Hardware-Feasible Learned Sampling Trajectories

This direction is split into other directions: hardware-constrained sampling. Its last version stays readable and citable here.

Version 1 of 148 papers64 conceptsfrozen 2026-10-04

In the mathematical theory of Compressed Sensing (CS), the recipe for fast imaging is simple: sample a small, random subset of data points and use optimization to fill in the blanks. However, in the physical world of Magnetic Resonance Imaging (MRI), hardware cannot "teleport" between random points in frequency space. To collect data, the scanner’s magnetic gradients must move along continuous, smooth paths.

This creates a fundamental tension: the sampling patterns that are mathematically optimal for reconstruction are often physically impossible for the hardware to execute. A new research direction has emerged to bridge this gap by jointly learning acquisition trajectories and reconstruction algorithms while strictly enforcing the physical laws of the imaging hardware.

The Problem of Physical Constraints

Standard CS theory often assumes independent and identically distributed (i.i.d.) sampling. As Nicolas Chauffert, Philippe Ciuciu, and Pierre Weiss noted in From variable density sampling to continuous sampling using Markov chains, physical hardware is constrained by gradient limits. In MRI, these are defined by two parameters:

  1. Maximum Gradient Strength ($G_{max}$): The maximum "speed" at which the scanner can move through frequency space (k-space).
  2. Slew Rate ($S_{max}$): The maximum "acceleration" or rate of change of the gradient.

If a learned sampling pattern ignores these, it may require "jumps" that cause mechanical vibrations, image artifacts, or even hardware damage. Early attempts to solve this, such as those by Oliver M. Jeromin and Marios S. Pattichis, focused on deterministic geometries like spirals, which are naturally smooth but lack the flexibility of data-driven optimization.

Joint Learning of Acquisition and Reconstruction

The modern approach, pioneered by researchers like Alexander Bronstein and Tomer Weiss, treats the sampling trajectory as a learnable parameter within a deep learning pipeline.

In frameworks like PILOT (Physics-Informed Learned Optimal Trajectories for Accelerated MRI), the system doesn't just learn how to reconstruct an image from a fixed mask. Instead, it learns the optimal path to take through k-space and the best way to turn that specific path's data into an image. By making the sampling process differentiable, the entire acquisition-to-reconstruction chain can be optimized end-to-end.

Embedding Physics into the Optimization Loop

To ensure these learned paths are actually executable, researchers embed hardware constraints directly into the neural network's loss function or architecture. There are three primary ways this is achieved:

1. Differentiable Penalty Terms

In the PILOT framework, Tomer Weiss and colleagues use hinge functions in the loss function to penalize any trajectory segment that exceeds the scanner's maximum gradient or slew rate. This acts as a form of regularization, forcing the network to "discover" smooth, efficient paths that stay within the hardware's operational boundaries.

2. Projection onto Feasible Sets

In 3D FLAT (Feasible Learned Acquisition Trajectories for Accelerated MRI) and TEAM-PILOT (TEAM PILOT - Learned Feasible Extendable Set of Dynamic MRI Acquisition Trajectories), the authors use projection algorithms. After the network suggests a trajectory update, the path is mathematically projected back onto the "feasible set" of trajectories that satisfy physical constraints. This ensures that the final output is guaranteed to be safe for the scanner.

3. Neural Ordinary Differential Equations (Neural ODEs)

Wei Peng and Fang Liu proposed using Neural ODEs to model the sampling trajectory in Learning Optimal K-space Acquisition and Reconstruction using Physics-Informed Neural Networks. Because ODEs naturally describe continuous motion, they provide a mathematically elegant way to enforce smoothness and physical consistency throughout the learning process.

Beyond Medical Imaging: Path-Based Optimization

While MRI is the most prominent application, the concept of "hardware-feasible" optimization is spreading to other fields where experiments have physical "movement" costs.

In chemical engineering, Jose Pablo Folch and Behrang Shafei have developed Path-based Bayesian Optimization (Practical Path-based Bayesian Optimization). In a lab, you cannot instantaneously change a reactor's temperature from 20°C to 200°C; you must follow a path. By treating optimization as a sequential decision process (often using Markov Decision Processes), these algorithms plan multi-step experimental trajectories that respect physical limits like ramp times and cleaning requirements.

Similarly, in quantum computing, Andy J. Goldschmidt and Steven L. Brunton apply Model Predictive Control to prepare quantum states while respecting pulse amplitude and slew rate limits (Model predictive control for robust quantum state preparation).

Conclusion

The shift from "mathematically optimal" to "physically feasible" represents a maturing of the field. By integrating hardware constraints—whether they are MRI gradient limits, robot motor torques (Elisa Alboni), or chemical ramp rates—into the core of the machine learning model, researchers are creating algorithms that don't just work in simulations, but are ready for immediate deployment on real-world physical systems.

Papers in this version

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_hardware_feasible_learned_sampling_traje_1e65772d,
  title = {Hardware-Feasible Learned Sampling Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hardware-feasible-learned-sampling-trajectories/rdr_1e65772d342a4378b34028f0c0e621fa},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_hardware_feasible_learned_sampling_traje_1e65772d_v1,
  title = {Hardware-Feasible Learned Sampling Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hardware-feasible-learned-sampling-trajectories/rdr_1e65772d342a4378b34028f0c0e621fa/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}