Covariate adjustment in partial least squares for the extraction of the spatial-temporal pattern from positron emission tomography data

Partial least squares (PLS) is a path modeling technique with latent variables that reduces data dimensions. The power of PLS for Positron Emission Tomography (PET) data analysis is found in its superior feature characterization capabilities and statistical flexibility. No assumptions are required for the underlying distribution of the data (relative to other commonly applied methods: SPM and MANCOVA-CVA). However, the PLS methodology as it exists, can not adjust for covariate effects in the spatial-temporal patterns. The aim of this dissertation work was to extend the PLS methodology to allow for covariate adjustment in the extracted spatial-temporal pattern from image data. This was accomplished by introducing multi-dimensional latent variable for image data to adjust for covariate effects. The extended PLS method was applied to both a simulated data set and an observed sequential PET data set acquired to study the temporal and spatial effects of amphetamine (AMPH) on cerebral blood flow.;PET image data were simulated for baseline and post-stimulation (post-AMPH) states. Data were simulated for a single transaxial PET image plane that consisted of two voxel groups: cerebellum and non-cerebellum with different brain activities. The non-cerebellar data were assumed not to be responsive to the AMPH, while three different temporal patterns (parabolic, constant, linear increment) were assumed for cerebellum. Simulation studies were conducted for circumstances of no covariate effect, different temporal patterns caused by covariate (gender), and different spatial patterns caused by covariate (gender). Parameters were based on the observed sequential PET studies. Five hundred realizations were generated for each parameter combination.;For the experimental data, the temporal and spatial patterns identified by the basic and the extended PLS methods were consistent with those identified by other methods. In simulations when no covariate effect was assumed, both methods performed equally. In simulations when covariate effects were present, the extended PLS was found to characterize the differential spatial and temporal patterns across genders compared to the basic PLS method. It is concluded that the proposed extension of the PLS method enhanced its feature characterization ability when different patterns exist due to covariate effects.