Diffeomorphic Flow for Topology Preserving Mesh Reconstruction

This direction is merged into another direction: topology-preserving deformation. Its last version stays readable and citable here.

Version 1 of 115 papers21 conceptsfrozen 2026-10-04

In 3D computer vision and medical imaging, representing complex shapes like the human cortex or abdominal organs is a balancing act. You want a model that is flexible enough to capture intricate folds but rigid enough to remain "physically plausible."

Traditional deep learning approaches often predict direct vertex displacements—essentially telling each point on a template mesh where to jump. However, if a vertex jumps too far, it can pass through another part of the surface, creating "self-intersections" or "tangled" geometry. This breaks the topology of the object, making the resulting model useless for simulations or downstream anatomical analysis.

To solve this, researchers have converged on a shared mechanism: Diffeomorphic Flow. Instead of predicting where a point should be, these models predict a velocity field that describes how a point should move over time.

From Displacement to Continuous Flow

The core intuition behind this research direction is to treat mesh deformation as a fluid-like process rather than a single static jump. In CorticalFlow, Léo Lebrat and colleagues model the deformation of the cerebral cortex by integrating a predicted velocity field via an Ordinary Differential Equation (ODE).

Mathematically, if $x$ is a point on a template mesh, its position at the end of the deformation is found by solving: $$\frac{d\phi_t}{dt} = v(\phi_t, t)$$ where $v$ is the velocity field. By integrating this field from $t=0$ to $t=1$, the template "flows" into the target shape. Because the transformation is continuous and smooth, it is a diffeomorphism—a mapping that is both differentiable and invertible. This ensures that if you start with a sphere, you end with something that is still topologically a sphere, no matter how much it is stretched or folded.

Why Flow Prevents Self-Intersection

The "magic" of the flow-based approach lies in its ability to prevent self-intersections. In a standard displacement model, two distant vertices might be assigned destination coordinates that happen to cross each other.

In a diffeomorphic flow, as explored by Fabian Bongratz and Christian Wachinger in their work on UNetFlow, the mesh moves through a continuous velocity field. Because the field is smooth, two points that start near each other will stay near each other, and points can never "pass through" one another. This is critical for segmenting thin, complex structures like the pancreas or the folded ribbons of the brain, where standard voxel-based methods often produce fragmented or "leaky" results.

Maintaining Dense Correspondence

A significant byproduct of this approach is point-to-point correspondence. Because every target shape is just a warped version of the same starting template, vertex #500 on a reconstructed liver is mathematically the same "spot" as vertex #500 on a reconstructed kidney or another patient's liver.

Shanlin Sun and Kun Yeun Han leverage this in Hybrid Neural Diffeomorphic Flow (HNDF). By encoding these deformations into "triplane" features—axis-aligned 2D planes—they can generate new 3D shapes that are guaranteed to share the same underlying structure. This makes it possible to perform statistical shape analysis or transfer textures across different objects effortlessly.

Scaling and Efficiency

While integrating ODEs can be computationally expensive, recent work has focused on making these flows efficient for real-time applications:

  1. Stationary Velocity Fields: Instead of a field that changes over time, UNetFlow often uses a stationary field, which simplifies the integration process while maintaining the diffeomorphic property.
  2. Cage-Based Deformation: In a different take on the problem, Keke Tang and Ligang Liu use cage-based structures. Instead of moving every point in a dense cloud, they move the vertices of a surrounding "cage." Using barycentric coordinates, the movement of the cage smoothly propagates to the interior points, ensuring the object's integrity is maintained even during adversarial attacks.

By shifting the focus from "where the points go" to "how the space flows," this research direction provides a robust framework for 3D reconstruction that respects the fundamental laws of geometry and anatomy.

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_diffeomorphic_flow_for_topology_preservi_bcbc7087,
  title = {Diffeomorphic Flow for Topology Preserving Mesh Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diffeomorphic-flow-for-topology-preserving-mesh-reconstruction/rdr_bcbc7087631d450a91b2c5e3d5b0461c},
  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_diffeomorphic_flow_for_topology_preservi_bcbc7087_v1,
  title = {Diffeomorphic Flow for Topology Preserving Mesh Reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/diffeomorphic-flow-for-topology-preserving-mesh-reconstruction/rdr_bcbc7087631d450a91b2c5e3d5b0461c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}