TOPICS

Technical Program:

Registration:

ICASSP2000 Connections:

Tutorial 4: 08:30 - 12:00, Monday, June 5, 2000

Signal Processing Strategies in Neuroinformatics
by
Lars Kai Hansen (Technical University of Denmark, Lyngby, Denmark)

Abstract

     Neuroinformatics is a new interdisciplinary research field which encompasses classical disciplines such as biology, physics, computer science, and engineering. It combines ``learning from the brain'' with ``learning about the brain''.  One of the major efforts in neuroinformatics concerns neuroimaging - mapping brain function by PET and fMRI (positron emission tomography and functional magnetic resonance imaging). Neuroimaging is opening a new window to the working human brain and brain mapping will have immense influence on the way we think about the brain both in research and in the clinic. Adaptive systems and neural networks in particular construct abstract models by learning. The vast databases created by current brain scanning techniques calls for new innovative signal processing strategies and for new visual and interactive means of communication and form a significant engineering challenge for the next century. From an application point of view such knowledge gains credibility by being expressed in intuitively accessible forms, say by computer graphics, and interactively in the form of hypertext. With the advent of visual hypertexts such as clickable maps or VRML (Virtual Reality Modeling Language) graphics and hypertext are combined.

Examples of work in this area can be found at:

http://hendrix.imm.dtu.dk
http://www.nimh.nih.gov/neuroinformatics/index.htm
http://www.neuroguide.com/bestbets.html   

 

     The tutorial will review the basic neuroimaging data acquisition schemes and signal processing needs for neuroinformatics. It will also state the engineering challenges associated with data analysis: normalization, intersubject co-registration by warping, modeling using principal and independent component analysis (PCA, ICA) an linear and non-linear representations, information visualization and communication. The presentation includes illustrations with experimental data involving PET and fMRI studies of motor activity, visual stimulation as well as cognitive tasks. A short outline of the course is as follows:

 

* Neuroinformatics: definitions, background and major projects around the world
* Neuroimaging: 3+1 D, the spatio-temporal structure of PET and fMRI data sets.
* Signal processing challenges and current strategies for preprocessing and representation: PCA, wavelet analysis, ICA. The problem of extremely ill-posed learning (learning high-dimensional models from small databases).
* Learning linear and non-linear relations between blood flow and function in the brain.
* Information visualization: VRML * Perspectives and open research issues

 

A comprehensive Matlab toolbox for fMRI signal processing including example neuroimaging data sets is offered free of charge for the scientific community.