TL;DR
This paper introduces a powerful PDE-based framework for finding the "shortest" paths for a car-like robot (the Reeds-Shepp car) in both 2D and 3D spaces. By lifting image data into a higher-dimensional position-plus-orientation space and solving anisotropic Eikonal equations, the authors provide a robust way to track complex tubular structures—like blood vessels and brain fibers—even when they cross, branch, or disappear in low-contrast regions.
Introduction: The Challenge of Tubular Structure Tracking
In medical image analysis, tracking thin, tube-like structures such as blood vessels, nerves, or white matter fibers is a major challenge. Traditional methods often suffer from the "leakage problem," where a tracking algorithm gets confused at a crossing and jumps onto the wrong path.
To solve this, researchers use models like the Reeds-Shepp car, which doesn't just care about the shortest distance between two points, but also the "cost" of turning (curvature). However, standard Reeds-Shepp models allow for a "reverse gear," leading to sharp "cusps" in the path that don't make sense for biological structures like blood vessels. This paper addresses the mathematical and numerical hurdles of extending these models to 3D and creating a "no-reverse-gear" variant that better mimics biological growth and branching.
Methods: Lifting and Eikonal PDEs
The authors treat pathfinding as a curve optimization problem on a manifold ($\mathbb{R}^d \times \mathbb{S}^{d-1}$), which represents every possible position and orientation.
- Lifting to Higher Dimensions: By adding orientation as a coordinate, the "leakage" at crossings is mathematically impossible because two crossing vessels have different orientations at the same spatial point. In the lifted space, these paths are distinct and do not intersect.
- Finsler Metrics: They define highly anisotropic Finsler metrics to approximate the singular distances of the Reeds-Shepp model. This allows them to penalize curvature and enforce "forward-only" motion.
- Fast-Marching (FM) Algorithm: To solve the Hamilton-Jacobi-Bellman (Eikonal) equation efficiently, they use an advanced FM method. The "secret sauce" is the use of adaptive discretization stencils that align with the preferred directions of the metric, ensuring the algorithm remains stable and accurate despite extreme anisotropy.
Key Insights: Cusps vs. Keypoints
In the standard Reeds-Shepp model (with reverse gear), a car might perform a three-point turn to reach a target, creating a sharp "cusp." In the new "no-reverse" model, these are replaced by keypoints—locations where the path includes an "in-place rotation." This is a much more natural way to model bifurcations in vessels, where a branch might split off at a sharp angle.
Results and Applications
The paper evaluates the method on several types of data, demonstrating high numerical accuracy and robustness:
- 2D Retinal Images: The "no-reverse" model proved superior for vessel tracking. While the standard model might create a non-physical cusp to reach a branch, the new model correctly identifies the branching point as a keypoint for rotation.
- 3D Synthetic dMRI: In simulated brain data, the sub-Riemannian approach successfully tracked fibers through crossings and areas of high torsion where standard isotropic methods failed by taking "shortcuts" through empty space.
Conclusion
This PDE-based framework provides a robust, mathematically grounded solution for tracking complex structures in medical images. By incorporating orientation and removing the reverse gear, the model mimics biological patterns more closely. Future work will focus on scaling to real, noisy 3D dMRI datasets and exploring locally adaptive metrics to further improve tracking in highly tortuous regions.
Summary updated