Hematological image analysis for acute lymphoblastic leukemia detection and classification

Abstract

Microscopic analysis of peripheral blood smear is a critical step in detection of leukemia.However, this type of light microscopic assessment is time consuming, inherently subjective, and is governed by hematopathologists clinical acumen and experience. To
circumvent such problems, an efficient computer aided methodology for quantitative analysis of peripheral blood samples is required to be developed. In this thesis, efforts are therefore made to devise methodologies for automated detection and subclassification of Acute Lymphoblastic Leukemia (ALL) using image processing and machine learning methods.Choice of appropriate segmentation scheme plays a vital role in the automated disease recognition process. Accordingly to segment the normal mature lymphocyte and malignant lymphoblast images into constituent morphological regions novel schemes have been proposed. In order to make the proposed schemes viable from a practical and real–time stand point, the segmentation problem is addressed in both supervised and unsupervised framework. These proposed methods are based on neural network,feature space clustering, and Markov random field modeling, where the segmentation problem is formulated as pixel classification, pixel clustering, and pixel labeling
problem respectively. A comprehensive validation analysis is presented to evaluate the performance of four proposed lymphocyte image segmentation schemes against manual
segmentation results provided by a panel of hematopathologists. It is observed that morphological components of normal and malignant lymphocytes differ significantly. To automatically recognize lymphoblasts and detect ALL in peripheral blood samples, an efficient methodology is proposed.Morphological, textural and color features are extracted from the segmented nucleus and cytoplasm regions of the lymphocyte images. An ensemble of classifiers represented as EOC3 comprising of three classifiers shows highest classification accuracy of 94.73% in comparison to individual members. The subclassification of ALL based on French–American–British (FAB) and World
Health Organization (WHO) criteria is essential for prognosis and treatment planning. Accordingly two independent methodologies are proposed for automated classification of malignant lymphocyte (lymphoblast) images based on morphology and phenotype. These methods include lymphoblast image segmentation, nucleus and cytoplasm feature extraction, and efficient classification.