New decision support tool for acute lymphoblastic leukemia classification

Monica Madhukar, Sos S. Agaian, Anthony T. ChronopoulosImage Processing: Algorithms and Systems / Parallel Processing for Imaging Applications 2012 · DOI

TL;DR

This paper introduces a novel decision support tool for classifying Acute Lymphoblastic Leukemia (ALL) from microscopic blood smear images. By employing a robust segmentation of entire images (not just cropped nuclei) and extracting a comprehensive set of shape and texture features, their system achieves high classification accuracy, offering a more automated and potentially more reliable alternative to manual diagnosis.

What Problem Does This Paper Address?

Diagnosing Acute Lymphoblastic Leukemia (ALL), especially in children, is critical for timely and effective treatment. While advanced diagnostic techniques exist, microscopic examination of blood slides remains a standard method. However, this manual process is subjective, prone to operator fatigue, and can lead to diagnostic confusion due to the visual similarity of ALL to other blood disorders. Existing automated systems often struggle with segmenting multiple cells within a single image and tend to focus on cropped images of individual nuclei, which may miss crucial contextual information. This paper aims to bridge this gap by developing a cost-effective, automated system that can accurately classify ALL from complete microscopic blood images, thereby improving diagnostic efficiency and consistency.

Methods

The proposed system follows a four-step pipeline:

  1. Pre-Processing: Images are first pre-processed to reduce background non-uniformities and enhance color information. A key step here is converting the RGB color space to the CIELAB color space. CIELAB is chosen for its perceptual uniformity and reduced color dimensionality, making it more suitable for color-based clustering and segmentation.
  2. Segmentation: The system employs a K-means clustering algorithm on the CIELAB images to segment the cell nuclei. Unlike many previous methods that segment individual nuclei from cropped images, this approach segments nuclei from the entire blood smear image, allowing for the analysis of multiple cells simultaneously. The algorithm identifies clusters corresponding to nuclei (high saturation, blueish hue) and other cellular components.
  3. Feature Extraction: A comprehensive set of features is extracted from the segmented nuclei. These features are categorized into:
  4. Shape Features: Including area, perimeter, compactness, solidity, eccentricity, and elongation. These capture the geometric properties of the nuclei.
  5. Texture Features: Derived from the Gray Level Co-occurrence Matrix (GLCM), such as energy, contrast, entropy, and correlation. These describe the spatial arrangement of pixel intensities within the nuclei.
  6. Fractal Dimension (Hausdorff Dimension): This feature is highlighted as particularly crucial for distinguishing cancerous from non-cancerous cells, as it quantifies the roughness of the nucleus boundary. The paper emphasizes that features are extracted from whole images containing multiple nuclei, addressing a significant limitation of prior work.

  7. Classification: A Support Vector Machine (SVM) classifier is used to distinguish between ALL and normal cells based on the extracted features. The paper discusses various cross-validation techniques (K-fold, Hold-Out, Leave-One-Out) for evaluating the classifier's performance.

Datasets

The method was evaluated using a dataset of 98 JPG images provided by Charles Fabio Scotti from the Università degli Studi di Milano. These images were captured using an optical laboratory microscope coupled with a Canon Power Shot G5 camera, with a resolution of 512x512 pixels and 24-bit color depth. The paper does not explicitly state if this dataset is newly introduced or a publicly available benchmark.

Key Insights and Ideas

  • Whole Image Segmentation is Crucial: The paper strongly advocates for segmenting and analyzing entire blood smear images rather than cropped nuclei. This approach allows for the capture of more contextual information and potentially more robust classification, especially when dealing with multiple cells.
  • CIELAB for Robust Color Analysis: Converting images to the CIELAB color space is a key pre-processing step that helps in handling variations in illumination and staining, leading to more reliable color-based segmentation.
  • Fractal Dimension as a Discriminator: The Hausdorff Dimension is identified as a particularly effective feature for differentiating cancerous from non-cancerous cells, showing a steep difference in box-count between the two cases.
  • Comprehensive Feature Set: The combination of shape, texture, and fractal features provides a rich representation of the cell nuclei, contributing to improved classification accuracy.

Key Results

The proposed system achieved a classification accuracy of approximately 93.5% using Leave-M-out cross-validation. The paper highlights that their system, by implementing on whole images with multiple nuclei, outperforms existing systems that primarily work on sub-images (single nuclei). Figure 9(b) visually compares their performance against several other systems, demonstrating a competitive or superior accuracy.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Training Data: For improved classification accuracy, it is suggested to include only "typical" images in the training set.
  • Classifier Complexity: Alternative training procedures, such as those accounting for the continuous nature of class membership or hierarchical classifiers, could be explored.
  • Training Samples: Increasing the number of training samples is desirable to further reduce overfitting.
  • Feature Selection: Further feature selection can be considered for future improvements.
  • Robustness: The system needs to be made more robust to excessive staining and touching cells.
  • Subtypes: Future work could involve classifying lymphoblasts into various subtypes.

Summary updated

BibTeX
@inproceedings{madhukar2012new,
  author = {Monica Madhukar and Sos S. Agaian and Anthony T. Chronopoulos},
  title = {{New decision support tool for acute lymphoblastic leukemia classification}},
  booktitle = {Image Processing: Algorithms and Systems / Parallel Processing for Imaging Applications},
  year = {2012},
  doi = {10.1117/12.905969},
}
Figure