Prognosis procedure for Acute Leukemia based on blood samples using artificial intelligent techniques

Leukemia is a cancer of blood affecting white blood cells and most commonly pediatric patients between the ages of 0 to 13 years old. Treatment protocols for leukemia disorders require testing procedures that can quantify tumor burden before and after therapy to assess responsiveness to therapy. Currently, genetic tests techniques such as molecular test, flow cytometry, cytogenetic analysis and Polymerase Chain Reaction (PCR) are also employed for specific leukemia detection after treatment. The need for automation of leukemia detection arises since the above specific tests are time consuming and costly. This study involves the development of prognosis procedure for acute leukemia based on blood samples that can be used for the classification of Acute Lymphoblastic Leukemia (ALL) blood sub-types after treatment. Segmentation based on the combination of image enhancement, color thresholding, geometric features, seed region growing and watershed algorithm has been proposed. There are five proposed methods for fixed color thresholding as well as Otsu thresholding and comparison segmentation performance between RGB, HSV and LAB color space are done. Fixed color thresholding based on Hue Component has proven to be the best in obtaining a fully segmented acute leukemia regions with accuracy of 99.28%, specificity of 99.71% and sensitivity of 88.69%. The study also includes feature extraction where 25 features are extracted in classifying ALL sub-types. Three feature categories are considered namely simple shape descriptors, color, and texture features. Besides that, two feature selection methods, Neighborhood Connected Component (NCA) and ReliefF are chosen to select prominent features and reduce feature dimensionality, resulting in the total of 16 selected features by NCA and 16 and 22 selected features by ReliefF. Two classification methods are tested which are multiclass classification and hierarchical classification method. An approach to select the optimal features was proposed using SLFN trained by ELM due to its simplicity and fasttraining. NCA method using 16 features are selected as input data for Multilayer Perceptron (MLP) and Support Vector Machine (SVM). The classification performance is based on the accuracy of the training and testing data. Overall, the performance comparison results indicate that MLP in hierarchical classification is the best option for ALL subtypes classification with testing accuracy of 99.79% compare to multiclass classification which is 99.09%.