Left Ventricle MR image segmentation Based on Locally Weighted Entropy Geodesic Active Contour model

Jianwei Zhang, Xiang Ma, Yunjie Chen, Fang Lin, Jin Wang

TL;DR

This paper introduces an improved active contour model for segmenting left ventricle MR images, particularly addressing challenges like noise, bias fields, and weak edges. By incorporating local information entropy and a novel signed pressure force function, the model achieves higher segmentation accuracy and robustness compared to existing methods.

What Problem Does This Paper Address?

Accurate segmentation of medical images, especially MRI, is crucial for diagnosis and treatment planning. However, medical images often suffer from artifacts like intensity inhomogeneity (bias fields) and weak boundaries, which pose significant challenges for traditional image segmentation algorithms. Existing active contour models (ACMs) often struggle with these issues: edge-based models fail with weak edges, while region-based models can be sensitive to intensity variations. While some models like LBF and RGAC attempt to address inhomogeneity, they still have limitations with noise and weak boundaries. This paper aims to bridge this gap by developing a more robust active contour model that can effectively handle these common image imperfections.

Methods

The core contribution of this paper is an enhanced Geodesic Active Contour (GAC) model. The authors introduce two key innovations:

  • Window Function for Local Information: Instead of relying solely on global image properties, the model uses a window function to analyze the local neighborhood of each pixel. This allows for a more nuanced understanding of the image's intensity distribution.
  • Locally Weighted Entropy Geodesic Active Contour (LW-EGAC) Model: A novel "signed pressure force function" is constructed based on local information entropy. This function leverages the entropy of the local region to guide the contour towards the true boundary. The entropy is calculated using a Gaussian distribution assumption for local pixel intensities. The model also incorporates a two-valued level set method for improved efficiency and stability.

The paper builds upon the GAC model's energy functional and level set evolution equation, replacing the standard edge-stopping function with their new entropy-based pressure force. They also discuss how their approach can be simplified by approximating the gradient of the pressure force and using a Gaussian filter for regularization, leading to a more efficient gradient descent flow.

Datasets

The authors evaluated their method using 20 groups of left ventricle MR images, each with a size of 110x110 pixels. They also used a "true segmentation result" as a ground truth for comparison. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Entropy as a Boundary Indicator: The paper highlights that local image entropy can effectively capture information about the underlying tissue boundaries, especially in the presence of noise and inhomogeneity. Regions with uniform intensity (low entropy) are less informative for boundary detection than regions with varying intensities.
  • Local Context is Key: By focusing on local image information through a window function, the model becomes more resilient to global intensity variations (bias fields) that can mislead purely global methods.
  • Balancing Accuracy and Efficiency: The adoption of a two-valued level set method and the simplification of the gradient descent flow demonstrate a conscious effort to balance segmentation accuracy with computational efficiency.

Key Results

The proposed LW-EGAC model demonstrated superior performance compared to several established methods:

  • High Segmentation Accuracy: The method achieved segmentation accuracy of over 90% on their test dataset.
  • Robustness to Bias Fields and Weak Edges: The model effectively segmented images with bias fields and weak boundaries, outperforming C-V, GAC, and RGAC models in these challenging scenarios.
  • Improved Noise Handling: While LBF and LGD models also showed good performance with bias fields, they struggled with noisy images, an area where the proposed method showed better results.
  • Reasonable Efficiency: The average iteration time was reported to be under 3 seconds with fewer than 200 iterations, which is considered a reasonable range for practical applications.

Limitations and Future Work

  • Parameter Sensitivity: While the paper mentions that their method has "more choice of kernel parameters" compared to LBF and LGD, it doesn't explicitly detail the sensitivity of their own parameters (like sigma and alpha) or provide a systematic way to tune them.
  • Computational Cost: Although deemed "reasonable," the paper doesn't provide a detailed comparison of computational complexity against all benchmarked methods. The use of window functions and entropy calculations can still add to the computational load.
  • Future Work: The paper doesn't explicitly outline future research directions, but the success in handling bias fields and weak edges suggests potential applications in other medical imaging modalities or more complex anatomical structures. Further exploration into automatic parameter selection would also be a valuable next step.

Summary updated

Abstract

In this paper, an improved Geometric Active Contour model is presented. This model introduced a window function to research the mean information of each pixel's neighbor region, and constructed a novel signed pressure force function based on the entropy to drive the contour towers the boundary. In order to improve the efficiency and stability of the algorithm, this paper adopts two valued level set method to realize the segmentation process. The experimental results show that this method can obtain satisfied result even when the images have intensity inhomogeneity and weak edges.
BibTeX
@article{zhang2014left,
  author = {Jianwei Zhang and Xiang Ma and Yunjie Chen and Fang Lin and Jin Wang},
  title = {{Left Ventricle MR image segmentation Based on Locally Weighted Entropy Geodesic Active Contour model}},
  journal = {International Journal of Multimedia and Ubiquitous Engineering},
  year = {2014},
  doi = {10.14257/ijmue.2014.9.4.14},
}
Figure