Analysis of Hematopoietic cells for Leukemia Detection and Classification

Abstract

Cancer has been menacing the human race so much so that in most of the nations, it is a predominant cause of death. The word “Cancer” is a generic term being used for a large group of diseases that can affect any part of the body. Each type of cancer has its own symptoms, rate of growth, treatment, and cure. The possibility of healing a cancer patient depends on at what stage the disease has been detected; early detection increases the chances of survival.
In last two decades, Computer Aided Classification (CAC) Systems have gained quite a popular for their fast, efficient, robust, and reliable models that help in detection of a wide variety of diseases. This research work addresses the main problems associated with hematological analysis, Acute Lymphoblastic Leukemia (ALL) detection in particular. This thesis helps in the development of algorithms for the extraction of useful information for classification of pattern of tissues.
An improved watershed algorithm using marker scheme (IWAMS) is proposed to separate the grouped cells (agglomerates) present in the microscopic images. This framework intends to divide the entire image or tissue from the background by using marker scheme. A shadowed cmeans (SCM) technique is further employed to isolate the tissue structure in different regions or the cells in their parts (nucleus and cytoplasm).
It is witnessed that the textural component of a normal and an affected blood cell varies significantly due to the change of their chromatin distribution. To automatically detect abnormalities and speed up the process of detection of ALL in the Peripheral Blood Smear (PBS), efficient feature extraction methodologies have been proposed in the subsequent chapters to extract features from the nucleus and cytoplasm region. Those features are then fed to the different classifiers for obtaining the accuracy of the model.
A multiresolution analysis based feature extraction approach is developed which uses TwoDimensional Discrete Wavelet Transform. Subsequently the relevant features are
selected using a combination of PCA and Bhattacharyya distance technique. The resultant feature set is of substantially lower dimension. On application of various classifier, it is observed that Back Propagation Neural Network (BPNN) gives better classification accuracy as compared to others.
A Gray Level Cooccurrence Matrix (GLCM) based framework is proposed which is used to differentiate the spatial relationship between pixels of an image. A probabilistic principal component analysis (PPCA) approach is employed to reduce the feature set. Of all classifiers, Random Forest (RF) results in providing greater accuracy.
Another multiresolution ALL detection using TwoDimensional Stationary Wavelet Transform is developed which uses the shift invariant properties to extract the texture features from microscopic images. This scheme allows the use of principal component analysis and linear discriminant analysis for reduction of the most appropriate features. This features with Support Vector Machine (SVM) classifier outperform other classifiers concerning accuracy.
Each model is studied independently, and tests are carried out to evaluate their performances. Each scheme is validated using the available dataset ALLIDB.
Performance measures, i.e., accuracy, sensitivity, and specificity are very promising and are used to compare the efficacy of proposed automated systems with that of standard diagnostic
procedures.