Finn Å rup Nielsen
Sun Feb 25 19:24:29 PST 1996
A single subject functional magnetic resonance imaging (fMRI) study is analysed using a preproccessing of singular value decomposition principal component analysis and a feed forward neural network analysis, using the entropic costfunction. The analysis is a classification of brain scans into classes according to an induced paradigm: sequential finger-to-thumb opposition. The analysis is ''backwarded'' to construct a saliency map giving the neural networks answer to what it finds important in the brain scans to explain the paradigm.
The hidden units of the neural network is given especially attention, and is used to reveal non-paradigm structure in the data. An analysis strategy is presented based on second moment statistics.
The visualization of the analysis results is aimed against a presentation in 3D. A world has been build in the describtion language VRML containing Talairach grid lines and labels, where isosurface brain images --- anatomical as well as functional --- can be put.