Abstract
We study a data-driven sub-Riemannian (SR) curve optimization model for connecting local orientations in orientation lifts of images. Our model lives on the projective line bundle $\mathbb{R}^{2} \times P^{1}$, with $P^{1}=S^{1}/_{\sim}$ with identification of antipodal points. It extends previous cortical models for contour perception on $\mathbb{R}^{2} \times P^{1}$ to the data-driven case. We provide a complete numerical analysis of the dynamics of the 1st Maxwell-set with growing radii of SR-spheres, revealing the cut-locus. Furthermore, a comparison of the cusp-surface in $\mathbb{R}^{2} \times P^{1}$ to its counterpart in $\mathbb{R}^{2} \times S^{1}$ reveals a general and strong reduction of cusps in spatial projections of geodesics. Experiments show that the projective line bundle structure greatly reduces the presence of cusps and computational time.
Introduction: Orientations vs. Directions
Extracting thin, curvilinear structures like blood vessels from medical imagery is a significant challenge due to noise, low contrast, and complex topological features like crossings. A standard approach involves "lifting" 2D images into a 3D space that incorporates local angular information. Traditionally, this has been modeled on the space of positions and directions, $\mathbb{R}^2 \times S^1$.
However, blood vessels typically lack an inherent "forward" or "backward" direction; they possess an orientation. Forcing a directional model often introduces "cusps"—unnatural, sharp 180-degree turns in the tracked path where the model attempts to maintain mathematical optimality at the expense of physical realism. This paper proposes shifting the framework to the projective line bundle $\mathbb{R}^2 \times P^1$, where an angle $\theta$ is identified with $\theta + \pi$.
Methodology: Data-Driven Sub-Riemannian Geometry
The authors implement a data-driven sub-Riemannian (SR) model. The process begins by transforming the 2D image into a 3D orientation score, highlighting vessel-like structures at specific angles. A sub-Riemannian metric is then defined on $\mathbb{R}^2 \times P^1$, where "distance" is a path that respects local orientation and penalizes excessive bending.
To find optimal paths (geodesics), the model solves an Eikonal partial differential equation (PDE) using wavefront propagation methods, such as anisotropic fast-marching. By identifying antipodal points, the search space is effectively halved, leading to significant computational gains.
Maxwell Set Analysis and the Cut-Locus
A core theoretical contribution of this work is the analysis of the "1st Maxwell set"—the locus of points where two distinct shortest paths of equal length intersect. This analysis is vital for identifying the "cut-locus," the point beyond which a geodesic ceases to be the unique global minimizer.
The study provides a detailed numerical investigation into how these Maxwell sets evolve as the search radius $R$ increases. In the projective space $\mathbb{R}^2 \times P^1$, the topology of the reachable set transitions from a simple sphere to a more complex torus-like structure, reflecting the underlying symmetries of the orientation bundle.
Key Results and Improvements
The $\mathbb{R}^2 \times P^1$ model was evaluated on 235 retinal vessel segments and compared against the standard $S^1$ directional model. Key findings include:
- Computational Efficiency: The new model is approximately 4 times faster than the $\mathbb{R}^2 \times S^1$ model. In practical tests, computation time for a standard image patch was reduced from 14.4 seconds to 3.4 seconds.
- Cusp Reduction: By merging antipodal directions, the model eliminates the "reverse gear" artifact. This allows geodesics to pass through orientation changes smoothly, significantly reducing non-physical cusps in the spatial projection.
- Tracking Accuracy: Qualitative results indicate that the geodesics produced by this model follow vessel centerlines more accurately, particularly in low-contrast regions and at vessel crossings.
Conclusion
By adopting the projective line bundle $\mathbb{R}^2 \times P^1$, this research demonstrates that modeling orientations rather than directions provides a more physically consistent and computationally efficient framework for vessel tracking. Future work aims to combine this model with explicit "cusp-free" constraints to further refine the tracking of complex vascular networks.
Summary updated