| dc.description.abstract | The capability to screen for leukemia based on bone marrow samples could facilitate the doctors in confirming the occurrence of leukemia from blood test. However, the images of the bone marrow slide have several drawbacks such as the appearance of unwanted regions and lack of contrast. The acquired images of the bone marrow slide could be better improved if these drawbacks are reduced. Due to these matters, a digital image processing system with classification capability is built up in this research which aims to reduce the drawbacks arise from manual screening of bone marrow slide. In this research, two enhancement techniques were used to improve the appearance of the acquired bone marrow slide images. These techniques include partial contrast stretching (PCS) and dark contrast stretching (DCS). Although both techniques produced good results, DCS has been chosen to be utilized in this research due to several reliable reasons. The elimination of unwanted regions leaving only the white blood cells (WBCs) in the bone marrow slide images is initiated by using the saturation component of the HSI (Hue, Saturation, Intensity) color space. Some noises (unwanted small particles) that still appeared were removed with median filter. Simultaneously, median filter fills „holes‟ (white color pixel) which are enclosed within the WBCs. Subsequently, seed-based region growing (SBRG) algorithm is used to remove unwanted regions based on predefined criteria and at the same time extract the area of WBCs in the bone marrow slide images. SBRG technique is capable to maintain the original size and shape of the WBCs in the image. Unlike in blood, the cells present in bone marrow are packed and often overlapped with each other. The watershed algorithm is used to separate the overlapped white blood cells in the bone marrow slide images. Besides area, several other geometrical features were also extracted from the WBCs include circularity, radius and perimeter. Other color features include mean, standard deviation and variance were also extracted for red, green and blue color respectively. These features were used as input data to the MLP network (standard and hierarchical MLP) and k-NN to be classified as Normal, Abnormal type M3 and Other Abnormal bone marrow slide images. Levenberg Marquardt (LM) and Bayesian Regularization (BR) training algorithms were used to train the MLP networks. Standard MLP, network (1A)-BR has managed to achieve the highest accuracy, which is 93.9% on testing dataset in classification of bone marrow slide images into Normal, Abnormal (M3) and Other Abnormal, outperformed MLP network (1A)-LM and k-NN classifier (2A). Hierarchical classifier, MLP network (1B) and (1C) has managed to achieve an average accuracy of 99.8% on training and 98.57% in testing outperformed the k-NN (2B) and (2C). In general, MLP networks have outperformed the KNN classifiers in the classification tasks of the bone marrow slide images. | en_US |
