Segmentation of White Blood Cells through Nucleus Mark Watershed Operations and Mean Shift Clustering

Zhi Liu, Jing Liu, Xiaoyan Xiao, Hui Yuan, Xiaomei Li, Jun Chang, Chengyun Zheng

TL;DR

This paper introduces a novel method for segmenting white blood cells (WBCs) in peripheral blood and bone marrow images, even under challenging lighting conditions and with cell adhesion. By combining color space transformations, mean shift clustering, and a nucleus-marked watershed operation, the algorithm achieves high accuracy and robustness, outperforming existing methods.

What Problem Does This Paper Address?

Accurate identification and segmentation of white blood cells (WBCs) are crucial for diagnosing various blood-related diseases. Traditional manual counting methods are time-consuming and prone to errors. While automated methods exist, they often struggle with variations in lighting, similar appearances between cell types, and the common problem of cell adhesion, where multiple cells clump together, making individual segmentation difficult. This paper aims to address these limitations by developing a robust automated system for WBC segmentation.

Methods

The proposed method is a multi-phase approach:

  • Phase I: Nucleus and Inside Seed Acquisition: This phase focuses on identifying the nucleus of WBCs. It involves:
  • Color Space Conversion: Transforming the RGB images into the rg chroma space and HSI color space to better isolate nucleus features.
  • Image Enhancement: Applying techniques to enhance the nucleus visibility.
  • Adaptive Thresholding and Morphology: Using Otsu's method and morphological operations to create a binary image of the nucleus and then refine it to obtain "inside seeds" (nucleus groups). A novel centroid-connected operation is introduced to handle multi-core nuclei.
  • Phase II: Outside Seed Acquisition: This phase aims to identify the outer boundaries of WBCs, including potential impurities. It utilizes:
  • Mean Shift Clustering: A non-parametric clustering algorithm applied to the C channel component of the CMYK color space. This is chosen because the C channel effectively distinguishes between nucleus, cytoplasm, and background based on brightness differences. The algorithm is applied in a 3D feature space (C channel, x-coordinate, y-coordinate).
  • WBC Enhancement: Further image enhancement techniques are applied to improve the visibility of WBCs, especially in weaker areas.
  • Adaptive Thresholding: To obtain "outside seeds" representing WBCs with potential impurities.
  • Phase III: Nucleus Mark Watershed Operation (NMWO): This is the core of the method for handling cell adhesion.
  • Watershed Transformation: A region-based segmentation approach is employed. The nucleus (inside seeds) acts as markers for the watershed algorithm.
  • Adhesion Detection: The algorithm determines if cell adhesion has occurred. If so, a second watershed transformation is applied, often using distance maps and adaptive iteration corrosion, to separate the adhered cells.
  • Phase IV: Post-Processing: This final step refines the segmentation by applying logical operations ("AND") between the nucleus and the segmented WBC regions to remove impurities and ensure accurate separation of individual WBCs. Morphological reconstruction is used to clean up the results.

Datasets

The study utilized two datasets:

  • Dataset 1: Collected from the Second Affiliated Hospital of Shandong University. This dataset consists of 306 images with 774 WBCs, captured under different lighting conditions from three types of smears (normal peripheral blood, M3 bone marrow, and M5 bone marrow). The images have a resolution of 2080 × 1542 pixels.
  • Dataset 2: Consists of 108 images downloaded from the ALL-IDB1 dataset. These images were captured in RGB format with varying resolutions (2592 × 1944, 1712 × 1368, and 1226 × 652 pixels).

Key Insights and Ideas

  • Nucleus as a Reliable Seed: The nucleus of a WBC is consistently used as a strong "inside seed" for watershed segmentation, providing a stable starting point even in challenging conditions.
  • CMYK C-Channel for Clustering: The C channel of the CMYK color space proves effective for mean shift clustering due to its ability to differentiate cell components based on brightness.
  • Multi-Stage Watershed for Adhesion: Employing a two-stage watershed transformation, guided by nucleus markers and distance maps, is crucial for effectively separating adhered cells.
  • Centroid-Connected Operation for Multi-Core Nuclei: A novel approach to consolidate multi-core nuclei into a single marker, improving the robustness of nucleus detection.
  • Robustness to Lighting: The combination of color space transformations and illumination adjustment strategies makes the algorithm resilient to variations in lighting conditions.

Key Results

The proposed algorithm demonstrated superior performance compared to traditional methods like color clustering, color and shape transformation, and SLIC.

  • Quantitative Comparison (Table 3 & 4): The proposed algorithm achieved significantly higher True Positives (TP) and a lower False Positive (FP) and False Negative (FN) rate. Its F-measure was consistently higher across various metrics. For instance, the proposed algorithm achieved 99% precision and 97% recall, outperforming other methods.
  • Cell Type Segmentation (Table 5): The algorithm showed improved accuracy in segmenting different WBC types across both datasets.
  • Visual Results (Figures 15-19): The visual comparisons clearly show that the proposed method provides more accurate and complete segmentation of WBCs, especially in cases of cell adhesion and varying lighting.

Limitations and Future Work

  • Watershed Ridge Line Matching: In some images, the watershed ridge line did not perfectly match the cell boundaries.
  • Complex Adhesion and Morphology: For extremely complex cases of cell adhesion and morphology, further refinement using advanced morphological characteristics might be necessary.
  • Distinguishing Similar Cell Types: While the algorithm is robust, distinguishing between specific subtypes of leukemia (e.g., ALL vs. AML, M3 vs. other AML) via automatic identification remains a challenge and requires further research, as mandated by current WHO standards.

The authors suggest that future work should focus on addressing these limitations, particularly in handling highly complex cell adhesion scenarios and refining the differentiation of specific disease-related cell subtypes.

Summary updated

BibTeX
@inproceedings{liu2015segmentation,
  author = {Zhi Liu and Jing Liu and Xiaoyan Xiao and Hui Yuan and Xiaomei Li and Jun Chang and Chengyun Zheng},
  title = {{Segmentation of White Blood Cells through Nucleus Mark Watershed Operations and Mean Shift Clustering}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=yuQrjoSbX7b},
}
Figure