1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Ensemble learning model pruning based on bees algorithm for breast cancer classification
Authors:	Ashwaq Mukred Saeed Qasem (P77884)
Supervisor:	Siti Norul Huda Sheikh Abdullah, Assoc. Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Computer Aided Diagnostic (CADx) ]Mammogram images Diagnostic imaging
Issue Date:	20-Feb-2020
Description:	Computer Aided Diagnostic (CADx) framework is able to assist radiologists to achieve better diagnosis for mammogram images of breast cancer. As reliable strategy, ensemble learning has shown better performance in CADx. Ensemble learning consists of two stages: multiple based-classifier generation and aggregation subsequently. In dealing with immense ensemble subset combinations, ensemble pruning can select the best combination of ensemble member from a set of generated-classifiers. Bees Algorithm (BA) is a popular swarm-based optimisation method that can exploit and explore in high dimensional search space with regards to attain the best solution. Therefore, the first objective of this thesis is to enhance ensemble pruning using Bees algorithm (BA) for breast cancer CADx framework (BA_Ens). A pool of different classifier models such as support vector machine, k-nearest neighbour and linear discriminant analysis classifiers have been trained using 12 groups of multi-features. Then, BA is used to exploit and explore random uniform combination subsets of the trained-classifiers set from a pool of generated classifiers fused by Majority Voting (MV) at the aggregation stage. Consequently, the best subset is selected as the optimal ensemble. On the other hand, the initial random locations of foragers are relatively inconsistent affect the quality of the resource reached to the optimal ensemble solution. On top of that, an implication of MV based ensemble fusion assigning equal weights to all ensemble members may lead to lower degree of voting power among expert members. Therefore, the second objective of this thesis is to enhance both the initialization and local search processes in the BA_Ens by introducing a selective Random Start Best step (RSB) initialization and Weighted MV based on BA Local search of (L-WMV). RSB initialization method starts by selecting the first member of the ensemble randomly while the remaining members are selected using a greedy search process. In L-WMV, adaptive weights are assigned to the ensemble members during local search process in BA. The most successful classifiers are assigned more weight to boost up the ensemble learning performance. In addition, diversity is an important factor in ensemble pruning process. The state of the art of ensemble diversity measurements such as Correlation (C), Q-statistic (Q) and Double Fault (DF) endure more computation time to achieve the mature ensemble. Hence this thesis proposes the third objective by enhancing ensemble pruning based on RSB(L-AMV) using information theory-based method. Here, Mutual Information (MI) is adopted in the ensemble pruning process with the attention for addressing a simpler ensemble pruning CADx framework based on RSB(L-AMV). The self-collection mammogram image dataset from Hospital Kuala Lumpur (HKL) consists of 263 breast masses (benign and malignant), are used to evaluate the proposed ensemble pruning model based on three proposed methods. As a result, proposed BA_Ens method gained about 76 %, 81%, 92% and 60% of Area Under Curve (AUC), accuracy, specificity and sensitivity subsequently. While, RSB(L-WMV) obtained 77%, 82%, 91% and 64% of AUC, accuracy, specificity and sensitivity respectively. The proposed MI- RSB(L-AMV) achieved a comparable performance as compared to RSB(L-AMV). Consequently, BA based ensemble pruning accomplished the optimal ensemble solution with less number of generations after applying MI method. In general, all three proposed methods have shown significant performance.,Ph.D
Pages:	226
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.