An Evolutionary Framework for Microstructure-Sensitive Generalized Diffusion Gradient Waveforms

Raphaël Truffet, Jonathan Rafael-Patino, Gabriel Girard, Marco Pizzolato, Christian Barillot, Jean-Philippe Thiran, Emmanuel CaruyerMICCAI 2020 · DOI

TL;DR

This paper introduces a novel framework for designing advanced MRI gradient waveforms that are specifically optimized to better detect microstructural features in the brain, like the volume of axons. By using a genetic algorithm and Monte Carlo simulations, they demonstrate that these custom waveforms can outperform traditional methods in their sensitivity to these crucial biological details.

What Problem Does This Paper Address?

Diffusion-weighted Magnetic Resonance Imaging (DW-MRI) is a powerful tool for non-invasively probing the microscopic structure of biological tissues, particularly the brain's white matter. The white matter is characterized by highly organized structures, primarily axons, which are essentially the "wires" of the brain. DW-MRI's ability to reveal information about these structures, such as the density and organization of axons, is crucial for understanding brain development, aging, and diseases like Alzheimer's or multiple sclerosis.

However, the quality of the information extracted from DW-MRI is heavily dependent on the design of the magnetic field gradients used during the scan. Traditional DW-MRI often employs simple, pulsed gradient waveforms. While effective, these may not be optimal for capturing subtle microstructural details. Designing "generalized" gradient waveforms that are specifically tailored to be sensitive to particular microstructural parameters (like the intra-axonal volume fraction, or IAVF) is a complex challenge. This paper aims to address this gap by proposing a systematic and automated framework for designing such optimized, microstructure-sensitive gradient waveforms.

Methods

The core of this work lies in a novel framework that combines Monte Carlo (MC) simulations with a genetic algorithm (GA).

  • Monte Carlo Simulations: Instead of relying on simplified analytical models, the authors use MC simulations to model the diffusion of water molecules within a realistic, three-dimensional representation of brain tissue. This approach offers greater flexibility and accuracy in capturing complex microstructural environments. The signal attenuation in DW-MRI is then computed based on these simulations.
  • Fisher Information: To quantify how sensitive a given gradient waveform is to a specific microstructural parameter (in this case, the IAVF), the researchers employ the Fisher Information (FI). The FI is a statistical measure that indicates how much information a random variable (the MRI signal) carries about an unknown parameter. A higher FI means the signal is more sensitive to changes in that parameter, leading to more precise estimations.
  • Genetic Algorithm: To search for the optimal gradient waveforms, a GA is used. This is an evolutionary computation technique inspired by natural selection.
  • Initialization: The process starts with a population of randomly generated gradient waveforms. These waveforms are constrained by realistic physical limits of MRI scanners (maximum gradient strength and slew rate).
  • Scoring: Each waveform is evaluated by computing its Fisher Information score, averaged over different orientations to ensure rotation invariance.
  • Evolution: Waveforms with higher scores are more likely to be selected for "reproduction." This involves cross-over (combining parts of two parent waveforms) and mutation (small random changes) to create a new generation of waveforms.
  • Iteration: This process is repeated for multiple generations, with the population of waveforms gradually evolving towards those that yield higher Fisher Information scores for IAVF estimation.

Datasets

The paper does not use pre-existing datasets in the traditional sense. Instead, they generate their own numerical substrates for simulation:

  • Numerical Substrates: These are 3D models of brain white matter. They consist of two compartments: intra-axonal space (modeled as parallel cylinders with varying radii) and extra-axonal space. These substrates are designed to represent different intra-axonal volume fractions (IAVF).
  • Benchmark: They compare their optimized waveforms against the Pulsed Gradient Spin Echo (PGSE) sequence, a standard in DW-MRI. They generate a large family of feasible PGSE sequences with varying timing and amplitude parameters to establish a baseline score for comparison.

Key Insights and Ideas

  • Beyond PGSE: The study strongly suggests that generalized gradient waveforms, designed with specific microstructural targets in mind, hold significant potential beyond the capabilities of conventional PGSE sequences.
  • IAVF as a Target: Focusing on the intra-axonal volume fraction (IAVF) is a clinically relevant choice, as it's a sensitive biomarker for neurodegenerative diseases.
  • Fisher Information as a Metric: Using Fisher Information provides a robust, statistically grounded way to quantify the sensitivity of a waveform to a specific microstructural parameter.
  • MC + GA Synergy: The combination of detailed Monte Carlo simulations for signal modeling and a genetic algorithm for waveform optimization is a powerful approach for tackling complex acquisition design problems.
  • Waveform vs. Sampling Direction: The framework separates the optimization of the gradient waveform itself from the optimization of sampling directions, allowing for a more focused search.

Key Results

  • Improved Sensitivity: The genetically optimized waveforms consistently achieved higher Fisher Information scores compared to a large set of feasible PGSE sequences, indicating greater sensitivity to the IAVF.
  • Evolutionary Improvement: The genetic algorithm demonstrated a clear progression, with the average Fisher Information score of the waveform population increasing significantly over generations.
  • Waveform Diversity: Even within the optimized generation, there was diversity in the waveforms that achieved high scores, suggesting multiple optimal solutions might exist.
  • Computational Efficiency: By pre-generating and storing Monte Carlo particle dynamics, they achieved a significant speedup in the simulation process, reducing generation time from 20 hours to about 30 minutes.

Limitations and Future Work

  • In Vivo Validation: The current study is based on simulations. The authors acknowledge that in vivo experiments are a necessary next step to validate the efficiency of these generated waveforms in actual human subjects.
  • Ground Truth Challenge: Designing experiments to definitively prove the efficiency of these new waveforms without having the ground truth microstructure is noted as a challenging issue.
  • Other Microstructure Parameters: While the framework is presented for IAVF, it can be generalized to optimize for other microstructural parameters. Future work is planned to target parameters like angular dispersion and spherical pore size estimation.
  • Substrate Complexity: While the current work uses a two-compartment model, the framework's adaptability to various compartment shapes and sizes suggests it can be extended to more complex microstructural models.

Summary updated

BibTeX
@inproceedings{truffet2020evolutionary,
  author = {Raphaël Truffet and Jonathan Rafael-Patino and Gabriel Girard and Marco Pizzolato and Christian Barillot and Jean-Philippe Thiran and Emmanuel Caruyer},
  title = {{An Evolutionary Framework for Microstructure-Sensitive Generalized Diffusion Gradient Waveforms}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  doi = {10.1007/978-3-030-59713-9\_10},
}
Figure