Preprocessing
Preprocessing in the context of the NPAIRS system refers to the manipulation of the data matrix prior to its analysis. Preprocessing does NOT refer to intra-subject alignment, inter-subject alignment, brain masking (aka, stripping), spatial smoothing, temporal smoothing, or any other post-reconstruction pre-analysis data manipulations. It is assumed these steps have been previously performed. See data requirements .
The current data matrix preprocessing options include:
These preprocessing methods are used to remove unwanted sources of variance from the data matrix prior to any analysis.
Design Matrix
Within the NPAIRS framework, the General Linear Model (GLM) can be used as a stand alone analysis method, or as a tool for removing unwanted variance from the data matrix prior to some other analysis method such as PCA/CVA. In either case, a design matrix is required that contains in its columns the linear effects that "describe" the columns with the data matrix.
The creation of a design matrix within the NPAIRS software is accomplished by the specification of fields within a volume list file (VLF). The VLF holds the names of the data files that are loaded into the rows of the data matrix, and also information about each row (volume) within this data matrix. This includes such information as to what subject each volume belongs to, whether or not the volume was a baseline or activation, the patient population (e.g, normals or disease) for each volume, any performance measurements taken during the scan, etc. With this information a fairly general design matrix can be constructed.
Univariate
Currently, the only univarite analysis method employed by NPAIRS is the General Linear Model (GLM). Each split of the data results in two independent data sets which are then fed separately, along with the user defined design matrix and contrast vector(s), to a GLM, resulting in T-statistics and c*beta spatial patterns.
Multivariate
The multivarite analysis methods employed by NPAIRS are Principal Component Analysis (PCA) and Canonical Variate Analysis (CVA).
The PCA paritions the total variance of the data matrix into orthogonal components consisting of PC scores and eigenimages. Each NPAIRS split creates a set of these eigenimages which are then compared (the comparison is done for each dimension).
The CVA uses the output from the PCA. The PC's from the PCA, or some subset of them, are used as the input variables to the CVA which performs discrimination based on some user defined group structure. The group structure is defined by selecting field(s) within the volume list file. Each disjoint half of an NPAIRS split is then fed through this PCA/CVA analysis method, resulting in canonical eigenimage spatial patterns and canonical variates.
Combination
It is possible to combine models (analysis methods) within the NPAIRS framework. This is done by using the GLM method as a preprocessor of the data matrix, and feeding the results to the PCA/CVA method. More specifically, a design matrix is constructed that holds effects of no interest (e.g, global effects, block effects). A GLM is run on the data matrix and a residual matrix is formed by subtracting away the uninteresting effects from the original data. The residual matrix replaces the original data matrix which is then fed to the PCA/CVA for further analysis.