Discriminating one finger from another: Support vector classification of event-related fMRI

LaConte, S.; Chen, J.; Peltier, S.; Hu, X.

by S. LaConte, J. Chen, S. Peltier, X. Hu

Abstract:

In this paper, we describe our work in classifying brain state using event-related (ER) fMRI data, which, unlike block design, span several time points. We construct hyper-images by concatenating images within ER epochs and use feature selection to reduce dimensionality and improve accuracy. We evaluated two feature selection strategies and investigated the limits of detecting brain state events using support vector machine classification (SVC). For right vs. left finger movement, we obtained good classification, achieving nearly perfect accuracy with feature selection. Discrimination of right-handed index vs. pinky finger movement required more training data, but achieved remarkable results (70% accuracy).

Reference:

abstract S. LaConte, J. Chen, S. Peltier, X. Hu. Discriminating one finger from another: Support vector classification of event-related fMRI. In Proceedings 13th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Miami, page 1583, 2005. [bibtex]

Bibtex Entry:

@inproceedings{Miami1583,
   Author = {LaConte, S. and Chen, J. and Peltier, S. and Hu, X.},
   Title ={Discriminating one finger from another: Support vector classification of event-related f{M}{R}{I}},
   BookTitle = {Proceedings 13th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Miami},
   Pages = {1583},
   Abstract = {In this paper, we describe our work in classifying brain state using event-related (ER) fMRI data, which, unlike block design, span several time points. We construct hyper-images by concatenating images within ER epochs and use feature selection to reduce dimensionality and improve accuracy. We evaluated two feature selection strategies and investigated the limits of detecting brain state events using support vector machine classification (SVC). For right vs. left finger movement, we obtained good classification, achieving nearly perfect accuracy with feature selection. Discrimination of right-handed index vs. pinky finger movement required more training data, but achieved remarkable results (70% accuracy).},
 Keywords = {Miami1583},
   Year = {2005} }