MULTICLASS CLASSIFICATION OF WHITE BLOOD CELLS FROM HISTOLOGICAL IMAGES USING DEEP CONVOLUTIONAL NEURAL NETWORKS

Abstract

The white blood cell (WBC), also called leukocytes, is a cellular component of the blood with a nucleus and without a haemoglobin. As an essential part of the immune system, it moves from blood to tissue and provide defence for fighting against the invasion of the foreign microorganisms, e.g., bacteria, viruses, and germs, by ingesting them, destroying infectious agents or by producing antibodies. The leukocyte can be categorized into five genres: Eosinophils, lymphocytes, Neutrophils, Monocytes and Basophils. Neutrophils are the most abundant, and they are responsible for defending the bacteria or fungal infection. Eosinophils occupy around 2–4% of WBC, and act in response to allergies and parasite infection. Lymphocytes undertake the task of the specific recognition of foreign agents and the consequent removal from the host. Monocytes are effective in direct destruction of pathogens and cleanup of the debris from the infection sites. The counter of different white blood cells plays a significant role in the clinical diagnosis and test: it is an indicator that reflects the hidden infection within the body and alerts the hematologists as a signal, i.e., the abnormal increase in WBC is the so-called leukocytosis. It also helps doctors monitor the effectiveness of chemotherapy or radiation treatment in people with cancer. The detection and distinguishment of diverse WBC and the further counting of the corresponding proportion is critical due to the richness of clinical meaning behind it. It is painstaking and of low efficiency if we manually differentiate the leukocytes under the microscopes, from which the automatic classification based on the images of WBC emerges.

Usually, the automatic classification approaches are present with several main steps: preprocessing, segmentation, feature extraction and classification. The preprocessing procedure primarily refers to the attempt of removing the noises or some artifacts from the images to output the contrast images. The segmentation can be considered as the operation of segmenting the WBC from the background of the smear images or extracting the region of interest (ROI). The consequent step is to build a representative feature vector for every type of WBC, and the classification will work based on it. In this very step, the hematologist sometimes may be involved to determine the features. However, the traditional classification approaches consume more time with the compromise in accuracy too.

Recent years, the emerging field of deep learning has powered many successful real-life applications. Deep neural networks, particularly convolutional neural networks (CNNs), have been widely applied to perform computer vision tasks such as image classification. Compared to machine learning algorithms, which use hand-crafted features as inputs, CNNs typically take raw images as inputs and learn hierarchical feature representations in an end-to-end fashion. Therefore, this project aimed to implement the detection of subtype blood cells using the advancement of neural networks known as deep learning CNN.