Automatic White Blood Cell Differential Classification

A Master of Science in Mechatronics Submitted to the School of Engineering by Juma A. Bin Darwish Al-Muhairy, "Automatic White Blood Cell Differential Classification," June 2005. Thesis Advisor Dr. Yousef Al- Assaf. Available are Both Soft and Hard Copies of the Thesis.

Abstract

The characteristics (quantity, shape and color) of the white blood cell (WBC) can give vital information about a patient's health. Hematologists, with the aid of microscopes, use their experience to classify WBCs and make appropriate reporting and recommendations to physicians. Automating the segmentation and classification of WBC could provide a useful tool in medical diagnoses. In this work, computer-based segmentation and classification of the four main classes of WBC (Neutrophils, Eosinophils, Lymphocytes, and Monocytes) were completed. Soft computing algorithms including neural network (NN) and polynomial classifiers (PC) were used for WBC classification, while watershed and thresholding based on size, shape, color and texture characteristics were used to segment WBC from Red Blood Cells RBC, platelets, cell fragments and stains. Furthermore, characteristics of the WBC were utilized in association with the intelligent systems to classify these WBC's to different classes. The number and distribution of different classes of WBC has medical indications (e.g. high Neutrophils count may possibly imply cancer, whereas high Lymphocytes count could lead to AIDS). To classify WBC morphological based features, Discrete Cosine Transform (DCT) based features and Discrete Wavelets Transform (DWT) based features were used as input feature vectors to the Neural Networks NN and Polynomial Classifiers PC. Various iv feature extraction modalities and classifiers were tested on blood data obtained from patients at Twam Hospital in UAE. Combining color morphological and DWT features in association with the PC second order classifier achieved a Classification Accuracy (CA) of 99.3%. Other advantages of using PC was that it needed less computational requirements and classification was independent on various user-set parameters as it was the case of NN. In addition, PC proven to be more reliable and consistent in terms of results compared to NN.