Statistical Neighbor Distance Influence in Active Contours

Jing Yang, Lawrence H. Staib, James S. DuncanMICCAI 2002 · DOI

Abstract

Accurate segmentation of medical images is frequently hampered by poor contrast, sensor noise, and diffuse boundaries. Traditional active contour models often fail in these environments because they lack global anatomical context. This paper introduces a multi-object segmentation framework that leverages the statistical relationship between neighboring structures. By modeling the distribution of distances between adjacent objects from a training set and incorporating this as a prior in a level-set energy functional, the proposed method significantly improves boundary detection in noisy and low-contrast images.

Introduction

Image segmentation is a fundamental task in medical image analysis, yet it remains challenging due to inherent image artifacts. Standard "snake" or active contour models evolve curves based on local image features like gradients. However, when boundaries are missing or obscured by noise, these models lack the necessary guidance to converge correctly. This research addresses this limitation by introducing "spatial context"—the consistent relative positions and shapes of neighboring anatomical structures—as a prior to guide the segmentation process.

Methodology

The proposed model is built upon the Level Set method and the Chan-Vese "active contours without edges" framework. The core innovation is the inclusion of a "neighbor distance" term in the energy functional.

The process begins with Binary Shape Alignment, where training images are aligned using similarity transformations to ensure distance measurements reflect relative anatomy. The Neighbor Distance Representation defines a metric $d_{ijk}$ representing the distance from a point on the boundary of object $i$ to the boundary of object $j$ along the normal direction.

Diagram illustrating the neighbor distance representation $d_{ijk}$ between two objects.

These distances are modeled using a Gaussian distribution. The resulting Energy Functional combines standard image intensity information and regularization terms with this new neighbor influence term, which pulls evolving curves toward the maximum likelihood estimation of their distance from neighbors. This allows for the Simultaneous Evolution of multiple objects, where each contour dynamically influences its neighbors.

Experimental Evaluation

The authors evaluated their model using three primary datasets:

  • Synthetic Data: A training set of 9 images containing two elliptic objects with added Gaussian noise.
  • Cardiac MRI: 16 2D MR heart images used to segment the left and right ventricles.
  • Brain MRI: 12 2D MR brain images used to segment eight different subcortical structures (including the lateral ventricles, caudate nucleus, and putamen).

Results and Discussion

The inclusion of neighbor influence led to significant quantitative improvements. In synthetic cases, the undirected distance error dropped from 5.3 to 2.1. More impressively, in complex brain MRI segmentation involving eight structures, the error was reduced from 9.6 to 2.8.

Comparison of level set evolution on noisy synthetic data without (top) and with (bottom) neighbor influence.

A key insight is that an object's neighbor often serves as its best landmark. Even if a specific boundary is blurry, its known spatial relationship to a clearer neighbor provides sufficient information for accurate localization. This mutual influence allows the system to stabilize as clearer edges are found, providing better reference points for more difficult regions. Qualitatively, the brain MRI results were particularly striking; the model successfully segmented eight distinct structures in approximately 20 minutes on 2002-era hardware, whereas standard models failed to lock onto low-contrast subcortical structures.

Conclusion

While the model demonstrates robust performance in 2D, extending it to 3D presents challenges in computational complexity and shape alignment. Furthermore, the model's reliance on a training set means its accuracy is tied to how well the training data represents anatomical variation. Despite these limitations, the integration of statistical neighbor distances provides a powerful mechanism for overcoming the limitations of local-only segmentation methods, offering a more context-aware approach to medical image analysis.

Summary updated

BibTeX
@inproceedings{yang2002statistical,
  author = {Jing Yang and Lawrence H. Staib and James S. Duncan},
  title = {{Statistical Neighbor Distance Influence in Active Contours}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2002},
  doi = {10.1007/3-540-45786-0\_73},
}

Loading related content...

Figure