Neural network-based analysis of MR time series

Abstract

Clustering has been introduced to analyze fMRI data by means of partitioning data into time series of similar temporal behavior. It is hoped that one of these clusters represents a dynamic effect of interest, like functional activation. Using self-organizing maps for clustering, additional information can be obtained by ordering cluster centers on a two-dimensional projection plane. The map's capability of data visualization is used to summarize all dynamic effects of an experiment by means of data partitioning. The map does allow differently sized and populated clusters in the data by forming "superclusters" on the map. The method is introduced as a conceptual extension to clustering. Applications to fMRI and to MR mammography are discussed.