Some Problems in the Statistical Analysis of Neuroimages

McCormack, Alan Graham (1990) Some Problems in the Statistical Analysis of Neuroimages. PhD thesis, University of Glasgow.

Full text available as:
[img]
Preview
PDF
Download (16MB) | Preview

Abstract

This thesis discusses some of the problems which arise in the statistical analysis of neuroimages. In particular, we develop and evaluate statistical methods which, we hope, will provide greater and more reliable insight into the biological processes which are illustrated in images generated by positron emission tomography (PET), single photon emission computerized tomography (SPECT) and quantitative autoradiography. In chapter one, a mathematical model is developed to characterise the kinetics of the drug MK-801 in normal and ischaemic tissue and to explore the use of radiolabelled MK-801 as an in vivo ligand for studying glutamenergic mechanisms. Using a three compartment model and assuming negligible dissociation from the specific receptor site, kinetic constants are found to be numerically identifiable in four of the nine brain regions in ishaemic tissue (frontal parietal cortex, frontal cortex, occipital cortex and striatum). Convergence to a unique set of parameters is not obtained for normal central nervous system tissue using this model. However, in all ischaemic and normal tissue a two compartment model can be fitted to the data. Thus, in pathological states in which extracellular concentrations of glutamate are elevated and levels of cerebral blood flow are reduced (e.g. ishaemia) during the period of measurement, it would appear that MK-801 has some potential as an in vivo ligand for imaging glutamate release. In chapter two, we address the problems of ranking the response to a drug over a set of brain regions and comparing the patterns of response between drugs or treatments. In the first instance a theoretical approach is taken for the case of ranking three brain regions. We aim to identify the covariance structure for maximising and minimising the probability of a correct ranking, assuming multivariate normality. For higher dimensions, the probability of correctly ranking the observation vector is investigated using the Bonferroni inequality. Due to the complex nature of the response vector, these theoretical approaches are seen to have severe limitations. As an alternative approach, we have investigated, empirically, the performance of a simple measure to characterise the response to drug treatment over a large number of brain regions. In a simulation study, we establish that, within the set of covariance matrices studied, fairly reliable measures of association can be computed. Moreover, doubling the between animal variance component or increasing the within animal variability appeared to have little effect on the reliability of the derived rankings. In chapter three, four univariate repeated measures ANOVA techniques (the traditional F-test, the Huynh-Feldt and Greenhouse-Geisser adjusted tests, and a conservative test based on the lower bound of Box's correction factor) are studied. This form of analysis is important in studying, for instance, the relationship between local cerebral blood flow and local cerebral glucose utilisation. A common feature of the data from these experiments is the high dimensionality of the observation vector on a, relatively speaking, small number of experimental units. Ten multiple comparison procedures are also studied within the same framework. Four of these methods (the Tukey, Scheffe, Bonferroni and Sidak procedures) are constructed under the assumption that the covariance matrix displayed sphericity. Of the remaining procedures, two Scheffe-type pairwise intervals, based on the Greenhouse-Geisser and Huynh-Feldt correction factors, take account of departures from sphericity by adjusting the degrees of freedom of the ANOVA F distribution. Four other methods (Bonferroni, Sidak, Greenhouse-Geisser and Hunyh-Feldt adjusted intervals) are based on the specific estimated variance for each contrast rather than on a pooled estimate based on the assumption of sphericity. In the hypothesis testing situation, the test based on the Huynh- Feldt correction factor has a significance level close to the nominal 5% level over a wide range of covariance matrices. However, if a conservative test is necessary, the Greenhouse-Geisser approach is preferable. Of the multiple comparison procedures, the Bonferroni and Sidak approaches, based on a specific error term for each contrast perform consistently well, giving joint confidence levels close to the nominal 95% level. In chapter four, we examine the spatial patterns in 'control - minus - control' subtractions in the case of six phantom images using a NOVOSPECT scanner. The results indicate that the clusters of 'high noise' which are present in the differenced images may lead to difficulties in the interpretation of any apparent effects observed in images obtained in activation type studies. In chapter five further work in the area of neuroimaging is discussed. In particular, we emphasise the need for interval estimation in non-linear regression and the determination of the signal to noise ratio, in activation-type studies, which is required to have some assurance that any apparent effects in the difference image are not artefacts of the measurement process.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Additional Information: Adviser: Ian Ford
Keywords: Statistics, Medical imaging, Neurosciences
Date of Award: 1990
Depositing User: Enlighten Team
Unique ID: glathesis:1990-76501
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 19 Nov 2019 14:15
Last Modified: 19 Nov 2019 14:15
URI: http://theses.gla.ac.uk/id/eprint/76501

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year