Detection of fine-scale activity patterns by integration of information, in local regions

The widespread statistical parametric mapping standardly performs spatial smoothing of the data with a Gaussian kernel (GK) to improve signal to noise ratio and statistical power. However, the best filtering is dependent on the shape of the activation regions, which is irregular in nature and not well matched by a constant GK. As a result, smoothing the data with a GK will obscure fine-scale patterns of weak effects that contain neuroscientifically relevant information. To improve the sensitivity of activation detection, in the presented work, multivariate statistical technique (PCA) and univariate statistical technique (GLM) were combined together to discover the fine-grained activity patterns. The time courses from every local homogenous regions were first integrated with PCA; then, GLM was used to construct the interests of statistic. The approach has implicitly taken account of the structures of both BOLD signal and noise existed in local regions. Therefore, it can highlight details of different regions. Experiments with real fMRI data, demonstrate that proposed technique can dramatically increase the sensitivity of the detection of the fine-scale brain activity patterns which contain subtle information about the experimental conditions.