@MISC\{IMM2002-05151, author = "F. {\AA}. Nielsen", title = "Virtual brain mapping: Meta-analysis and visualization in functional neuroimaging", year = "2002", month = "sep", publisher = "Informatics and Mathematical Modelling, Technical University of Denmark", address = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby", note = "Talk at Brain Forum {7,} Department of Health Science and Technology, Aalborg University", url = "http://www2.compute.dtu.dk/pubdb/pubs/5151-full.html", abstract = "Results from functional neuroimaging such as positron emission tomography and functional magnetic resonance are often reported as sets of {3-}dimensional coordinates in Talairach stereotactic space. By utilizing data collected in the BrainMap database and from our own small {XML} database we can automatically model and visualize several studies at once. We model a set of {3-}dimensional coordinates by a voxelization step where flexible probability density models such as kernel density estimators produce a voxel-volume representation of a study, allowing us to represent all coordinate data in one single data matrix. By conditioning on elements in the databases other than the coordinate data, e.g., anatomical labels associated with many coordinates we can make conditional novelty detection identifying outliers in the database that might be errorneous entries or seldom occuring patterns. In the BrainMap database we found errors, e.g., stemming from confusion of centimeters and millimeters during entering and errors in the original article. Conditional probability density modeling also enables generation of probabilistic atlases and automatic probabilistic anatomical labeling of new coordinates. By conditioning on the behavioral domains associated with each study, e.g, the words `word' and 'visual{',} we can make virtual brain activations. Voxelization also permits us to find related volumes, where query volumes are matched with database items and the most related volumes are found and returned in sorted lists. Image-based indices can be created by singular value decomposition and by matching individual volumes against eigenimages. Individual experiments, sets of experiments as well as results from meta-analyses can be rendered as glyphs, cut-planes or isosurfaces in {3-}dimensional Corner Cube Environments or exported as VRML97 and made available on the Internet, see http://hendrix.imm.dtu.dk." }