Visualizing Fluid Flows via Regularized Optimal Mass Transport with Applications to Neuroscience

Xinan Chen, Anh Phong Tran, Rena Elkin, Helene Benveniste, Allen R. Tannenbaum

TL;DR

This paper introduces a significantly faster and more efficient method for visualizing fluid flows, particularly in the brain's glymphatic system, by enhancing the Optimal Mass Transport (OMT) framework with a diffusion term (rOMT). The improved algorithm achieves substantial runtime reductions (up to 99%) while maintaining accuracy, making it a powerful tool for analyzing complex biological fluid dynamics.

What Problem Does This Paper Address?

Understanding fluid flow dynamics within biological systems, especially the brain, is crucial for diagnosing and treating neurological diseases. The glymphatic system, responsible for waste clearance in the brain, involves complex fluid movements driven by both advection (bulk flow) and diffusion (spreading). Traditional methods for modeling and visualizing these flows often struggle with computational expense, especially for high-resolution 3D data. Existing Optimal Mass Transport (OMT) models, while powerful for fluid dynamics, typically focus on advection and can be computationally intensive. This paper addresses the need for a more efficient and accurate method to visualize these advection-diffusion dominated flows, particularly in the context of neuroscience applications.

Methods

The core of this work is the development of a regularized Optimal Mass Transport (rOMT) model. This model extends the original OMT formulation by incorporating a diffusion term into the continuity equation, creating an advection-diffusion equation. This allows for the simultaneous modeling of both bulk flow and diffusive spreading, which is essential for accurately representing biological fluid dynamics.

The authors also present a modified numerical method for solving the rOMT problem. This method focuses on optimizing computational efficiency by:

  • Pre-computing intermediate results: This avoids redundant calculations within nested functions, a common bottleneck in complex algorithms.
  • Combining nested functions: Streamlining the computational pipeline.
  • Implementing a free end-point condition: This makes the model more robust to noisy data, which is common in medical imaging like DCE-MRI.
  • Lagrangian Coordinates: For visualization, they employ a Lagrangian approach to track particle trajectories (pathlines) and derive speed-lines and Péclet-lines. The Péclet number, a ratio of advection to diffusion, is a key metric for understanding the dominant transport mechanism.

Datasets

The authors evaluate their methodology on two types of datasets:

  • Synthetic Geometric Dataset: Five 3D Gaussian spheres were created to simulate advection and diffusion. This dataset allows for controlled testing and comparison of different algorithm versions.
  • Real DCE-MRI Rat Brain Dataset: This dataset consists of 55 rat brains, with 12 successive images per rat, captured over a 110-minute period using Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI). This real-world data is crucial for demonstrating the method's applicability to biological systems.

Key Insights and Ideas

  • Diffusion is Crucial for Brain Fluid Dynamics: The inclusion of a diffusion term in the OMT model is not just an algorithmic enhancement but a biological necessity for accurately modeling brain fluid transport, where both advection and diffusion play significant roles.
  • Péclet Number as a Diagnostic Tool: The Péclet number provides a quantitative measure to distinguish between advection-dominated and diffusion-dominated regions, offering valuable insights into how substances move within the brain.
  • Significant Computational Speedup: The optimized rOMT algorithm achieves dramatic reductions in computation time, making it feasible to analyze large 3D datasets that were previously intractable.
  • Free End-Point Condition for Robustness: Using a free end-point condition in the optimization problem improves the model's ability to handle noisy input data, a common issue in medical imaging.
  • Lagrangian Visualization for Intuition: The use of pathlines, speed-lines, and Péclet-lines in a Lagrangian framework provides an intuitive and comprehensive way to visualize complex fluid dynamics.

Key Results

  • Runtime Reduction: The upgraded rOMT algorithm achieved a significant reduction in runtime compared to the previous version. For the Gaussian sphere dataset, this reduction was 91.25% ± 0.51%. For the DCE-MRI rat brain dataset, the reduction was 91.21% ± 1.21%, and further improved to 99.08% ± 1.21% when run in parallel.
  • Accuracy: The authors report that the upgraded algorithm maintained high accuracy, with no significant difference in Mean Squared Error (MSE) compared to the parallelized version and only minor differences compared to the previous version, especially for the original image sizes.
  • Visualization: The method successfully visualized fluid flows in both synthetic and real brain data, highlighting regions of advection-dominated and diffusion-dominated transport. For instance, in the rat brain data, higher speeds were observed along large vessels, while diffusion dominated deeper brain tissue.

Limitations and Future Work

  • Image Noise Impact: While the free end-point condition helps, image noise can still affect the results, as observed in the comparison of parallelized and unparallelized algorithms on pathlines (Fig. 6). Future work could focus on improved image pre-processing to mitigate data noise.
  • Constant Diffusion Coefficient: The current model uses a constant diffusion coefficient ($\sigma$). Realistically, diffusion in the brain is spatially and potentially time-dependent. Future research could explore non-linear or spatially dependent diffusion terms.
  • Multi-Marginal Models: For highly emphasized continuity and smoothness of velocity fields, multi-marginal models might be considered, though they would likely increase computation time.
  • Data Availability: While the code and some datasets are available, the full rat brain dataset is available upon reasonable requests.

Summary updated

Abstract

Regularized optimal mass transport (rOMT) problem adds a diffusion term to the continuity equation in the original dynamic formulation of the optimal mass transport (OMT) problem proposed by Benamou and Brenier. We show that the rOMT model serves as a powerful tool in computational fluid dynamics (CFD) for visualizing fluid flows in the glymphatic system. In the present work, we describe how to modify the previous numerical method for efficient implementation, resulting in a significant reduction in computational runtime. Numerical results applied to synthetic and real-data are provided.
BibTeX
@article{chen2023visualizing,
  author = {Xinan Chen and Anh Phong Tran and Rena Elkin and Helene Benveniste and Allen R. Tannenbaum},
  title = {{Visualizing Fluid Flows via Regularized Optimal Mass Transport with Applications to Neuroscience}},
  journal = {J. Sci. Comput},
  year = {2023},
  doi = {10.1007/s10915-023-02337-9},
}
Figure