Abstract
275
Objectives: While principal component analysis (PCA) has been explored extensively in brain imaging to identify patterns associated with a particular clinical condition, we here propose using it to identify patterns that are neither primarily related to that condition nor orthogonal to the pattern of interest in order to reduce noise before statistical image interpretation. In particular, we here propose identifying principal components from healthy controls and removing them from patient data before calculating the degree (pattern expression score) to which a known pattern associated with a particular clinical condition is present in the patient data. A theoretical framework is provided to estimate the benefit of denoising as well as the cost. Furthermore, the effect of denoising is shown for FDG-PET data from 180 controls from the ADNI-database, which were used to identify PCs (n=130) and manipulated (n=50) to contain a random degree of a pattern known as PDCP (Parkinson disease cognition related pattern: widespread cortical hypometabolism, subcortical hypermetabolism, particularly in cerebellum and brainstem, here derived from our own FDG-PET-data). The effect of denoising was measured in terms of altered Pearson’s correlation between the pattern expression score derived from the manipulated images and the random score used for manipulation. For our pattern, Monte Carlo simulation showed best results when using all available PCs for denoising (R² = .77 as opposed to R² = .40 without denoising) while removing single PCs often worsened the correlation. We conclude that the proposed method is a powerful tool to improve statistical image interpretation. The proposed procedure differs from the usual procedure in that the PCs that are interpreted as noise and thus removed from the data are not orthogonal to the (known) pattern of interest so that, in combination, they may explain a part of the noise component of the pattern expression score. The proposed method can be applied before calculation of an index for individual diagnosis but should be evaluated for each pattern. To our knowledge, it hasn’t been described before.