The Best Texture Features for Leukocytes Recognition

Area of Science:

• Medical Imaging
• Computational Pathology
• Biomedical Engineering

Background:

• Accurate white blood cell (WBC) or leukocyte differential counting is crucial for diagnosing diseases like leukemia and infections.
• Manual microscopic evaluation of blood smears is time-consuming and labor-intensive, necessitating automated solutions.
• Leukocyte recognition is a critical component of automated Computer-Aided Design (CAD) systems for cell counting.

Purpose of the Study:

• To identify optimal texture features for recognizing five types of leukocytes (Monocyte, Lymphocyte, Neutrophil, Eosinophil, Basophil).
• To evaluate the performance of different feature categories and classifiers in an automated cell counting system.
• To establish a foundation for future systems that maximize accuracy for individual leukocyte types.

Main Methods:

• Analysis of seven texture feature categories: GLCM, Haralick, Spectral, Wavelet-based, Gabor-based, CoALBP, and RICLBP.
• Feature selection using stepwise regression to identify the most effective features for leukocyte detection.
• Classification using three established algorithms: K-Nearest Neighbors (K-NN), Linear Discriminant Analysis (LDA), and Naive Bayes (NB).

Main Results:

• The proposed system was evaluated on a dataset of 200 cell images.
• RICLBP features, when used with the LDA classifier, achieved the highest mean accuracy of 85.53%.
• Individual leukocyte type accuracies varied, indicating room for improvement beyond overall mean accuracy.

Conclusions:

• While RICLBP features with LDA yielded the best overall mean accuracy, they did not maximize the accuracy for all individual leukocyte types.
• Future research should focus on developing systems that integrate multiple features and multiple classifiers.
• The goal of future systems is to enhance accuracy for each specific cell type, leading to more robust diagnostic tools.