The aim of this research is to develop a computer-aided detection system for prostate cancer within the peripheral zone, based on single modality T2-W Magnetic Resonance Imaging. As the most diagnosed and second leading cause of death from cancer in men, prostate cancer is a significant health problem globally. In fact, considering the deficiencies in current clinical screening methods, there is a need for a rapid development of MRI technologies and computer algorithms to better detect prostate cancer. In this thesis, we developed one unsupervised, and two supervised computer algorithms, using different texture descriptors involving different resolutions, filters, techniques and orientations. For classification purposes, we investigated 11 machine learning algorithms to build predictive models. This thesis also investigated the effects of window sizes for all performance metrics. In the proposed unsupervised method a small number of texture descriptors were used to differentiate benign and malignant regions. Subsequently, the fuzzy c-means clustering algorithm was employed to segment malignant regions. The resulting binary segmentations were then combined to find overlapping regions with the highest probability of being malignant. In contrast, the two supervised methods were based on 215 texture descriptors and textons. Both methods used the same machine learning algorithms in the training and classification phases. To evaluate the performance of the proposed methods, the unsupervised and supervised methods were tested based on 37 (275 MRI images) and 45 (418 MRI images) patients, respectively. The performance evaluations showed that the results of all methods were comparable with the state-of-the-art in the literature in terms of area under the curve, classification accuracy, sensitivity and specificity. The contributions of this thesis are three-fold. First, we developed novel supervised and unsupervised computer-aided detection systems for prostate cancer, which are different from those produced hitherto, in the sense of our methods exploit different combinations of texture descriptors and machine learning algorithms. In fact, this thesis is the first study which has thoroughly investigated textons in prostate cancer detection. Secondly, this thesis made an extensive investigation into the effects of window size, both on the methods and feature's performance itself. Thirdly, we provide an extensive quantitative comparison of the performances of 11 machine learning algorithms and a thorough qualitative comparison between our results and the state-of-the-art in the literature.
Thesis, 47 MB, PDF
Show more files.. Show less files..
Thesis, 47 MB, PDF
|This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License|