dc.contributor.author	Lim, Jia Qi
dc.date.accessioned	2023-04-16T03:41:23Z
dc.date.available	2023-04-16T03:41:23Z
dc.date.issued	2020
dc.description	Thesis (Ph.D (Mathematics))
dc.description.abstract	Early detection of brain tumor is crucial for successful treatment and could increase the chances of patients’ survival. Magnetic resonance imaging (MRI) is a popular tool for diagnosis because it is non-invasive, and the images produced are of high resolution and high contrast between tissues. In view of this, computer aided diagnosis (CAD) system based on data-driven predictive models with MRI scans as inputs is an excellent paradigm to automate the diagnosis process and resolve the variation in opinions among radiologists and clinical experts. This research proposes a complete pathway incorporating detection, segmentation and characterization of brain MRI. Firstly, two-dimensional (2D) brain MRI scans are identified as either normal or tumorous. This objective is achieved by a sequence of steps: image preprocessing, feature extraction, feature reduction, classification models and classifiers performance evaluation. Extreme learning machine (ELM) with 150 latent variables trained with 55 linear discriminant analysis (LDA) features show the highest accuracy when tested with data from different sources. Furthermore, interpretable rule-based classifier using gray level run length matrix (GLRLM) and local binary pattern (LBP) features are constructed for this purpose. Secondly, brain tumor segmentation is conducted by unsupervised clustering-based image segmentation approaches such as k-means clustering, Gaussian mixture model (GMM), and fuzzy C-means (FCM). The results of this study show that clustering based segmentation approaches based solely on single grayscale pixel intensity are not robust and effective in brain tumor segmentation. Thirdly, manually segregated tumor region is classified to be pituitary tumor, meningioma or glioma. The research framework in discriminating the types of tumor is similar to procedures in binary classification task. ELM model with 300 latent variables trained with four LDA attributes achieves the best generalization performance compared to other multi-class predictive models. This shows the discriminative power and relevance of the extracted LDA features. This study is concluded with the introduction of a novel CAD system as outlined above.
dc.description.sponsorship	Faculty of Science
dc.identifier.uri	http://openscience.utm.my/handle/123456789/164
dc.language.iso	en
dc.publisher	Universiti Teknologi Malaysia
dc.subject	Tumors
dc.subject	Magnetic resonance imaging
dc.title	Predictive models for brain tumor classification and segmentation in magnetic resonance imaging and their performance assessment analyses
dc.type	Thesis
dc.type	Dataset

Predictive models for brain tumor classification and segmentation in magnetic resonance imaging and their performance assessment analyses

Files

Original bundle

License bundle

Collections