Shape Registration with Spherical Cross Correlation

Boris A. Gutman, Yalin Wang, Tony F. Chan, Paul M. Thompson, Arthur W. Toga2008

TL;DR

This paper introduces a novel framework for aligning 3D shapes by mapping them to a sphere and using spherical cross-correlation. This method generalizes existing techniques, offers a global search that guarantees optimal alignment, and is significantly faster than previous approaches, making it more practical for complex shape analysis tasks.

What Problem Does This Paper Address?

Aligning 3D shapes is a fundamental problem in many fields, especially medical imaging, where it's crucial for comparing anatomical structures, creating atlases, and analyzing variations across populations. While many specialized methods exist for common shapes like the cortex or hippocampus, a robust and general method for arbitrary genus-zero closed surfaces remains an open challenge. Existing general methods often struggle with local minima, require manual initialization, or are computationally expensive. This paper aims to bridge this gap by providing a general, robust, and efficient shape registration framework.

Methods

The core of the proposed framework is a spherical cross-correlation method. The process involves:

  1. Spherical Parameterization: Each 3D shape (assumed to be a closed genus-zero surface) is mapped to a 2-sphere using a diffeomorphic surface mapping that preserves rotation. This transforms the 3D shape problem into a 2D problem on the sphere.
  2. Shape Attributes: Surface-intrinsic scalar attributes are defined on the 3D shape. These can be simple (like distance to mass center) or more complex (like mean and Gaussian curvature). These attributes are then mapped to the sphere.
  3. Spherical Cross-Correlation: A global spherical cross-correlation function is computed between the mapped attributes of two shapes. This function measures the similarity between the shapes across all possible rotations.
  4. Parameterization Distortion Weighting: The cross-correlation is weighted by the cross-correlation of the parameterization distortion (often referred to as the conformal factor). This accounts for how the spherical mapping stretches or shrinks different regions of the surface, ensuring that regions of similar size on the sphere correspond to regions of similar size on the original 3D shape.
  5. Fast Computation: The authors leverage the Fast Fourier Transform (FFT) to significantly speed up the computation of the spherical cross-correlation, making the method computationally feasible.

What's novel is the combination of a global search strategy (avoiding local minima), the use of orientation-invariant shape attributes, and the efficient FFT-based computation, which generalizes and improves upon prior work.

Datasets

The paper evaluates its method on a dataset of 45 right hippocampal surfaces extracted from healthy elderly subjects. This dataset is used to demonstrate the effectiveness of the proposed registration method and compare it with existing techniques like FOE and SHREC. No new benchmarks are introduced; rather, existing datasets are used for validation.

Key Insights and Ideas

  • Global Optimization Guarantee: By framing the problem as a global spherical cross-correlation, the method guarantees finding the optimal alignment, unlike iterative methods (like ICP variants) that can get stuck in local minima.
  • Rotation Invariance: Using orientation-invariant shape attributes on the sphere makes the alignment independent of the object's initial position and size.
  • Efficiency through FFT: The application of FFT for spherical cross-correlation dramatically reduces computation time, making complex analyses practical.
  • Flexibility in Similarity Measure: The framework allows for customization of the similarity measure by selecting different surface-intrinsic scalar functions, enabling tailoring to specific applications.
  • Parameterization Distortion Matters: Explicitly weighting by parameterization distortion is crucial for accurate alignment, especially when dealing with non-area-preserving spherical maps.

Key Results

  • The proposed method achieved superior alignment compared to FOE and a modified SHREC algorithm on the hippocampal dataset.
  • The average running time for a single subject registration with a bandwidth of 64 was approximately 44.6 ± 3.6 seconds on a standard laptop. This is significantly faster than estimates for SHREC (around 60 minutes for a comparable bandwidth).
  • Increasing the bandwidth (level of detail in spherical harmonics) improved alignment quality, with a bandwidth of 64 generally outperforming a bandwidth of 16.
  • The method demonstrated a significant improvement in weighted point-wise distance to the target subject compared to FOE, with improvements ranging from 8% to 117%.

Limitations and Future Work

  • Genus-Zero Assumption: The framework currently assumes shapes are closed genus-zero surfaces. Extending it to handle more complex topologies (e.g., surfaces with holes) would be a valuable next step.
  • Parameterization Choice: While the method works with various parameterizations, the quality of alignment can be influenced by the chosen spherical map, especially if it exhibits significant area distortion. The authors suggest exploring area-preserving spherical maps in future work.
  • Shape Attributes: The current experiments primarily use distance to mass center. Future work could explore a wider range of shape attributes and their impact on alignment quality.
  • Automated Patch Selection: The authors express interest in extending the technique to automated patch selection and matching, which could allow for aligning specific regions of interest even if the overall shapes differ significantly.

Summary updated

Abstract

We present a framework for shape alignment that generalizes several existing methods. We assume that the shape is a closed genus zero surface. Our framework requires a diffeomorphic surface mapping to the 2-sphere which preserves rotation. Our similarity measure is a global spherical cross-correlation function of surface-intrinsic scalar attributes, weighted by the cross-correlation of the parameterization distortion. The final similarity measure may be customized according to the surfaceintrinsic scalar functions used in the application. 1
BibTeX
@misc{gutman2008shape,
  author = {Boris A. Gutman and Yalin Wang and Tony F. Chan and Paul M. Thompson and Arthur W. Toga},
  title = {{Shape Registration with Spherical Cross Correlation}},
  year = {2008},
}
Figure