Hippocampal Morphometry Study by Automated Surface Fluid Registration and its Application to Alzheimer's Disease

Jie Shi, Yuting Wang, Paul M. Thompson, Yalin Wang2011

TL;DR

This paper introduces a novel automated system for analyzing hippocampal shape changes in Alzheimer's Disease (AD) by combining conformal surface parameterization with fluid registration. The method achieves more robust and sensitive detection of subtle atrophic changes compared to existing techniques, paving the way for earlier and more accurate AD diagnosis.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) is a major public health concern, and early detection is crucial for effective management. Hippocampal atrophy, a well-established biomarker for AD, is often assessed using MRI. However, accurately quantifying these subtle shape changes across individuals is challenging. Traditional methods often rely on volumetric measurements or require manual landmarking for surface alignment, which can be time-consuming and prone to error. This paper addresses the need for an automated, robust, and sensitive method to align and compare hippocampal surfaces, enabling more precise identification of AD-related atrophy patterns.

Methods

The core of the proposed system lies in a two-pronged approach:

  • Conformal Surface Parameterization: The authors leverage conformal mapping to transform complex 3D hippocampal surfaces into 2D planar domains. This process is "angle-preserving," meaning it preserves local geometric relationships. Crucially, they introduce a novel and stable method for computing surface mean curvatures directly from the derivatives of this conformal parameterization. This "conformal representation" (combining conformal factor and mean curvature) captures intrinsic geometric features of the hippocampus.
  • Mutual Information Regularized Fluid Registration: Once the surfaces are represented in their 2D parameter domains, the problem of surface registration is transformed into an image registration problem. They employ a fluid registration technique, driven by mutual information (MI). MI is a powerful metric for image alignment as it measures the statistical dependence between two images. By regularizing the fluid flow with MI, the system ensures that the resulting correspondences are "diffeomorphic," meaning they are smooth and one-to-one, preserving the topological integrity of the shapes.

The combination of conformal mapping and fluid registration allows for automated alignment of surface features without the need for manual landmarking.

Datasets

The paper evaluates the proposed system on three distinct hippocampal datasets:

  1. Synthetic Surfaces: Used to visually demonstrate the workflow and the effectiveness of feature matching in the 2D parameter domain.
  2. Clinical Dataset (12 AD vs. 14 Control): This dataset was manually traced by experienced neurologists and was previously used in a similar study [19].
  3. ADNI Baseline Dataset: A larger dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) comprising 40 AD, 40 Mild Cognitive Impairment (MCI), and 40 control subjects. The hippocampal surfaces were automatically segmented in a prior work [14].

Key Insights and Ideas

  • Conformal Representation for Robust Feature Extraction: Using conformal factor and mean curvature derived from conformal parameterization provides a stable and intrinsic representation of hippocampal surface geometry, less sensitive to noise or errors in normal estimation compared to other methods.
  • Bridging 3D Surface Registration to 2D Image Registration: Transforming the complex 3D surface registration problem into a more manageable 2D image registration problem via conformal mapping is a key conceptual advance.
  • Diffeomorphic Correspondences for Accurate Shape Analysis: The use of fluid registration regularized by mutual information ensures that the generated mappings are diffeomorphic, which is critical for accurately capturing and quantifying subtle shape deformations associated with disease.
  • Automated Workflow: The system eliminates the need for manual landmarking, making the analysis process more efficient and reproducible.

Key Results

The proposed system demonstrated superior performance across multiple evaluations:

  • Synthetic Data: Visualizations showed improved matching of features between synthetic surfaces.
  • Clinical Dataset: The system detected larger areas with more significant group differences in shape compared to a previous method [23], achieving statistically significant results (p=0.0063 for left hippocampus, p=0.0298 for right) that outperformed the prior method (p=0.0205 for left, p=0.1026 for right).
  • ADNI Dataset: The system showed better detection of MCI, an intermediate stage of cognitive decline, compared to both AD and control subjects.
  • Comparison with SPHARM: When compared to the popular SPHARM tool on the ADNI dataset, the proposed system exhibited higher robustness, with significantly fewer subjects failing the registration process (e.g., 5 control, 17 AD, 14 MCI failed in SPHARM vs. 1 subject in the proposed method). The FDR-corrected p-values also indicated better performance in detecting group differences.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the current system is effective for AD detection, it is a general method that could be applied to other applications. The specific limitations of the conformal parameterization or fluid registration itself are not extensively detailed, but the improved performance over prior methods suggests a high degree of robustness.
  • Future Work: The authors plan to extend their system to automatically map lateral ventricle enlargements in Alzheimer's disease and in individuals at risk. This indicates a broader ambition to apply their automated morphometry techniques to other brain structures and disease states.

Summary updated

Abstract

Alzheimer’s disease (AD) is a severe and growing public health crisis. Efforts are underway to look for AD early detection in an efficient manner. Among all the AD biomarkers, hippocampal atrophy assessed on high-resolution T1-weighted MRI is the best established and validated. Hippocampal morphometry is increasingly used in the AD research, with modeling the hippocampus as a 3D parametric surface mesh. However, a major question in the analysis is how to align corresponding surface regions across subjects. Here we develop a system for detecting AD symptoms on hippocampal surfaces with an automated surface fluid registration method, which is based on conformal surface representation and mutual information regularized image fluid registration. Since conformal mappings are diffeomorphic and the mutual information method is able to drive a diffeomorphic flow that is adjusted to enforce appropriate surface correspondences in the surface parameter domain, combining conformal and fluid mappings will generate 3D shape correspondences that are diffeomorphic. We also incorporate in the system a novel method to compute curvatures using surface conformal parameterization. Experimental results in three hippocampal datasets show that the new system outperformed an early similar method and the popular SPHARM tool. 1
BibTeX
@misc{shi2011hippocampal,
  author = {Jie Shi and Yuting Wang and Paul M. Thompson and Yalin Wang},
  title = {{Hippocampal Morphometry Study by Automated Surface Fluid Registration and its Application to Alzheimer's Disease}},
  year = {2011},
}
Figure