Robustness of correlations between PCA of FDG-PET scans and biological variables in healthy and demented subjects

In neuroimaging it is helpful and useful to obtain robust and accurate estimates of relationships between the image derived data and separately derived covariates such as clinical and demographic measures. Due to the high dimensionality of brain images, complex image analysis is typically used to extract certain image features, which may or may not relate to the covariates. These correlations which explain variance within the image data are frequently of interest. Principal component analysis (PCA) is used to extract image features from a sample of 42 FDG PET brain images (19 normal controls (NCs), 23 Alzheimer's disease (AD) patients). For the first three most robust PCs, the correlation of the PC scores with: i) the Mini Mental Status Exam (MMSE) score and ii) age is examined. The key aspects of this work is the assessment of: i) the robustness and significance of the correlations using bootstrap resampling; ii) the influence of the PCA on the robustness of the correlations; iii) the impact of two intensity normalization methods (global and cerebellum). Results show that: i) Pearson's statistics can lead to overoptimistic results. ii) The robustness of the correlations deteriorate with the number of PCs. iii) The correlations are hugely influenced by the method of intensity normalization: the correlation of cognitive impairment with PC1 are stronger and more significant for global normalization; whereas the correlations with age were strongest and more robust with PC2 and cerebellar normalization.