Lung vessel segmentation in CT images using graph cuts

Zhiwei Zhai, Marius Staring, Berend C. StoelImage Processing 2016 · DOI

TL;DR

This paper introduces a novel graph cuts-based method for segmenting lung vessels in CT scans by combining appearance (CT intensity) and shape (vesselness) information. This approach achieves competitive performance on a challenging benchmark dataset, outperforming simpler thresholding methods and offering a robust solution for medical image analysis.

What Problem Does This Paper Address?

Accurately segmenting lung vessels in CT images is crucial for diagnosing and monitoring various pulmonary diseases. Traditional methods, like those based on Hessian filters, struggle with vessel bifurcations and boundaries, leading to inaccurate segmentations when simply thresholding the "vesselness" output. While graph cuts offer a way to incorporate neighborhood information for better segmentation, effectively combining different sources of information (like appearance and shape) into a cost function remains a challenge. This paper aims to address this by proposing a more sophisticated graph cuts cost function that leverages both the intensity of the CT scan and a derived shape descriptor.

Methods

The proposed method builds upon existing techniques but introduces a novel cost function for graph cuts. It consists of three main steps:

  1. Strain Energy Filter for Vessel Enhancement: This step uses a strain energy filter, an advancement over traditional Hessian-based filters, to enhance vessel-like structures. It's designed to overcome the limitations of Hessian filters at bifurcations and boundaries by considering strain energy density and intensity continuity.
  2. Graph Representation with Pruning: To manage the computational complexity of segmenting large 3D CT volumes, the authors employ a strategy to reduce the number of nodes in the graph. Voxels with very low "vesselness" are pruned as they are likely background. The remaining voxels form the nodes of the graph, with edges connecting neighboring voxels. A sparse matrix representation is used for efficiency.
  3. Graph Cuts Cost Function: This is the core innovation. The energy function for graph cuts combines:
  4. Data Term: This term uses both appearance (CT intensity) and shape (vesselness from the strain energy filter) to determine the likelihood of a voxel belonging to a vessel. Crucially, they use sigmoid functions instead of Gaussian functions for the data term, which they argue better handles voxels far from the center of the distribution.
  5. Smoothness Term: This term penalizes differences in labels between neighboring voxels, encouraging smooth segmentations. The cost is weighted by the spatial distance between voxels.

Datasets

  • Manually Labeled Sub-volumes: Two sub-volumes were manually annotated by an expert for initial evaluation and parameter optimization. These were used to compare the proposed method against thresholding Frangi's vesselness, thresholding the strain energy filter's vesselness, and a Freiman's based method.
  • VESSEL12 Challenge Dataset: This dataset consists of 20 anonymized CT scans from three hospitals. The manual labeling was performed on pre-generated points, and the segmentation vessels were evaluated against these annotations. This dataset was used for independent evaluation.

Key Insights and Ideas

  • Combining Appearance and Shape is Key: The paper strongly advocates for integrating both intensity (appearance) and vesselness (shape) information within the graph cuts framework for improved segmentation accuracy.
  • Sigmoid Functions for Data Term: Using sigmoid functions for the data term in graph cuts is a clever way to handle voxels with high or low intensity/vesselness values more effectively than traditional Gaussian distributions, especially when dealing with the wide range of values in medical images.
  • Pruning for Efficiency: Reducing the graph size by pruning background voxels based on a low vesselness threshold is a practical necessity for handling large 3D datasets and makes the graph cuts approach computationally feasible.
  • Graph Cuts as a Unifying Framework: The paper demonstrates how graph cuts can elegantly combine diverse information sources and neighborhood constraints into a single optimization problem.

Key Results

  • Sub-volume Evaluation: The proposed method achieved better segmentation results on the manually labeled sub-volumes compared to thresholding-based methods and a Freiman's based method, particularly in terms of F1 score. For Dataset1, it achieved an F1 score of 0.761, and for Dataset2, 0.690.
  • VESSEL12 Challenge Performance: On the VESSEL12 dataset, the method achieved a competitive area under the ROC curve (Az) of 0.975, outperforming many binary submissions in the challenge. This indicates a high level of accuracy and robustness across a diverse set of clinical images.

Limitations and Future Work

  • Potential for Improvement: The authors suggest that the method could be further improved by incorporating additional information, such as the radius of vessels into the data term.
  • Reducing False Positives: Designing a more accurate boundary cost function that considers differences in multiple features and explicitly removing airway walls and nodules are suggested as ways to reduce false positives.

Summary updated

BibTeX
@inproceedings{zhai2016lung,
  author = {Zhiwei Zhai and Marius Staring and Berend C. Stoel},
  title = {{Lung vessel segmentation in CT images using graph cuts}},
  booktitle = {Image Processing},
  year = {2016},
  doi = {10.1117/12.2216827},
}
Figure