Brain Tumor Segmentation Based on Structuring Element Map Modification and Marker-controlled Watershed Transform

Xiaopeng Wang, Shengyang Wan, Tao LeiJ. Softw 2014 · DOI

TL;DR

This paper introduces a novel method for segmenting brain tumors in medical images by adaptively modifying the image using a "structuring element map" before applying a marker-controlled watershed transform. This approach aims to overcome the limitations of traditional methods, such as over-segmentation and inaccurate contour detection, leading to more precise tumor localization.

What Problem Does This Paper Address?

Accurate brain tumor segmentation is crucial for diagnosis, treatment planning, and monitoring. However, medical images like MRI and CT scans often suffer from noise and intensity variations, making it difficult to precisely delineate tumor boundaries. Traditional segmentation methods, such as the watershed transform, are prone to over-segmentation, where a single tumor region is divided into multiple parts, or under-segmentation, where parts of the tumor are missed. Existing solutions like image pre-filtering or region merging have their own drawbacks, including loss of detail or inaccurate contour positioning. This paper tackles the challenge of achieving accurate and robust brain tumor segmentation by addressing the issues of noise, fuzzy boundaries, and intensity heterogeneity.

Methods

The proposed method combines two main techniques:

  1. Structuring Element Map Modification: This is the core novelty. Instead of using a fixed-size structuring element (SE) for morphological operations like opening and closing, this method constructs a "structuring element map." This map assigns a unique SE size to each pixel based on the weighted variance of its neighborhood in the morphological gradient image. This adaptive modification allows the algorithm to:
  2. Eliminate noise and small details without significantly shifting object contours.
  3. Preserve larger object contours more accurately.
  4. The morphological gradient is calculated using a circular SE of radius 1.
  5. The weighted variance is computed considering the distance of neighboring pixels to the current pixel.
  6. The structuring element map, $M(x,y)$, is derived from this weighted variance, with its values normalized to represent the maximum SE size ($R_{max}$).
  7. The original image is then modified using an adaptive opening-closing operation, where the SE size for each pixel is dictated by the corresponding value in the structuring element map.
  8. Marker-Controlled Watershed Transform: After the adaptive modification, the marker-controlled watershed transform is applied. This technique is used to segment the modified image. The process involves:
  9. Tumor Marker Extraction: Identifying potential tumor regions by thresholding the modified image, assuming tumors generally have higher gray values.
  10. Background Marker Extraction: Determining background regions by calculating the watershed transform of the Euclidean distance of the inner tumor regions.
  11. Watershed Segmentation: Applying the watershed transform on the modified gradient image, guided by the extracted foreground and background markers, to produce the final segmentation.

Datasets

The paper evaluates its method using:

  • Synthetic images: A synthetic image with four distinct regions (labeled A1, A2, A3, A4) is used for initial testing and demonstration of the method's capabilities.
  • Clinical CT images: Several clinical CT images of brain tumors are used for validation. One specific image is referred to as "CT-1."
  • Noisy images: The synthetic image is also tested with added Gaussian noise (0.1%) and salt-and-pepper noise (5%) to assess robustness.

The paper does not mention the creation of a new benchmark dataset.

Key Insights and Ideas

  • Adaptive Morphological Operations: The key insight is that a fixed-size structuring element in morphological operations is a significant limitation. By creating a "structuring element map" that adapts the SE size to local image characteristics (gradient variance), the method can achieve a more nuanced modification of the image.
  • Balancing Noise Reduction and Contour Preservation: The adaptive modification effectively balances the removal of noise and small details with the preservation of crucial object contours, a common trade-off in image processing.
  • Improved Robustness to Noise: The proposed method demonstrates better resilience to noise compared to standard watershed transform and even region merging techniques, which is critical for real-world medical imaging.
  • Precise Localization: The combination of adaptive modification and marker-controlled watershed transform leads to more accurate localization of tumor boundaries, closely matching manual segmentations.

Key Results

  • Quantitative Accuracy: The Tanimoto Metric (TM) is used to evaluate segmentation accuracy. For synthetic images, the proposed method achieved TM values around 0.9187 on average. For clinical CT images, the proposed method consistently outperformed standard watershed, region merging, and marker-controlled watershed transforms, with TM values often exceeding 0.90 for various tumor regions. For example, for CT-1, the proposed method achieved a TM of 0.9745 for region A1, compared to 0.9376 for region merging and 0.8104 for marker-controlled watershed.
  • Qualitative Results: Visual comparisons show that the proposed method produces segmentations that are significantly closer to manual segmentations than other tested methods, especially in challenging cases with noise or complex boundaries. It effectively reduces over-segmentation and avoids contour offsets.
  • Time Complexity: While watershed transform is the fastest, the proposed method is slower than basic watershed but faster than region merging techniques, offering a reasonable trade-off between speed and accuracy.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the proposed method is slower than the basic watershed transform.
  • The accuracy of the method is dependent on the quality of the initial markers and the chosen parameters ($R_{max}$, $T$, and $\alpha$).
  • Future Work:
    • The authors suggest further research into optimizing the parameter selection for the structuring element map and marker extraction.
  • Exploring more advanced techniques for marker generation could further enhance segmentation accuracy.
  • The paper implies that further validation on larger and more diverse datasets would be beneficial.

Summary updated

BibTeX
@article{wang2014brain,
  author = {Xiaopeng Wang and Shengyang Wan and Tao Lei},
  title = {{Brain Tumor Segmentation Based on Structuring Element Map Modification and Marker-controlled Watershed Transform}},
  journal = {J. Softw},
  year = {2014},
  doi = {10.4304/jsw.9.11.2925-2932},
}
Figure