Share Email Print

Proceedings Paper

Unsupervised time course analysis of functional magnetic resonance imaging (fMRI) using self-organizing maps (SOMs)
Author(s): Stephan G. Erberich; Matthias Fellenberg; Timo Krings; Stefan Kemeny; Wolfgang Reith; Klaus Willmes; Walter Oberschelp
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Functional Magnetic Resonance Imaging (fMRI) data of the brain includes activated parenchymal voxels, corresponding to the paradigm performed, non-activated parenchymal voxels and background voxels. Statistical tests, e.g. using the general linear model approach of SPM or the Kolmogorov-Smirnov (KS) non-parametric statistic, are common 'supervised' techniques to look for activation in functional brain MRI. Selection of voxel type by comparing the voxel time course with a model of the expected hemodynamic response function (HRF) from the task paradigm has proven to be difficult due to individual and spatial variance of the measured HRF. For the functional differentiation of brain voxels we introduce a method separating brain voxels based on their features in the time domain using a self-organizing map (SOM) neural network technique without modeling the HRF. Since activation measured by fMRI is related to magnetic susceptibility changes in venous blood which represents only 2 - 5% of brain matter, preprocessing is required to remove the majority of non- activated voxels which dominate learning instead of real activation patterns. Using the auto-correlation function one can select voxels which are candidates of being activated. Features of the time course of the selected voxels can be learned with the SOM. In the first step the SOM is trained by the voxels time course, fitting its neurons to the input. After learning, the neurons have adapted to the intrinsic features space of the voxel time courses. Using the trained SOM, voxel time courses are presented again, now labeled by the neuron having the smallest Euclidean distance to the presented voxel time course. The result of the labeling and the learned feature time course vectors are compared visually with the p-value map of the KS statistic. With the SOM map one can visually separate the voxels based on their features in the time domain into different functional task related classes.

Paper Details

Date Published: 20 May 1999
PDF: 8 pages
Proc. SPIE 3660, Medical Imaging 1999: Physiology and Function from Multidimensional Images, (20 May 1999); doi: 10.1117/12.349596
Show Author Affiliations
Stephan G. Erberich, Univ. Hospital/Technical Univ. Aachen and Max-Planck-Institute for Biochemistry (United States)
Matthias Fellenberg, Max-Planck-Institute for Biochemistry (Germany)
Timo Krings, Univ. Hospital/Technical Univ. Aachen (Germany)
Stefan Kemeny, Univ. Hospital/Technical Univ. Aachen (United States)
Wolfgang Reith, Univ. Hospital/Technical Univ. Aachen (Germany)
Klaus Willmes, Univ. Hospital/Technical Univ. Aachen (Germany)
Walter Oberschelp, Technical Univ. Aachen (Germany)

Published in SPIE Proceedings Vol. 3660:
Medical Imaging 1999: Physiology and Function from Multidimensional Images
Chin-Tu Chen; Anne V. Clough, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?