An Automated Malignant Tumour Localization Algorithm for Prostate Cancer Detection in Trans-rectal Ultrasound Images
The goal of this thesis is to design, implement and evaluate an automated algorithm to detect cancerous tissues and segment the malignant tumour in ultrasound images of the prostate. To accomplish this goal, first, the important image features which would lead to the optimal segmentation are identified. This work focuses on the local texture feature and spatial features. Various approaches to extract the local texture feature are explored, including grey-level co-occurrence matrix (GLCM), recurrent random-pulsed neural networks (RNN), and a novel wavelet-based filter. The spatial features are represented using conventional one dimensional fuzzy membership functions and novel multi-dimensional fuzzy membership functions. The texture and spatial features are combined using the fuzzy inference system. Two of the techniques investigated in this thesis could potentially constitute the basis for key paradigm shifts in medical imaging research. One of these is the idea that medical images in general, and ultrasound images in particular, contain information which are hidden from medical professionals due to limitations in the human visual system. This thesis shows that this information could be extracted using a computerized approach by separating the deterministic components in the image from the indeterministic components, or noise. The other idea concerns the representation of multidimensional statistical distribution information with fuzzy membership functions with the corresponding dimensions. This thesis shows that increasing the number of dimensions with which to represent the statistical distributions results in a more accurate mapping of information that relates to human anatomy, which is essentially 3D in nature. In the thesis, the natures of the various techniques are explored by testing on synthesized images. Then, these approaches are adapted to the ultrasonic prostate cancer segmentation problem and are evaluated with trans-rectal ultrasound images (TRUS). The segmentation using only texture features yields results with high sensitivity. When the spatial features are incorporated using the fuzzy inference system, the specificity of the diagnosis improves dramatically and the overall classification accuracy is also increased. Clinically, this automated diagnostic system could be used as a decision support tool for radiologists when identifying suspicious regions in the prostate from which to draw biopsy samples. The proposed system improves the consistency of the cancer detection process and could provide savings in both time and cost in the prevention and treatment of prostate cancer.