A dynamic graph-cuts method with integrated multiple feature maps for segmenting kidneys in ultrasound images

Qiang Zheng, Steven Warner, Gregory Tasian, Yong FanarXiv 2017 · arXiv

Introduction

Ultrasound (US) imaging is a primary modality for diagnosing kidney disease due to its non-invasive and cost-effective nature. However, automated segmentation of kidneys in US images remains a significant challenge. Unlike CT or MRI, ultrasound images are characterized by high speckle noise, low contrast between the kidney and surrounding tissues, and "weak boundaries" where the organ's edges are poorly defined. Furthermore, the high variability in kidney appearance across different patients complicates the development of robust automated tools.

Traditional methods like Active Contour Models (ACM) often struggle with US images because they rely on local gradients, making them prone to getting trapped in local minima caused by noise. While Graph Cuts (GC) offer global optimization, applying them to entire noisy images is computationally intensive and often inaccurate. This paper presents a semi-automatic approach that combines multi-scale texture features with a localized, dynamic graph-cuts algorithm to overcome these limitations.

Methodology: Multi-Feature Dynamic Graph Cuts

The proposed pipeline transitions from simple intensity-based analysis to a sophisticated multi-feature approach involving three primary stages:

1. Multi-Scale Texture Extraction

To combat speckle noise and enhance boundary detection, the researchers utilize Gabor filters at multiple scales and orientations. This process extracts texture information that is more stable than raw pixel intensity. A novel fusion scheme identifies dominant directions across scales to create a robust texture feature map that suppresses noise while enhancing edges.

The proposed multi-scale Gabor feature extraction process identifies dominant directions to generate a robust texture feature map.

2. Localized Dynamic Graph Construction

Efficiency and accuracy are improved by restricting the graph-cut optimization to a "narrow band" around an initial user-defined contour. This localization prevents the algorithm from being distracted by unrelated abdominal structures, leading to a more globally optimal solution for the specific boundary of interest.

3. Hybrid Similarity Measure and Iterative Optimization

The graph edge weights are calculated using a hybrid measure that integrates pixel-wise intensity differences with regional information (comparing a pixel to the mean intensity of its local neighborhood). The segmentation is performed iteratively; the graph is dynamically updated based on the previous iteration's results until the boundary converges.

Experimental Results

The algorithm was validated using a clinical dataset of 85 subjects from The Children’s Hospital of Philadelphia. Twenty subjects were used for parameter tuning, while the remaining 65 were used for performance evaluation against expert manual segmentations.

The integration of Gabor texture maps proved essential. As shown in the comparison below, the texture maps significantly suppress noise and clarify the kidney's internal and external structures compared to raw ultrasound data.

Visual comparison of raw ultrasound images (top row) and their corresponding texture feature maps (bottom row), demonstrating enhanced structural clarity.

The proposed method achieved an average Dice Index of 0.9581 and a Jaccard Index of 0.9204. In comparative tests against state-of-the-art techniques—including Localizing Region-based Active Contours (LRAC) and Geodesic Active Contours (GAC)—the proposed model showed statistically significant improvements (p < 10⁻¹⁹). Specifically, the Mean Distance error in boundary placement was reduced to 1.7166 pixels, compared to 2.98–4.02 pixels for competing methods.

Conclusion

By combining localized dynamic graph cuts with multi-scale texture features, this method provides a highly accurate solution for kidney segmentation in challenging ultrasound images. While currently semi-automatic—requiring a rough initial contour from a user—the high Intraclass Correlation Coefficients demonstrate its robustness to different initializations. Future work will focus on full automation and applying these high-fidelity segmentations to clinical tasks, such as monitoring the progression of chronic kidney disease (CKD) through parenchymal area measurements.

Summary updated

BibTeX
@misc{zheng2017dynamic,
  author = {Qiang Zheng and Steven Warner and Gregory Tasian and Yong Fan},
  title = {{A dynamic graph-cuts method with integrated multiple feature maps for segmenting kidneys in ultrasound images}},
  year = {2017},
  eprint = {1706.03372},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1706.03372},
}
Figure