Theses & Dissertations

Permanent URI for this collection

https://hdl.handle.net/20.500.14170/209

Browse

Browsing Theses & Dissertations by Subject "Acute myeloid leukemia (AML)"
  • Publication
    Automated classification procedure for acute Myeloid Leukemia cells based on bone marrow samples
    ( 2019)
    Lim Huey Nee
    Leukemia is a cancer of blood that is the most common cancer among children between the ages of 0 to 13 years old. The development of acute leukemia progress rapidly and uncontrollably if the diagnosis and treatment is delayed. The conventional leukemia diagnosis begins with screening process on peripheral blood followed by bone marrow test. The test includes genetic and molecular tests which are expensive and time consuming. Moreover, the previous works were mostly focused on peripheral blood samples and studies on screening and diagnosis, i.e. the recognition between normal and abnormal cases but not the leukemia subtypes classification tasks. This study involves in the development of automated classification procedure for acute myeloid leukemia (AML) based on bone marrow samples. The challenge in this study is to find the best automated procedure from image segmentation to classification that can classify the subtypes of acute myeloid leukemia. Automated image segmentation technique based on the combination of multilayer k-mean clustering algorithm, watershed segmentation and seeded region growing algorithm has been proposed. A new concept of multilayer clustering procedure is proposed and presented to implement the idea of automated image segmentation process. The combination of multilayer clustering, watershed and seeded region growing method successfully eliminates background, red blood cells and unwanted cells while retaining the cells of interest. The proposed image segmentation method achieves the highest specificity of 98.17%, highest sensitivity of 99.40% and highest accuracy of 98.61%, where it is the highest percentage as compared to the conventional segmentation method. The study also includes feature extraction where 92 features are extracted in classifying AML subtypes. Four feature categories are considered namely simple shape descriptors, colour, texture, and Gabor features. Besides that, two feature selection methods, Filter-based Feature Selection (FFS) and Bagged Decision Tree (BDT) are chosen to select prominent features and reduce feature dimensionality. Two classification methods are tested, namely multiclass classification and hierarchical classification method. The extracted features and classification methods are fed into three classifiers which are Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Decision Tree (DT). The classification performance is based on the accuracy of training, validation and testing data. MLP has shown to yield good classification and generalization performance as it achieves accuracies of 98.56%, 99.55% and 97.78% for training, validation and testing data respectively. Hierarchical MLP classifier is used in feature selection process. 30 features are selected using FFS method while 48 features are selected using BDT method. Colour features have proven to be prominent and important as six of the colour features are among the top ten selected features. MLP with BDT feature set achieved the highest accuracy of 98.21%, 98.22% and 99.00% for training, validation and testing data respectively. Overall, the performance comparison results indicate that MLP implemented with BDT feature set is the best option for AML subtypes classification.
      1  6