Blind Source Separation of MRI

This project addresses the application of a geometric multiscale framework Blind Source Separation (BSS) on Magnetic Resonance Imaging (MRI).

Abstract
This project addresses the application of a geometric multiscale framework Blind Source Separation (BSS) on Magnetic Resonance Imaging (MRI). The motivation for such application emerges from the major presence of MRI signal originated from water and fat molecules, that suffers a crosstalk due to the close chemical shift of these body tissue major substances. We have implemented and explored several issues, including finding sparse representations of brain MRI of axial slices, centroids extraction and multi dimensional problems. As well, we have implemented a GUI ( Graphical User Interface) that enables the examination of relevant procedures on MRI data selected by the user.

 

The problem
Some MR images are mainly composed of several distinct sources. For example, brain MR images that are mainly composed of fat and water, which combine to exhibit the properties of the examined tissues. In such cases, the physician might be interested in two separated images (i.e. fat and non-fat components of the tissue, dominated by water). Therefore, BSS application to MRI may be very valuable.

 

The solution
We search for a representation in which the sources are sparse. We then project the data onto the transformed mixtures’ space to construct a scatter-plot. Under some conditions, such a scatter-plot reveals clustering along orientations indicating the parameters of the underlying linear mixing model, thereby enables the recovery of estimated sources.

1
Figure 1 – a Diagram of our approach

 

Tools
The entire project is implemented with Matlab.

 

Conclusions
We concluded that the geometric approach is applicable to some axial brain MRI, and may be used to facilitate BSS in medical imaging of other modalities as well. We also brought conclusions regarding finding sparse representations, solving maxima problems and working with fuzzy logic algorithms (FCM).

 

Acknowledgment
We are grateful to our project supervisor Ehud Orian for his help and guidance throughout this work. Also we are thankful to the Ollendorff Minerva Center which supported our project.