Measuring Size and Shape of the Hippocampus in MR Images Using a Deformable Shape Model

Dinggang Shen, Scott Moffat, Susan M. Resnick, Christos DavatzikosNeuroImage 2002 · DOI

TL;DR

This paper introduces an Adaptive-Focus Deformable Model (AFDM) for automatically segmenting and measuring the size and shape of the hippocampus in MRI scans. The method combines geometric properties, statistical shape priors, and user-defined landmarks to achieve accurate and reproducible results, significantly reducing manual effort compared to existing techniques.

What Problem Does This Paper Address?

Accurately segmenting the hippocampus in MRI scans is crucial for understanding its role in learning and memory and for diagnosing neurological and psychiatric conditions like Alzheimer's disease. However, the hippocampus presents significant challenges for automated analysis due to its small size, low contrast with surrounding tissues, and often indistinct boundaries. Traditional manual segmentation is time-consuming and prone to inter-rater variability. While fully automatic methods exist, they often struggle with the hippocampus's complex shape and subtle features. Semi-automatic methods offer a compromise but can still require substantial user interaction. This paper aims to bridge this gap by developing a robust, semi-automatic method that minimizes user burden while maintaining high accuracy and reproducibility.

Methods

The core of the proposed method is an Adaptive-Focus Deformable Model (AFDM). This model is built upon several key components:

  • Shape Model Construction: A 3D surface model of the hippocampus is created from a manually segmented image. This model is then enriched with affine-invariant attribute vectors. These vectors capture the geometric structure of the hippocampus at both local (e.g., curvature) and global scales, providing a rich representation of its shape. These attribute vectors are crucial for establishing correspondences between different individuals.
  • Statistical Shape Prior: To ensure that the segmented shapes remain anatomically plausible, the model incorporates a statistical characterization of normal shape variation learned from a training set of manually segmented hippocampi. This prior acts as a constraint, guiding the deformation to stay within the expected range of hippocampal shapes.
  • Landmark-Guided Deformation: To overcome challenges with weak or missing edges in MRI, the algorithm utilizes manually placed landmarks. These landmarks are not required to define precise anatomical correspondences but rather serve to guide the deformable model towards the correct boundaries, especially in difficult-to-define regions.
  • Energy Function Optimization: The segmentation process is framed as an optimization problem where an energy function is minimized. This function balances three terms:
  • Model Energy ($E_{model}$): Encourages attribute vector similarity between the model and the image, aiding in correspondence.
  • Data Energy ($E_{data}$): Pulls the model towards image-derived edges, primarily from a generated gray matter (GM) edge map.
  • Landmark Energy ($E_{landmark}$): Acts as a spring, pulling the model towards the manually placed landmarks.
  • Greedy Algorithm: A greedy algorithm is employed to iteratively minimize the energy function, deforming surface segments rather than individual points to avoid local minima and improve robustness. The focus shifts from smoother deformations to finer details as the process progresses.

Datasets

The paper uses MR images from the Baltimore Longitudinal Study of Aging (BLSA) for evaluation. These images were collected over a 9-year period from elderly subjects aged 55-85. The paper does not explicitly mention a separate training dataset for the statistical shape prior, but implies it is derived from manually traced hippocampi.

Key Insights and Ideas

  • Attribute Vectors as Geometric Signatures: The use of affine-invariant attribute vectors is a key innovation. They provide a robust way to characterize local and global shape features, enabling anatomically meaningful correspondences without explicit manual definition of homologous points.
  • Landmarks as Boundary Guides, Not Correspondences: The approach cleverly uses landmarks to guide segmentation without requiring them to define precise anatomical homologies. This significantly reduces the burden on the user.
  • Hierarchical Deformation for Robustness: Deforming surface segments hierarchically, starting with smoother movements and progressing to finer details, makes the algorithm more robust to noisy edges and spurious features in the MRI.
  • Statistical Prior as a "Look-alike" Constraint: The statistical shape prior is crucial for ensuring that the segmented hippocampus "looks like" a hippocampus, preventing the model from being misled by incorrect edges.

Key Results

  • High Accuracy: The algorithm achieved a mean volume error of 2.9% and a mean overlap error of 11.5% when compared to expert manual segmentations. This performance is statistically equivalent to manual segmentation.
  • Reproducibility: The algorithm demonstrated high reproducibility, with minimal variations in volume and overlap errors when subjected to slight shifts in initialization or variations in landmark placement.
  • Reduced User Time: The proposed method significantly reduces the time required for segmentation compared to manual methods, with landmark placement taking approximately 20 minutes for a skilled user.
  • Landmark Sensitivity: While the algorithm performs well with all landmarks, it shows a slight increase in volume error (up to 4.1%) when using only 50% or 25% of the landmarks, indicating the importance of sufficient landmark guidance for accurate volume estimation, though overlap error remains relatively stable.

Limitations and Future Work

  • Manual Landmark Placement: The primary limitation acknowledged is the need for manual placement of landmarks, which, while reduced, still requires user effort.
  • Inclusion of Surrounding Structures: The current model only includes the left and right hippocampi. Future work aims to incorporate surrounding structures to better differentiate hippocampal boundaries, especially in regions with poor edge information.
  • Image-Derived Forces: The reliance solely on edge maps for image-derived forces can be problematic due to noise. Future research will explore incorporating other types of image information.
  • Extension to Other Structures: The AFDM framework is generalizable and could be applied to segment other brain structures.

Summary updated

BibTeX
@article{shen2002measuring,
  author = {Dinggang Shen and Scott Moffat and Susan M. Resnick and Christos Davatzikos},
  title = {{Measuring Size and Shape of the Hippocampus in MR Images Using a Deformable Shape Model}},
  journal = {NeuroImage},
  year = {2002},
  doi = {10.1006/nimg.2001.0987},
}
Figure