Magnetic Resonance Imaging and the Fourier Transform

Patrick Foley - University of California Santa Barbara

March 13, 2016

Abstract
Magnetic resonance imaging is the process in which images of tissue are produced by taking advantage of the magnetic properties atoms, hydrogen in particular. Since people are mostly made
of hydrogen, magnetic resonance imaging works well to form images of human tissue.
Hydrogen is the most simple atom, consisting of just a proton and an electron. It is the behavior
of the proton in a magnetic field that makes magnetic resonance imaging possible. When in the
presence of a strong magnetic field, the proton absorbs and emits electromagnetic radiation at
specific frequencies. These frequencies can be detected and stored as data.
The magnetic resonance phase and frequency signal comes in as a complicated sum of many
simple signals and is stored that way as data. The way that this complicated data is organized
and turned into an image is by way of Fourier transform. Fourier transforming the data takes the
complicated signal and breaks it up into specific frequencies so it can be organized as pixel data.
This pixel data can then be displayed as an image.
The technology for magnetic resonance imaging has been being developed over the past century
starting with early research in nuclear magnetic resonance. The ability to produce images from
magnetic resonance signals would not be possible without the Fourier transform.

Introduction

absorb electromagnetic energy and re-emit the

energy when they would relax back to their origi1.1 Nuclear Magnetic Resonance
nal state in the magnetic field.[4] This absorption
and emission of electromagnetic energy turned
Magnetic resonance imaging came about from
out to be the key to magnetic resonance imagresearch in nuclear magnetic resonance. Nuing.
clear magnetic resonance is the the study of how
atomic nuclei absorb and emit electromagnetic
As nuclear magnetic resonance developed,
radiation in the presence of a magnetic field. many wondered whether or not you could create
This branch of study was pioneered by Walter an image out of magnetic resonance signals. In
Gerlach and Otto Stern in the 1920s. In 1924, 1975, Richard Ernst proposed the use of Fourier
they published the results of an experiment that transform of the phase and frequency data in
demonstrated the quantum nature of the mag- order to organize spatial data to reconstruct a
netic moment of silver atoms by molecular beam 2 dimensional image. This research led to the
deflection in an inhomogeneous magnetic field.[4] first images being reconstructed from magnetic
This research paved the the way for the study of resonance signals. Richard Ernst was awarded
nuclear magnetic resonance.
the 1991 nobel prize in chemistry for his Fourier
In the mid-1900s two physicists, Edward M. Transform method of image reconstruction.[4]
Purcell and Felix Bloch, found that when certain Without the Fourier transform, magnetic resonuclei were placed in a magnetic field they would nance imaging would not be possible.
1

1.2

Magnetic Resonance Imaging

2.1
It is the behavior of the proton in a magnetic field
that makes magnetic resonance imaging possible.
Protons have a magnetic moment which will line
up with the magnetic field of a strong magnet
and begin to precess with their angular velocity in the same direction of the field, and with
a frequency dependent on the strength of the
magnetic field at that point. If you then disturb the protons with a radio frequency pulse,
they begin to precess at a higher energy as a
result of absorbing the electromagnetic energy.
When the pulse is turned off, protons in a higher
energy state relax back down to a lower energy
state. When the proton relaxes to a lower energy
state, it gives off electromagnetic radiation at a
specific frequency dictated by its position in the
magnetic field gradient. A detector called a radiofrequency coil is used to turn the frequencies
given off by the relaxing protons into an electronic signal. A radiofrequency coil consists of
two electromagnetic coils, a transmitter and receiver that generate and receive electromagnetic
fields. [3,4]

Materials and Methods

The Fourier Transform

The Fourier transform is a mathematical tool

that is used to decompose a function or signal
into its constituent sinusoids. As it turns out, all
periodic functions can be represented as a sum of
simple sinusoids of different frequencies (fig. 1).
This is a great tool for magnetic resonance imaging since the incoming signal is a sum of simple
sinusoids of different frequencies.[2]

This signal comes in as a very complicated

wave form and is stored as data. The way in
which the data is organized is
by the use or the
R
Fourier transform, f() = f (x)e2ix dx,
which allows software to break up the signal into
individual frequencies so that they can be organized into pixel data that can be inverse transformed back into an image.
Fourier transform is used to decompose the
complicated signal received from magnetic resonance echoes. These echoes contain encoded
frequency and phase spatial information used to
construct an image. The Fourier Transform separates all encoded frequencies so that the data
can be stored and organized as pixels in what is
known as k-space. When inverse-transformed,
the pixel contains one spatial frequency contributing to the image. A spatial frequency corresponds to the number of alternating light and
dark bands per meter in the pixel.[3,5]

Figure 1: Composition of a complicated waveform from

simple sinusoids.[2]

The Fourier transform is typically used to analyze functions of time or space by looking at
them in their frequency domain then transforming them back to their time domain. Here is the
2

Fourier transform of a function in the time do- Fourier space. K-space is used to organize spamain and its inverse:
tial frequency and amplitude information into
pixels. A single pixel in k-space can be inversetransformed to contribute one specific spatial freZ
2it
quency to the image in the form of alternating
g(t)e
dt = G()
F{g(t)} =

light and dark bands.[1] The inverse transform

of all of the information in k-space combines all
Z
spatial frequencies resulting in the entire image.
2it
1
The frequency and orientation of the alternating
G()e
d = g(t)
F {G()} =

light and dark bands that make up the the image depend on where the pixel resides in k-space.
You can see here that the Fourier transform
Pixels with low spatial frequency are mapped
takes a function of time or space and transforms
near the origin and pixels with high spatial freit into a function of frequency or spacial frequency are mapped on the outskirts (fig. 3).[3]
quency. What might not be obvious is that the
frequency domain function will be represented in
a basis of sines and cosines. This will allow for
the separation of discrete frequencies that make
up a function of time or space (Fig. 2).

Figure 3: MRI. This coronal slice of a brain is interrogated for all its different spatial frequencies by successively
altering magnetic field gradients (open arrows in top three
images) during frequency- and phase-encoding. Although
only three examples are shown here, many different gradient combinations are necessary to fill k-space with enough
Figure 2: Fourier transform (FT) extracts the frequencies

phase and spatial information to create an image. In-

and relative amplitudes of the simpler waves hidden in a

verse Fourier transform (iFT) of k-space essentially adds

complicated wave g(t). Inverse Fourier transform (iFT)

the relative contributions of all spatial frequencies to give

restores the time domain. In this example, Fourier trans-

the final image.[3]

form of three cosine waves of different frequencies results

in three delta functions.[3]

2.2

2.3

Image Refinement

An important part of constructing an image is

encoding phase spatial information. Without
this phase spatial information it would be impossible to reconstruct an image. One way that
the phase information is encoded in the signal
is by using a phase encoding magnetic gradient.
This is achieved by quickly activating and deactivating the gradient. When the gradient is only

K-Space

When the signal of the magnetic resonance echo

comes in, it is incredibly complicated. As we
have seen from 2.1, the Fourier transform can
help us with this by breaking up the signal
into its discrete frequencies. Once dismantled,
the data is entered into k-space which is a 2D
3

on for a short time, some of the protons spins

precess faster or slower than others depending
on their position in the field gradient. When allowed to relax in state, a change of phase is imprinted on the protons due to the timing of turning the gradient on and off and that information
is carried over in their given signal.[3] The more
phase encoding steps that are taken, the greater
the image resolution will be due to the gathering
of more phase information.
In order to further refine images, a thresholding technique has been developed to reduce noise
and eliminate false correlation in pixels with very
large signal changes due to the flow of fluid in
the body. The mechanics of this method can
be best explained using abstract vectors containing time-course data information. The idea is to
form a correlation coefficient, which we will call
cc, for each pixel, and define a threshold value,
T H, where if |cc| > T H then the pixel is thrown
out. In order to define the correlation coefficient,
we must define the vector ~f which contains the
functional time domain magnetic resonance information and the vector ~r which is the corresponding reference vector. ~r may be derived
from the average of several experimentally derived ~f s. The correlation coefficient can be
written:

to remove bad data which in turn refines the

image.[1]

Results

Magnetic Resonance Imaging is made possible

by the behavior of a proton in a magnetic field.
It is the ability of the proton to absorb and emit
electromagnetic radiation, nuclear magnetic resonance, that allows us to extract information
from inside the body to create an image.
Because the processing of complicated signals is inherent to magnetic resonance imaging,
the Fourier transform is clearly necessary in order to make sense of the information contained
in the signals obtained from magnetic resonance
echoes.
There are many ways to improve the quality of the images obtained in magnetic resonance
imaging, one of those being the encoding of phase
information by quickly turning a magnetic field
gradient on and off. The more phase information encoded and obtained, the higher the image
resolution will be.
In order to refine magnetic resonance images
further, sophisticated mathematics have been
used to create a theory for image refinement. A
correlation coefficient is defined using abstract
vectors containing time course data. This corre~f ~r
cc =
lation coefficient can be compared to a threshold
~f ~r
value in order for software to decide how to keep
The angle between the vectors ~f and ~f can good data and throw out bad data. The timecourse data in each pixel can be examined usbe written:
ing this thresholding technique in both time and
f ~r
frequency domains by way of Fourier transforma1 ~
|
= cos |
~f ~r
tion in order to remove bad pixels from images.

The thresholding value then implies that

data will be accepted when < cos1 (T H).
The smaller the angle, the more closely shapes
resemble each other. The typical angle used for
thresholding is 4 . The goal here is to create software that will make a digital map of zeros and
ones with the same dimensionality as the image.
Ones are assigned to pixels with similar shapes
as defined by the threshold value, and zeros are
assigned to the others, essentially removing the
data from those pixels. This method is meant

Discussion

Before the Fourier transform method was created by Richard Ernst in 1975, a few people
tried their hand at image formation from magnetic resonance signals. In 1973, Physicist Peter
Mansfield proposed using magnetic field gradients to acquire spatial information in order to
produce an image. That same year at a conference in Poland he showed that he produced
4

Brain. Magnetic Resonance in Medicine,

30(1), 161-173.

a magnetic resonance interferogram at a 1mm

resolution.[4] This was not an actual image but
it was still a pretty good reconstruction of spatial
information.
The only way in which the magnetic resonance signal can be turned into specific frequencies and mapped to a pixel is by Fourier transformation. There are many methods of refining
magnetic resonance images and most of them use
Fourier transformation in one way or another.

2. Bevel, P. (2010).
Fourier Transforms. Retrieved February 08, 2016, from
http://www.thefouriertransform.com
3. Gallagher, Thomas A., Nemeth, Alexander J., Hacein-Bay, Lotfi.
(2008).
An Introduction to the Fourier Transform: Relationship to MRI. American
Journal of Roentgenology, 190(5), DOI
10.2214/AJR.07.2874

Acknowledgments

Special thanks to Jared Pagett and Dylan Beard

for their feedback on this paper.

4. Geva, T. (2006).
Magnetic resonance
imaging: historical perspective. Journal of
Cardiovascular Magnetic Resonance, 8(4),
573-580.

References

5. Oshio, K., and Feinberg, D. A. (1991).

GRASE (Gradient and Spin Echo) imaging: A novel fast MRI technique. Magnetic
resonance in medicine, 20(2), 344-349.

1. Bandettini, Peter A., Jesmanowicz A.,

Wong, Eric C., Hyde, James S. (1993).
Processing Strategies for Time-Course
Data Sets in Functional MRI of the Human