|M.Sc Student||Cohen Sarit|
|Subject||Improvement of Spatial Resolution of Source Estimation of|
Brain Activity Using Blind Source Separation
|Department||Department of Medicine||Supervisors||PROFESSOR EMERITUS Hillel Pratt|
|DR. Michael Zibulevsky|
The research aim was to improve the spatial resolution of LORETA by using techniques of Blind Source Separation such as ICA (Independent Component Analysis) and Sparse Component Analysis. The secondary aims were testing the BSS techniques as an alternative to averaging.
The ICA method was used to separate blindly mixed signals recorded from the subject's scalp. The learning algorithm was based on the information maximization in a single layer neural network. The sparsity algorithm was used to transform the ERP into sparse signals, with most of their values equal or close to zero. The sparse signal contains few values significantly different from their surrounding.
In this study we analyzed data from 12 subjects in an oddball paradigm in which subjects distinguished between two auditory stimuli of 1500 Hz and 1600 Hz. The ERP raw data was analyzed in two ways; the traditional way of averaging and the ICA with sparsity. The results of both techniques were compared.
In the first phase we removed noise sources such as the electric activity of the heart, eye blinks and eye movement. In The second phase we chose 5 ICA components, which have the largest contribution to data in a time interval of N1 component. Those 5 dominant components of ICA were reconstructed together to form a clear signal. In the third phase we compared between the ICA reconstructed signal and the averaged signal. The results show the same brain activity area (estimated by LORETA) in both methods. The fourth stage was comparing the activity areas in terms of spatial resolution.
The results show a significantly advantages of the ICA plus Sparsity compared to the traditional technique of averaging.
The conclusions are that BSS methods can be used to enhance SNR. The ICA algorithm succeeded in separating certain source signals that were artifacts or noise sources not corresponding to brain activity. In addition, using the ICA with sparsity algorithms improved the spatial resolution of LORETA. The final conclusion is that ICA offers a very promising method, which should be further investigated.