White Blood Cells Segmentation and Classification Using Swarm Optimization Algorithms and Multilayer Perceptron

Introduction

Blood consists of white blood cells that fight infection, red blood cells that carry oxygen to the tissues and platelets which help blood to clot. Any change in the number of blood cells indicates blood disease as the number and shape of white cells indicates diseases. The decrease in the number of white blood cells makes people more susceptible to infections and the increase in these cells is caused by cancers of the bone marrow (Mohamed et al, 2019).

White blood cells consist of five categories neutrophil, monocyte, lymphocyte, basophil and eosinophil (Agaian et al., 2018), that defend the body against diseases. Neutrophil, eosinophil and basophil are types of granulocytes that contain enzymes that destroy bacteria, but lymphocyte and monocytes are agranulocytes that produce antibodies and destroy foreign cells found in infectious mononucleosis.

Monocytes have a longer lifespan than many white blood cells and help to break down bacteria. Lymphocytes create antibodies to defend against bacteria, viruses, and other potentially harmful invaders.

Neutrophils that kill and digest bacteria, are considered the most numerous type of white blood cells and your first line of defense when an infection strikes. Basophils appear to sound an alarm when infectious agents invade your blood and consider the marker of allergic disease that help control the body's immune response. Eosinophils attack and kill parasites, destroy cancer cells, and help with allergic responses (AL-Dulaimi et al., 2018).

Leukemia or cancer develops due to the problems with blood cell production and affects white blood cells. The cancer in the blood is considered a deadly disease (Mohamed et al., 2018), there are four types of leukemia, according to French American British (FAB); chronic leukemia, which spreads slowly and causes the symptoms after years. It is sub-classified into Chronic Lymphocytic Leukemia (CLL) and Chronic Myelogenous Leukemia (CML) and acute leukemia, which spreads very fast and the patient feel sick immediately. It is sub-classified into Acute Myelogenous Leukemia (AML) and Acute Lymphoblastic Leukemia (ALL) (Agaian et al., 2018; Sahlol et al., 2019).

The automated identification (Walczak, 2016) of white blood cell types is vital as manual data analysis consumes time and is not always accurate and leukemia detection is a major issue for diagnosis and treatment as the fast and accurate specification of the leukemia type helps in providing the suitable therapy for each type.

Joshi et al. (Joshi et al., 2013) proposed an automatic segmentation technique using Otsu threading method to segment the leukocytes from the blood microscopic images, then they extract the area, perimeter and circularity features to classify the lymphocyte into infected or normal cells using K-Nearest Neighbor (KNN). Their experiment was done using 108 images from ALL-IDB dataset and gave overall accuracy 93%.

Nee et al. (Nee et al., 2012) proposed a segmentation method using erosion and dilation morphological operations to mark leukocytes and blast cells and Otsu's threshold for segmentation and the watershed algorithm to separate the connected cell. The proposed method shows average segmentation accuracy 94.5% for leukemia subtypes M2, M5, and M6.

Nazlibilek et al. (Nazlibilek et al., 2014) presented a system for counting and classifying the white blood cells into their five types: basophil, lymphocyte, neutrophil, monocyte and eosinophil. They used the Otsu's method to extract the white blood cells from the images, then used the mathematical operations and MATLAB functions for counting, labelling the cells and placed them in sub matrices to prepare for classification. They used Principal Component Analysis (PCA) to reduce the extracted feature for classification using Multilayer Perceptron (MLP). They tried the MLP with different number of inputs and hidden layers and compared between them and showed that the classification accuracy rate increased to 95% after using PCA in feature reduction.

Wang et al. (Wang et al., 2016) used spacial and spectral features to classify the white blood cells to their categories using mathematical morphology (spatial analysis) and SVM (spectral analysis). The experiment showed that the spatial features aren't accurate enough for abnormal WBCs, so they used SVM to train the spectral features and spatial features for better classification accuracy, the classification accuracy was more than 90%.