Sub-Riemannian Lifting for Curvature Aware Image Analysis

This direction is merged into another direction: orientation score lifting. Its last version stays readable and citable here.

Version 1 of 110 papers19 conceptsfrozen 2026-10-04

When tracking a blood vessel in a retinal scan or a neural fiber in a brain MRI, standard algorithms often take "shortcuts." In a 2D or 3D image, if two vessels cross, a simple shortest-path algorithm might jump from one vessel to the other at the intersection. This happens because the algorithm only sees spatial proximity, ignoring the fact that biological structures have an inherent "flow" or directionality.

To solve this, researchers have developed a framework that "lifts" images into higher-dimensional spaces where orientation is treated as a coordinate just as important as position.

Lifting Beyond Spatial Coordinates

The core intuition behind this research direction is that a curve in an image is more than just a sequence of points $(x, y)$; it is a sequence of points and their corresponding tangents. Remco Duits, Alexey Mashtakov, and Gonzalo Sanguinetti describe this in A PDE Approach to Data-driven Sub-Riemannian Geodesics in SE(2) as moving from the 2D plane to the 3D group of roto-translations, $SE(2)$.

In this lifted space, two crossing lines are no longer touching. Because they have different orientations at the crossing point, they are separated along the "orientation axis." By searching for paths in this higher-dimensional space, algorithms can follow a single structure through a crossing without getting "leaked" into the intersecting one.

The Sub-Riemannian Constraint

Simply adding a dimension isn't enough; the movement in that space must be restricted. This is where sub-Riemannian geometry comes in.

Imagine driving a car. You can move forward, and you can turn the steering wheel to change your orientation, but you cannot move sideways. This "non-holonomic" constraint means your velocity must always align with your current heading. In the context of image analysis, this constraint forces the extracted path to be smooth.

As explored by Antonin Chambolle and Thomas Pock in Total roto-translational variation, imposing a metric where spatial movement is coupled with local orientation naturally prevents sharp corners. If an algorithm wants to change direction, it must "pay" for the rotation, effectively penalizing curvature and encouraging geometric continuity.

Avoiding Non-Physical Singularities

A recurring challenge in modeling these paths is the appearance of "cusps"—points where a mathematical optimal path suddenly reverses direction to minimize distance. While these might be valid solutions for a robot that can shift into reverse, they are biologically impossible for structures like blood vessels or white matter tracts.

In On sub-Riemannian geodesics in $SE(3)$ whose spatial projections do not have cusps, Remco Duits and colleagues identify the exact conditions needed to ensure 3D curves remain smooth and "cuspless."

Building on this, Remco Duits and Jorg M. Portegies suggest in Optimal Paths for Variants of the 2D and 3D Reeds-Shepp Car that we should treat these structures like "forward-only" vehicles. By restricting the model to forward motion, sharp cusps are replaced by stationary rotation points at bifurcations. This more accurately reflects how biological vessels branch and grow.

Measuring Connectivity in 5D

This approach reaches its most complex form in brain imaging. In Brain Connectivity Measures via Direct Sub-Finslerian Front Propagation, Jorg M. Portegies and Stephan Meesters lift 3D brain data into a 5D space of positions and directions.

By propagating "fronts" through this 5D bundle using sub-Finslerian geometry, the model incorporates the "stiffness" of neural fibers directly into the distance metric. The resulting pathways are not just the shortest lines between two points, but the most biologically plausible routes that respect the physical constraints of the brain's architecture.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sub_riemannian_lifting_for_curvature_awa_5a586509,
  title = {Sub-Riemannian Lifting for Curvature Aware Image Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sub-riemannian-lifting-for-curvature-aware-image-analysis/rdr_5a586509bc69420494a076c97f8c0f32},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sub_riemannian_lifting_for_curvature_awa_5a586509_v1,
  title = {Sub-Riemannian Lifting for Curvature Aware Image Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sub-riemannian-lifting-for-curvature-aware-image-analysis/rdr_5a586509bc69420494a076c97f8c0f32/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}