Matching 3D Shapes Using 2D Conformal Representations

Xianfeng Gu, Baba C. Vemuri

Introduction

Matching 3D shapes—establishing a point-to-point correspondence between different objects—is a critical task in medical imaging. Applications range from longitudinal studies of tumor growth to cross-sectional comparisons of brain structures in neurodegenerative diseases like Alzheimer’s. However, non-rigid 3D matching is computationally demanding due to the high dimensionality of the search space. A significant challenge is ensuring that the resulting transformation is a diffeomorphism: a smooth, invertible mapping that preserves the topology of the surface without tearing or folding.

Methodology: A Conformal Approach

The proposed method by Gu and Vemuri addresses these challenges by transforming the 3D matching problem into a 2D optimization task. The framework relies on several key concepts from differential geometry:

1. Conformal Mapping and Canonical Domains

For surfaces with genus zero (topologically equivalent to a sphere), the authors map the 3D surface onto a 2D canonical domain, specifically a sphere. A conformal map is used because it preserves local angles, maintaining the geometric structure even as areas are stretched.

Commutative diagram illustrating the mapping between 3D surfaces ($S_1, S_2$) and their 2D canonical domains ($D_1, D_2$).

2. The $(\lambda, H)$ Representation

To ensure no geometric information is lost during dimensionality reduction, the surface is represented by a two-tuple signature:

  • Conformal Factor ($\lambda$): Captures the local stretching required to map the surface to the sphere.
  • Mean Curvature ($H$): Captures the physical "bumpiness" of the 3D object. According to the fundamental theorem of surface theory, this $(\lambda, H)$ pair uniquely determines a 3D surface up to a rigid motion.

3. Optimization via Möbius Transformations

Matching two 3D shapes is reduced to matching their $(\lambda, H)$ functions over the sphere. The algorithm searches for the optimal match within the group of Möbius transformations, which are the only conformal automorphisms of a sphere. This restriction inherently guarantees that the resulting 3D warp is a diffeomorphism. To initialize the search, three landmark points are used.

Experimental Results

The method was evaluated using both synthetic deformations of cortical surfaces and real anatomical data from human hippocampi.

Visualization of brain cortical surfaces mapped to a spherical domain, with color maps representing geometric signatures used for matching.

Accuracy and Efficiency

In synthetic tests, the algorithm demonstrated high precision, with mean errors in the range of $10^{-3}$ to $10^{-4}$. The computational efficiency was notable; on a Pentium 4 processor, computing the conformal representation took approximately 90 seconds, while estimating the 3D deformation for a surface with 5,000 faces took about 200 seconds. This performance is significantly faster than traditional volumetric warping techniques.

Robustness

The approach proved robust to significant distortions in conformal structure and is naturally invariant to rigid motions (rotation and translation) because it relies on intrinsic geometric properties.

Discussion and Limitations

While highly effective for genus-zero surfaces, the current framework cannot handle objects with holes (e.g., a torus) without further generalization. Additionally, the optimization depends on the accurate placement of at least three initial landmarks. Future work aims to extend these conformal representations to surfaces of arbitrary genus.

Summary updated

Abstract

Matching 3D shapes is a fundamental problem in Medical Imaging with many applications including, but not limited to, shape deformation analysis, tracking etc. Matching 3D shapes poses a computationally challenging task. The problem is especially hard when the transformation sought is diffeomorphic and non-rigid between the shapes being matched. In this paper, we propose a novel and computationally efficient matching technique which guarantees that the estimated non-rigid transformation between the two shapes being matched is a diffeomorphism. Specifically, we propose to conformally map each of the two 3D shapes onto the canonical domain and then match these 2D representations over the class of diffeomorphisms. The representation consists of a two tuple (λ, H), where, λ is the conformal factor required to map the given 3D surface to the canonical domain (a sphere for genus zero surfaces) and H is the mean curvature of the 3D surface. Given this two tuple, it is possible to uniquely determine the corresponding 3D surface. This representation is one of the most salient features of the work presented here. The second salient feature is the fact that 3D non-rigid registration is achieved by matching the aforementioned 2D representations. We present convincing results on real data with synthesized deformations and real data with real deformations.
BibTeX
@inproceedings{gu2004matching,
  author = {Xianfeng Gu and Baba C. Vemuri},
  title = {{Matching 3D Shapes Using 2D Conformal Representations}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2004},
  doi = {10.1007/978-3-540-30135-6\_94},
}
Figure