Adaptive Region Growing Based on Boundary Measures

Yu Qiao, Jie YangICONIP 2011 · DOI

TL;DR

This paper introduces an adaptive region growing algorithm for image segmentation that robustly handles challenging scenarios like weak boundary contrast and inhomogeneous backgrounds. By combining local boundary measures for region expansion with a global measure for boundary selection, the method effectively identifies object boundaries even when they are discontinuous or obscured.

What Problem Does This Paper Address?

Accurate image segmentation is crucial for many applications, but it's often hindered by several common issues. In medical imaging, for instance, boundaries between structures can be faint (weak boundary contrast), making them hard to distinguish. Additionally, the background itself might not be uniform, with varying intensity levels (inhomogeneous background), or the intensity ranges of the object and its background might overlap, creating ambiguity. Traditional methods like thresholding, edge detection, and even some region growing techniques struggle with these problems, leading to inaccurate segmentation, such as over-segmentation or under-segmentation. This paper aims to develop a segmentation method that is resilient to these specific challenges.

Methods

The proposed method is an "adaptive region growing" algorithm that operates in two main phases:

  1. Region Expansion: This phase starts with a seed point and iteratively grows the region. Instead of relying solely on simple intensity similarity, it uses "local boundary measures" to evaluate potential points for expansion. These measures consider the intensity of a contour point relative to both the current region and its background. Specifically, it uses:
  2. Local Region Competition Measure ($v_0$): Compares the intensity difference between a contour point and the current region's mean intensity versus the difference with the background's mean intensity.
  3. Application-Dependent Measures ($v_1$, $v_2$): These measures assess if the contour point's intensity falls within the intensity range or standard deviation of the current region, helping to avoid misclassifying internal region points as boundaries.
  4. Local Classification Function ($h(x_e)$): Combines the above measures to classify contour points as either "region candidates" (similar to the region) or "boundary candidates" (significant contrast).
  5. Growing Cost: A "growing cost" is assigned to each contour point, primarily based on its local contour contrast. The point with the minimum growing cost is selected for expansion, ensuring the region grows along the path of least resistance.
  6. Boundary Selection: After the region has expanded as much as possible, a "global boundary measure" is used to select the final object boundary from the set of step contours. This measure, denoted by $\rho(i)$, calculates the proportion of "boundary candidates" within a given step contour. The contour with the highest global boundary measure is chosen as the final object boundary. This approach prioritizes contours that exhibit more significant boundary evidence, effectively handling "boundary openings" (discontinuities) by selecting the most complete boundary.

Datasets

The paper evaluates its method on real medical images, specifically:

  • Magnetic Resonance (MR) images of the brain, including ventricles, caudate nucleus, and putamen.
  • The paper does not mention introducing new datasets or benchmarks.

Key Insights and Ideas

  • Local Measures for Robustness: Using local boundary measures that consider intensity differences with both the region and background, and also incorporate intensity range and standard deviation, is key to handling inhomogeneous backgrounds and overlapping intensity distributions.
  • Growing Cost for Guided Expansion: Prioritizing expansion based on minimum "growing cost" (local contour contrast) allows the region to grow intelligently, even when faced with weak local gradients.
  • Global Measure for Boundary Completeness: The global boundary measure ($\rho(i)$) is a novel way to select the final boundary by evaluating its "completeness" (proportion of boundary candidates). This is crucial for overcoming boundary openings and weak contrasts that plague other methods.
  • Synergy of Local and Global: The combination of local measures for intelligent expansion and a global measure for final selection creates a powerful segmentation strategy that addresses multiple segmentation challenges simultaneously.

Key Results

The paper demonstrates the effectiveness of its proposed algorithm through qualitative results on MR images:

  • Robustness to Weak Contrast: The method successfully segments structures like the putamen, which have very weak boundary contrast, where Canny and Sobel edge detectors fail.
  • Handling Inhomogeneous Background: The segmentation of the caudate nucleus, which has a darker ventricle and brighter white matter as background, is shown to be successful.
  • Overcoming Boundary Openings: The algorithm avoids over-segmentation into the white matter through boundary openings in the caudate nucleus, unlike snake and level set methods.
  • Comparison with Other Methods: Qualitative comparisons show that the proposed method outperforms Canny/Sobel edge detection, GVF-snake, and level set methods in handling the aforementioned challenges. For instance, the level set method is shown to over-expand due to boundary openings.

Limitations and Future Work

  • Parameter Tuning: While the paper mentions parameters like $\lambda$ (region intensity range parameter) and $\kappa$ (region intensity threshold), the process of selecting optimal values for these parameters is not extensively detailed. The experimental results use specific values ($\lambda=3, \kappa=3$ for caudate nucleus; $\lambda=2, \kappa=2$ for putamen), implying some level of tuning is required.
  • Computational Complexity: The paper mentions that the local state classification simplifies computational complexity, but a detailed analysis of the overall computational cost compared to other methods is not provided.
  • Future Work: Not explicitly discussed in the paper. However, the robustness demonstrated suggests potential for application in a wider range of medical imaging tasks and other image analysis domains. Further exploration into automatic parameter selection or adaptive parameter estimation could be a natural next step.

Summary updated

BibTeX
@inproceedings{qiao2011adaptive,
  author = {Yu Qiao and Jie Yang},
  title = {{Adaptive Region Growing Based on Boundary Measures}},
  booktitle = {ICONIP},
  year = {2011},
  doi = {10.1007/978-3-642-24955-6\_30},
}
Figure