fMRI is a diagnostic imaging method, which is based on the magnetic resonance phenomena.
Abstract
fMRI is a diagnostic imaging method, which is based on the magnetic resonance phenomena. The BOLD technique (Blood Oxygen Level Dependent) is a common technique which is used for fMRI as it identifies areas in the brain with a high level of oxygen in the blood. According to research, it was discovered that high oxygen level indicates of increasing activity level of the same area in the brain, thus the fMRI experiment aim is to find the active sources in the brain during a specific activity. The problem of separating the sources from the fMRI data sequence (without any prior knowledge of the sources and mixture qualities) can be treated as a Blind Source Separation problem (BSS). This project deals with different methods of achieving a solution for the Blind Source Separation problem using sparse representations and implementing them for genuine fMRI data sequences.
The problem
The active structure of the brain consists several independent sources, which are responsible for different kinds of activities. The problem of separating sources from fMRI data sequences may be treated as `Blind Source Separation`(BSS). We wish to separate blindly (without any prior knowledge of the sources and mixture qualities) the independent sources from the fMRI images while we assume that the images are linear combination (mixture) of the desirable sources.
The solution
To simplify the Blind Source Separation problem we have used the sparse representations technique. In order to solve the BBS problem, we have used two methods: Geometric separation method (effective with two sources). ICA_Newton algorithm using Newton optimization method (after using PCA algorithm).
The fMRI simulation:
The general results
The separation algorithm outcome of the 1st experiment (between the range of 1 to 128 images):
The sources from the 16 separated images:
Tools
The Project was programmed in Matlab 6.5, on a PC platform. The Matlab tools which were used: Wavelet Packet toolbox, PCA tools and The ICA toolbox (for implementing ICA_Newton separation).
Conclusions
The geometric method disadvantage is that it`s difficult to suit it for a larger number of mixtures in order to achieve satisfied results (it`s appropriated for low dimensions of several images). Therefore, we have used the ICA_Newton algorithm. When we have started to deal with the fMRI data, we have found that the gradient method doesn`t sparse well the mixtures since the mixtures don`t consist edges. Thus, we have noticed that the WP method was appropriated to sparse the fMRI mixtures. We have examined the active sources for every image range and we have planted an artificial source in order to examine the consistency of the algorithm. Finally, after achieving great and satisfied results, we have analyzed the separation outcome and determined the real sources.
Acknowledgment
We would like to thank our devoted supervisor Mr. Michael Bronstein for his professional guidance of developing this project. We are also grateful to the lab engineer Mr. Johanan Erez and to the VISL lab staff for their support and help.
Moreover, we are also grateful to the Ollendorf Minerva Center Fund for supporting this project.