Statistical Analysis of Spect Neuroimages in Alzheimer's Disease

McCrory, Stephen James (1992) Statistical Analysis of Spect Neuroimages in Alzheimer's Disease. MSc(R) thesis, University of Glasgow.

Full text available as:
[thumbnail of 13834190.pdf] PDF
Download (10MB)


We have considered a variety of problems in the statistical analysis of quantitative data extracted from Single Photon Emission Computed Tomography (SPECT) and Positron Emission Tomography (PET) neuroimages. Many of the questions of interest to investigators in this area of research were described and illustrated with analysis of some SPECT datasets. In chapter one we introduced the technique of tomograhic imaging, describing a common approach to extracting quantitative data from images and gave some background to Alzheimer's disease. In chapter two we described a major research study into Alzheimer's disease, from which most of the datasets used in this thesis were obtained. In chapter three we identified the broad categories of statistical problems into within and between group analysis of regional mean patterns and the interrelationships among regions. A SPECT dataset, consisting of Alzheimer and normal control subjects, was used to illustrate the use of univariate methods to study these problems. We saw from these analyses that it was difficult to extract clear biological interpretations with this approach due to certain features in the data extracted from the images. In particular, the presence of substantial random variation between subject data vectors meant that meaningful analysis could only be carried out after adjusting the regional data - to remove the between subject variation - prior to the analysis. Different methods of adjustment were seen to give different results here. Although not particularly evident from these data, another feature of typical imaging datasets was the large number of regions to be analysed. In chapter four we looked at the application of univariate and multivariate ANOVA type methods to compare regional mean profiles between groups and illustrated some approaches to follow-up analysis. Assumptions underlying these techniques, including normality and equality of covariance matrices were assessed as was the choice of scale for the analysis. The assumption of multivariate normality was reasonable on the square root scale in both groups, although equality of group covariance matrices was very strongly rejected. Even though many of the assumptions in the RM ANOVA may be violated for these datasets, the fact that global tests can be performed, even when the number of regions p exceeds the numbers of subject n, will make this the most viable approach. Adjusted F-tests will be appropriate in such circumstances. In chapter five, we looked at some approaches to investigating inter-relationships among regions. The most common approach here is to use simple correlation analysis among regions after adjusting data vectors for the subject effect. As in the analysis of means, the results will be strongly influenced by the form of the adjustment. Ford (1986) has shown that inferences from the results of correlation analysis are made difficult; with adjustment in the data resulting in confounding of parameters in a model of the correlation structure. Even so, between group comparisons may still provide valuable insight into a disease process. A testing scheme gave tentative evidence of differences in the correlation structure between the normal and Alzheimer groups. Multivariate exploratory techniques where used to study the interregional correlation structures. Principal components analysis demonstrated that just a few patterns accounted for most of the variation among subjects in each of the groups and that bilateral pairs of regions were very strongly correlated. Further canonical correlation analysis of the data suggested that regional profiles may be summarised into hemispheric sums and differences separately without too much loss of information. In studies with several regions being studied this would be a useful reduction of dimensions. Multi-Dimensional scaling highlighted a number of other features in our data including the measurement difficulties with smallish regions such as the basal- ganglia and some evidence of a spatial relationship between regional data. Formal analysis of the covariance structures using the spatial correlation model of Worsley et. al. (1992) gave some evidence for this feature in SPECT data, albeit using an estimated distance matrix.

Item Type: Thesis (MSc(R))
Qualification Level: Masters
Additional Information: Adviser: Ian Ford
Keywords: Statistics, Medical imaging
Date of Award: 1992
Depositing User: Enlighten Team
Unique ID: glathesis:1992-76410
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 19 Nov 2019 14:43
Last Modified: 19 Nov 2019 14:43

Actions (login required)

View Item View Item


Downloads per month over past year