Abstract
Alzheimer's Disease (AD) is a major concern for the elderly population, currently affecting over 670,000 people in the UK. With the continual increase in the age of the population the problem is expected to rise. There is no known cure to the condition and a definite diagnosis cannot be made in life. Clinical diagnosis is considered to be approximately 80% - 90% accurate, sometimes taking up to a year to assess. Early detection could aid in the care and possible development of better treatments or even a cure. AD has been shown to alter the structure and global texture of the brain. Studies using Magnetic Resonance imaging (MRI) and Computerised Tomography (CT) have been used to detect these changes with some success by some researchers. Positron Emission Tomography (PET) imaging is a functional imaging modality and in theory before structural changes are evident functional changes should be apparent. Therefore we utilise PET images for this study. This thesis will exploit the fact that AD alters the global texture of the brain. Texture features extracted from fluoro-deoxy-glucose (FDG) PET images and sinograms of the brain will be used. Most texture feature extraction methods fail, due to poor signal to noise ratio so we will use a novel texture feature extraction method known as the Trace transform - triple features, which can extract features directly from raw data acquired by PET scanners. Classifiers will be used to aid in the separation of the two groups, namely AD patients and normal controls. The Trace transform - triple feature method has proven its potential as a good feature extraction technique. It enabled us to achieve classification accuracy of up to 93% on raw sinogram data using a combination of five features. This result is very good compared with the clinical accuracy of 80% reported by most researcher. It is comparable to results obtained by Kippenhan et al [52, 53, 51, 50], who used regional metabolic activity using PET and a neural network classifier. Monomial features extracted from images achieved accuracies as high as 87%. These features are good discriminators, however, they suffer from lack of scaling invariance. This is problematic as brain sizes do vary considerably. The use of registration and extraction of regional information failed to produce fruitful results. This is principally due to poor registration. The registration failed primarily because a very small cross section of the brain was available. Also the effect of AD alters the structure of the brain. Since the registration relies on matching structure, it becomes questionable whether one can actually register automatically a very degraded AD brain. Gender and age are crucial to the progress of Alzheimer's disease. Age and gender matching is not sufficient to get the best results. This thesis has shown that performance gains of up to 11% can be attained by simply incorporating age and/or gender into the classification model. However, the maximum classification accuracy was not improved any further.