טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentOrian Ehud
SubjectBlind Separation of Brain Tissue Signatures in Magnetic
Resonance Imaging
DepartmentDepartment of Electrical Engineering
Supervisor Professor Emeritus Yehoshua Zeevi
Full Thesis textFull thesis text - English Version


Abstract

    Blind Source Separation is an essential processing technique in a variety of communication and signal processing applications. This research addresses the application of a geometric approach for BSS in Magnetic Resonance Imaging of human brain. It presents new aspects and applications of the geometric BSS approach, that concern BSS of brain MRI.

    Raw MRI data is acquired by sampling the image frequency domain. Therefore, we propose a variation of the geometric approach in polar coordinates to complex-valued unmixing.

    The BSS problem of brain MRI seems to pose an underdetermined problem. We show that this is due to parasitic sources originating from a peripheral connective and bone tissue. It is therefore solvable in a region of interest. We also address the problem that stems from sources of shared features, such as tissues that complement each other in space. We show that virtual sources may result from distributions of shared features in the sources. We discover emergence of extra clusters in the scatter-plots, indicating unrecognized sources and hence yielding what appears to be an ill-posed problem due to possible existence of more sources than observations.

    Due to the prolonged acquisition time of MRI scans, a set of MR images, acquired with different scanning parameters, may result in a lack of registration. We propose an application of the geometric approach to registration of unregistered mixtures.

    The recovery error is examined via a linear approximation. The analytic result indicates that the error in sources recovery depends also on the geometry dictated by the true mixing matrix. This lends itself to a BSS approach in a problematic case where the mixtures' representations are almost collinear. We suggest a method for decolinearization of mixtures, by a transformation to an ordinary BSS problem which is solvable in additional iteration.

    The second part of this research is devoted to search for separable MRI scans. We show that such scans are the set corresponding to spin-lattice and spin-spin relaxation times. Separation results show enhanced Gadolinium effect, improving contrast of tissue affected by Gadolinium. The second set is composed of Diffusion-Weighted MRI scans. Considering two distinct diffusion populations, the conventional separation method currently implemented, consists of acquiring data from 15 scans. As a possible alternative, we propose that the model can be reduced to a linear mixing model, which is solvable by blind separation of two DWMRI scans, acquired with only two diffusion weighting values.