Every Iteration, Seunghoon Nam et al. demonstrate that bypassing these repetitive interpolation steps can speed up reconstruction by 23-fold. However, their solution relies on diagonal approximations of the gridding operator, which degrade when the trajectory exhibits drastic variations in sampling density (such as in spiral scans).

2. Broken Mathematical Separability

On a Cartesian grid, multi-dimensional optimization problems can often be broken down into a series of independent, parallelizable 1D problems. As Çağdaş Bilen and Ivan Selesnick explain in High-Speed Compressed Sensing Reconstruction in Dynamic Parallel MRI Using Augmented Lagrangian and Parallel Processing, this separable, block-diagonal structure is what allows algorithms to run efficiently on GPUs. Non-Cartesian trajectories destroy this separability, drastically increasing the conditioning difficulty of the underlying optimization problem (Provable Preconditioned Plug-and-Play Approach for Compressed Sensing MRI Reconstruction).

3. The Calibration Problem

Advanced parallel imaging relies on combining data from multiple receiver coils, which requires knowing the precise sensitivity map of each coil. Cartesian scans usually reserve a small, fully sampled region in the center of $k$-space to calibrate these maps. Non-Cartesian trajectories, by design, rarely sample data uniformly. As Sebastian Rosenzweig et al. note in Simultaneous multi‐slice MRI using cartesian and radial FLASH and regularized nonlinear inversion: SMS‐NLINV, this creates a fundamental conflict between the desire for motion-robust trajectories and the data requirements of traditional calibration algorithms.


Strategies for Trajectory Generalization

To build models that work universally across clinical protocols, researchers are exploring several avenues to decouple architectures from fixed Cartesian grids.

Trajectory-Specific Physics Embeddings

Instead of forcing data onto a grid, some approaches embrace the specific geometry of the trajectory. In Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation, Qing Wu et al. utilize Implicit Neural Representations (INRs) paired with the Fourier-slice theorem to reconstruct radial data without external training sets. While highly effective, this strategy creates a different kind of specificity: the mathematical formulation is locked to radial lines and cannot easily ingest spiral or Cartesian formats.

Non-Grid Architectures

Other researchers are replacing CNNs with architectures that do not assume a rigid pixel grid. For example, Muhammed Fikret Yalcinbas et al. explore the use of Vision Transformers (ViTs) in Rosette Trajectory MRI Reconstruction with Vision Transformers to map irregular rosette sampling patterns. Because self-attention mechanisms operate on sets of tokens rather than local pixel neighborhoods, they offer a more flexible mathematical framework for handling arbitrary trajectories.

Joint Learning of Trajectory and Reconstruction

Rather than treating the trajectory as an unchangeable constraint imposed by the hardware, a growing body of work seeks to optimize the sampling path and the reconstruction network simultaneously.

┌─────────────────────────┐      ┌─────────────────────────┐
│  Learned Trajectory     │ ───► │  Reconstruction Model   │
│  (Differentiable Mask)  │ ◄─── │  (Loss Gradient)        │
└─────────────────────────┘      └─────────────────────────┘

By making the acquisition pattern a differentiable component of the training pipeline—as explored in Joint learning of cartesian undersampling and reconstruction for accelerated MRI and by Thomas Sanchez et al. in Scalable Learning-Based Sampling Optimization for Compressive Dynamic MRI—the system can learn trajectories that are mathematically optimized for the reconstruction network's specific capabilities. While currently restricted mostly to constrained Cartesian line-skipping or simplified 2D models (Constrained Probabilistic Mask Learning for Task-specific Undersampled MRI Reconstruction), extending this joint optimization to fully non-Cartesian space represents the ultimate goal for hardware-software co-design in medical imaging.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sampling_trajectory_generalization_b7e35dc8,
  title = {sampling trajectory generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sampling-trajectory-generalization/rdr_b7e35dc8b2e443a5b7cd3293385a865b},
  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_sampling_trajectory_generalization_b7e35dc8_v1,
  title = {sampling trajectory generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sampling-trajectory-generalization/rdr_b7e35dc8b2e443a5b7cd3293385a865b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake