Use of resampling techniques in assessment of robustness of complex image analysis: An example in FDG-PET of patients with Alzheimer’s disease

For a finite noisy sample, extraction of meaningful information which is representative of the population is a formidable task in which over-analysis of the data is not uncommon. Resampling techniques such as the bootstrap can be used to assess the robustness of the information to the particular sample from which this information is extracted often through complex image analysis. In this work, resampling techniques have been applied to principal component analysis (PCA) and Fisher discriminant analysis (FDA) of 2-Fluoro-2-Deoxy-D-Glucose (FDG) PET brain images of healthy volunteers (HVs) and Alzheimer’s disease (AD) patients. The objective of this work is to find those principal components (PCs) of the sample which would allow robust inferences about future brain images. The metric of angle between PCA subspaces has been used to evaluate the robustness of PCA subspaces formed by different PCs. Using a sample of 42 images (23 ADs and 19 HVs) only the first three PCs can be regarded as robust permitting reliable FDA discrimination of AD patients from HVs. In addition, resampling was used to generate standard error images highlighting the less robust brain regions in the discrimination task.