Segmentation of White Blood Cell from Acute Lymphoblastic Leukemia Images Using Dual-Threshold Method

Yan Li, Rui Zhu, Mi Lei, Yihui Cao, Di YaoComputational 2016 · DOI

TL;DR

This paper introduces a novel dual-threshold segmentation method that effectively combines RGB and HSV color spaces to accurately identify white blood cells (WBCs), specifically lymphoblasts, in microscopic images. By leveraging the complementary strengths of these color spaces, the method achieves a high segmentation accuracy of 97.85%, offering a promising advancement for automated leukemia diagnosis.

What Problem Does This Paper Address?

Accurate segmentation of white blood cells (WBCs) is a critical first step in automated blood cell morphology analysis, particularly for diagnosing diseases like Acute Lymphoblastic Leukemia (ALL). However, this task is challenging due to the inherent morphological diversity of WBCs and the complex, often noisy backgrounds in microscopic images. Existing methods, often relying on single-thresholding in either RGB or HSV color spaces, struggle with variations in illumination and the similar color intensities between WBC cytoplasm and red blood cells (RBCs). This leads to incomplete or inaccurate segmentation, hindering subsequent classification and diagnosis. The paper aims to overcome these limitations by developing a more robust and accurate segmentation technique.

Methods

The proposed method employs a three-stage approach: preprocessing, threshold segmentation, and postprocessing.

  • Preprocessing: The input RGB image is converted into a grayscale image and its H component from the HSV color space. Global contrast stretching is applied to the grayscale image to enhance details and improve contrast between foreground and background.
  • Threshold Segmentation: This is the core of the novel approach. It uses a dual-threshold method that strategically combines information from both the contrast-stretched grayscale image (RGB-based) and the H component image (HSV-based).
  • The grayscale image is used to extract the image background using a single threshold (Thresh1).
  • The H component image is used to separate red blood cells (RBCs) using another threshold (Thresh2). The H channel is chosen because it offers better separation between RBCs and WBC cytoplasm compared to other channels.
  • The final segmented image is obtained by intersecting the results of these two steps.
  • Crucially, the optimal values for Thresh1 and Thresh2 are determined using a golden section search method, which iteratively optimizes a performance metric (Dice Similarity Coefficient - DSC) to find the best thresholds.
  • Postprocessing: This stage refines the segmented WBCs. It involves binarization, morphological erosion, median filtering to reduce noise and remove small artifacts, and finally, maximum connected region (MCR) extraction to isolate complete WBCs.

Datasets

The proposed method was evaluated using the ALL_IDB (Acute Lymphoblastic Leukemia Image Database). Specifically, the study utilized the ALL_IDB2 subset, which contains 260 images, each featuring a single WBC. The first half of these images are from ALL patients (lymphoblasts), and the second half are from non-ALL patients (normal WBCs). The dataset was captured using an optical microscope and a Canon Power Shot G5 camera. For evaluation, manual segmentations were performed by two expert oncologists to serve as ground truth.

Key Insights and Ideas

  • Complementary Strengths of Color Spaces: The paper highlights that no single color space is perfect for WBC segmentation. RGB is good for general image processing, while HSV's H channel offers better discrimination for certain cellular components like cytoplasm and RBCs. Combining them is key.
  • Dual-Thresholding for Robustness: By using two distinct thresholds, one for background extraction (from grayscale) and another for RBC removal (from HSV's H channel), the method becomes more robust to variations in illumination and the challenging color similarity between cytoplasm and RBCs.
  • Golden Section Search for Optimal Thresholds: Instead of relying on fixed or heuristic thresholds, employing an optimization technique like golden section search to find the optimal threshold values significantly improves segmentation accuracy and consistency.
  • Importance of Postprocessing: Even with advanced segmentation, postprocessing steps like morphological operations and MCR extraction are vital for cleaning up the results and ensuring accurate final segmentation.

Key Results

  • High Segmentation Accuracy: The proposed dual-threshold method achieved an overall single WBC segmentation accuracy of 97.85% (measured by DSC).
  • Superior Performance: The method outperformed two single-threshold-based methods (one using RGB, the other using HSV) in terms of mean DSC. The proposed method achieved a mean DSC of 0.9795, compared to 0.9542 for the RGB-based method and 0.7737 for the HSV-based method.
  • Robustness: The standard deviation of the DSC values for the proposed method was significantly lower (0.0144) than the other two methods (0.0403 and 0.2163), indicating greater robustness across different images and conditions.
  • Qualitative Success: Visual results demonstrated that the proposed method successfully segmented both nucleus and cytoplasm of WBCs, even in challenging images with varying backgrounds and the presence of RBCs, where other methods struggled.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the method performs well, it is primarily tested on images with single WBCs. The performance on images with multiple overlapping WBCs might require further refinement, particularly in the postprocessing stage (e.g., removing MCR if multiple WBCs are present). The selection of initial parameters for the golden section search method was based on prior knowledge, which could potentially be automated or further optimized.
  • Future Work: The authors suggest that this segmentation method shows a "good prospect" for further steps in automated WBC analysis, including feature extraction, ALL classification, and diagnosis. Future work could focus on extending the method to handle multi-WBC images more effectively and integrating it into a complete diagnostic system.

Summary updated

Abstract

We propose a dual-threshold method based on a strategic combination of RGB and HSV color space for white blood cell (WBC) segmentation. The proposed method consists of three main parts: preprocessing, threshold segmentation, and postprocessing. In the preprocessing part, we get two images for further processing: one contrast-stretched gray image and one H component image from transformed HSV color space. In the threshold segmentation part, a dual-threshold method is proposed for improving the conventional single-threshold approaches and a golden section search method is used for determining the optimal thresholds. For the postprocessing part, mathematical morphology and median filtering are utilized to denoise and remove incomplete WBCs. The proposed method was tested in segmenting the lymphoblasts on a public Acute Lymphoblastic Leukemia (ALL) image dataset. The results show that the performance of the proposed method is better than single-threshold approach independently performed in RGB and HSV color space and the overall single WBC segmentation accuracy reaches 97.85%, showing a good prospect in subsequent lymphoblast classification and ALL diagnosis.
BibTeX
@inproceedings{li2016segmentation,
  author = {Yan Li and Rui Zhu and Mi Lei and Yihui Cao and Di Yao},
  title = {{Segmentation of White Blood Cell from Acute Lymphoblastic Leukemia Images Using Dual-Threshold Method}},
  booktitle = {Computational},
  year = {2016},
  doi = {10.1155/2016/9514707},
}
Figure