Vous êtes sur la page 1sur 25

SYDE 770 Image Processing

Course Project

Non-Parametric Power
Spectrum Estimation
Methods
SYDE 770 Image Processing Course Project

Prof E. Jernigan
Eric Hui - 97142203
erichui@alumni.uwaterloo.ca

Thursday, December 12, 2002


Table of Contents
1 POWER SPECTRUM ESTIMATION (PSE) ......................................................................2
1.1 PERIODOGRAM................................................................................................................. 2
1.1.1 Spectral Resolution........................................................................................................... 4
1.1.2 Asymptotic Bias............................................................................................................... 4
1.1.3 Variance........................................................................................................................ 4
1.1.4 Complexity ..................................................................................................................... 5
1.2 MODIFIED PERIODOGRAM............................................................................................... 5
1.3 BARTLETT S METHOD ...................................................................................................... 7
1.4 WELCH S METHOD........................................................................................................... 8
1.5 BLACKMAN-TUKEY METHOD........................................................................................... 9
2 FEATURE EXTRACTION ................................................................................................. 12
2.1 E XPERIMENT O UTLINE .................................................................................................. 12
2.2 COMPARISON OF PSE METHODS.................................................................................... 14
2.3 CANCER DETECTION ..................................................................................................... 16
3 CONCLUSIONS...................................................................................................................20

- ii -
List of Figures

Figure 1 Triangular Window w (k ) and its DTFT W (e j ) ....................................................... 3

Figure 2 Prostate Ultrasound Image ............................................................................................ 12

Figure 3 Outlined and Identified Prostate Ultrasound Images..................................................... 12

Figure 4 Linearization of Power Spectrum. Left: P ( e j ) . Right: 20 log P ( e j ) .................... 13

Figure 5 Feature Image for Slope Feature.................................................................................... 13

Figure 6 Feature Image for Y-Intercept Feature .......................................................................... 14

Figure 7 Cancerous part identified by Dr. Fenster....................................................................... 18

Figure 8 Feature Images. Left: slope; Right: y-intercept............................................................. 18

Figure 9 Processed Images. Left: before thresholding; Right: after thresholding (33/255).......... 18

Figure 10 Cancerous part identified by Dr. Fenster..................................................................... 19

Figure 11 Feature Images. Left: slope; Right: y-intercept........................................................... 19

Figure 12 Processed Images. Left: before thresholding; Right: after thresholding (95/255) ........ 19

- iii -
List of Tables
Table 1 Comparison of Truncation Windows................................................................................ 6

Table 2 Expected Feature Values ................................................................................................ 14

Table 3 MSE for Slope Feature ................................................................................................... 15

Table 4 MSE for Y-Intercept Feature.......................................................................................... 15

Table 5 Interpretation of Table 3 and Table 4............................................................................. 15

Table 6 MSE Comparisons.......................................................................................................... 16

- iv -
Abstract
There are two purposes for this paper. One purpose is to analyze several common non-parametric
power spectrum estimation (PSE) methods. Another purpose is to analyze how power spectrum
can be used for feature extraction to detect cancer in an ultrasound image of a prostate.
This paper first analyzes five non-parametric PSE methods. They are: Periodogram Method,
Modified Periodogram Method, Bartletts Method, Welchs Method, and Blackman-Tukey Method.
The methods are evaluated and compared by four characteristics: spectral resolution, asymptotic
bias, variance, and complexity.
The five PSE methods were implemented in Matlab and were applied to a test ultrasound image.
The resulting estimated power spectrum was linearized. The slope and the y-intercept of the
linearized power spectrum were used as two feature values. The mean-squared-error was calculated
in order to evaluate the performance of the five methods for this feature extraction. It was found
that this particular feature extraction does not depend much on the spectral resolution and the
variance of the power spectrum. Consequently, the Periodogram Method, which has the lowest
complexity, is best for this feature extraction.
Lastly, the texture and the overall brightness of the image are reflected on the slope and the y-
intercept of the linearized power spectrum. This paper proposes a simple but effective approach
which uses these two features jointly to detect the cancerous part of the prostate.

- 1 -
1 Power Spectrum Estimation (PSE)
There are mainly two types of power spectrum estimation (PSE) methods: parametric and non-
parametric. In contrast to parametric methods, non-parametric methods do not make any
assumptions on the data-generating process or model (e.g. autoregressive model). This paper
analyzes five common non-parametric PSE methods. They are: Periodogram Method, Modified
Periodogram Method, Bartletts Method, Welchs Method, and Blackman-Tukey Method. All these
are estimation methods. The ideal power spectrum of a signal x (n) can be computed by first
finding the ideal autocorrelation rx (x ) :
N
1
rx ( ) = E[ x( n) x( n + )] = lim
N 2N + 1
x(n) x(n + ) .
n = N

This is an average computed over the infinite interval ( , ) . However, in image processing, the
signal is often of finite length because the size of any image is finite. Even with a large image, in
order to make the assumption of stationarity, often only a small window of the image is considered
at a time. In this paper, it is assumed that the window is small enough so that the assumption of
stationary holds, allowing the autocorrelation and the power spectrum to be a Fourier Transform
pair.
1.1 Periodogram
A simple-minded method of estimating the autocorrelation of a signal of finite length is called the
Periodogram Method. This estimation method assumes the observed signal x N (n) is a truncated
version of the originally intended signal x (n) :

1, 0 n < N
x N ( n) = x (n ) wN ( n) , where wN ( n) =
0, otherwise
The autocorrelation is calculated simply by computing the average over the finite interval as follows:
N 1
1
rPer ( ) =
N
x(n) x(n + ) .
n =0

This is mathematically equivalent to



1 1
rPer ( ) =
N
x N ( n) x N ( n + ) =
N
x N ( ) x N ( ) .
n =

Taking the Discrete Time Fourier Transform (DTFT), the power spectrum estimated using the
Periodogram Method is:
1 2
PPer ( e j ) = X N (e j ) ,
N
where X N (e j ) is the DTFT of the observed signal x N (n) .

The expected value of the estimated autocorrelation rPer ( ) is:

- 2 -
1 N 1
1 N 1
E[rx ( )] = E
N

n =0
x( n) x( n + ) =
N
E[x (n)x (n + )]
n =0

1 N 1
= rx ( )
N n= 0
N
= rx ( )
N
Thus, the estimated autocorrelation rPer ( ) is the product of the ideal autocorrelation rx ( )
multiplied by the triangular function w ( ) as follows:
N
rPer ( n) = w ( )rx ( ) , where w ( ) = .
N
The power spectrum can then be estimated by taking the DTFT of the estimated autocorrelation
rPer ( ) as follows:

[ ]
E PPer ( e j ) = E [DTFT {rPer ( )}] = DTFT {E[rPer ( )}
= DTFT {w ( ) rx ( )}
1
= W (e j ) Px ( e j )
2

1 sin ( N / 2 )
2

where W (e j ) = DTFT {w ( )} = and Px ( e j ) = DTFT {rx ( )} .


N sin( / 2)

Thus, the estimated power spectrum PPer (e j ) is equivalent to the ideal power spectrum Px (e j )
convoluted with the low pass filter W (e j ) . This low pass filter W (e j ) , as shown in Figure 1, is
generally undesirable because of two problems. Firstly, the ideal power spectrum is blurred or
smoothened by the low pass filter, and the amount of blurring is mainly determined by the width
of the main lobe. Secondly, the power spectrum at any particular frequency is leaking into the
side lobes, possibly masking the power spectra at the nearby frequencies. Section 1.2 describes
these problems in details with an example.
2
1 sin( 12 N )
W (e ) =
j

N k N sin( 12 )
w (k ) =
N

k DTFT k
-N 0 N 0
Figure 1 Triangular Window w (k ) and its DTFT W (e j )

- 3 -
In the following sections, several PSE methods, which try to improve the Periodogram Method, are
described. In order to evaluate these methods, four characteristics are analyzed. They are: (1)
spectral resolution, (2) asymptotic bias, (3) variance, and (4) complexity.
1.1.1 Spectral Resolution
The spectral resolution can be defined by the frequency at which the power of the low pass filter
drops to half of the total power. Since 20 log( 0.5) 3 , the spectral resolution is equivalent to the
3dB bandwidth of low pass filter. The spectral resolution measures the amount of blurring effect on
the ideal power spectrum. In general, the larger the bandwidth of the main lobe, the higher spectra
resolution is, and the more blurring effect the low pass filter has.

[ ]
Res P (e j ) = ( )3dB .
(Note this definition of spectral resolution might be counter-intuitive since one would normally
expect the higher the spectral resolution, the sharper the spectrum would be.)
For the low pass filter W (e j ) defined above, the spectral resolution using the Periodogram
Method is:

[ ]
Res PPer (e j ) = 0.89
2
N
Pleaser refer to [1] for derivations. For this paper, the exact derivation of this spectral resolution is
not important since the spectral resolutions of all other methods are derived from this for
comparison.
1.1.2 Asymptotic Bias
Another characteristic of an estimation method is asymptotic bias. One would expect that as the
data length N increases, the estimated power spectrum P ( e j ) estimated should converge to the
ideal power spectrum Px (e j ) . An estimation method is asymptotically unbiased if

[ ]
lim E P (e j ) = Px ( e j ) .
N

Using the Periodogram Method,

[ j
] 1 j j
lim E PPer (e ) = lim E W (e ) Px (e )
N N
2
1
Since the low pass filter W (e j ) converges to an impulse as N , the desirable condition
2

[ ]
lim E PPer (e j ) = P( e j )
N

is satisfied. Consequently, the Periodogram Method is asymptotically unbiased.


1.1.3 Variance
As the data length N increases, the variance of a desirable estimated power spectrum should vanish.
Mathematically, this condition can be expressed as:

- 4 -
[ ] [ ]
lim Var P (e j ) = E P (e j1 ) P (e j 2 ) = 0
N

Because the variance of the Periodogram Method is of fourth order, it is difficult to compute the
variance for a general signal x (n) . It is shown, however, in [1] that the variance is proportional to
the square of the power spectrum.

[ ]
Var PPer ( e j ) P 2 (e j )

For example, for a Gaussian white noise process with variance x2 , the variance is

[ ]
Var PPer ( e j ) x4 > 0 .
That is, even with a very large data length N, there are still some fluctuations or jitters in the ideally
flat power spectrum.
1.1.4 Complexity
The complexity of an algorithm can be analyzed by inspecting the loop structure of the
implementation. The Periodogram Method can be implemented as follows:
1) Compute the estimated autocorrelation
1 N 1
rPer ( ) =
N n =0
x( n) x( n + ) which is an O( n ) operation because there are two
2

nested loops, one for n and one for .


2) Take the Discrete Fourier Transform (DFT) of the estimated autocorrelation
2
1 N 1 j u
P per ( u) = rx ( ) e N which is an O( n 2 ) operation because there are two
N =0
nested loops, one for and one for u .
Thus, the overall complexity of the Periodogram Method is O( n 2 ) + O( n 2 ) = O(n 2 ) .

1.2 Modified Periodogram


The Periodogram Method described above assumes that the observed signal x N (n) is a truncated
version of the original intended signal x (n) , with a rectangular truncation window:

1, 0 n < N
wN ( n) = .
0, otherwise
The Modified Periodogram generalizes the truncation window. By choosing a different truncation
window, the width of the main lobe and the amplitudes of the side lobes of the low pass filtered can
be varied. On one hand, the width of main lobe determines the spectral resolution. The larger the
width of the main lobe, the more blurring effect the low pass filter has on the power spectrum. On
the other hand, the amplitudes of the side lobes determine the spectral masking effect. The higher
the amplitudes of the side lobes, the more masking effect of the power spectrum at a given
frequency has on the nearby frequencies.

- 5 -
Table 1 shows several common truncation windows which are defined over the finite interval
0 n N 1 .
Table 1 Comparison of Truncation Windows
Window Mathe matic al Ex pr es sion Side Lob e 3dB
Name Level ( dB) Band width
Recta ngul ar 2
w( n) = 1 -13 0.89
Window
N
Tr iangul ar 2n N + 1 2
Window w( n) = 1 -27 1.28
N +1 N
Han ning 2n 2
w( n) = 0.51 cos -32 1.44
Window N 1 N

Hammi ng 2n 2
Window
w( n) = 0.54 0.46 cos -43 1.30
N 1 N
Bl ackman 2n 4n 2
Window
w( n) = 0.42 0.5 cos + 0.8 cos -58 1.68
N 1 N 1 N
As shown from the Table 1, there is a trade-off between the width of the main lobe and the
amplitudes of the side lobe. Although the Rectangular Window provides the best spectral resolution
with the narrowest 3dB bandwidth for the main lobe, it also gives the worst spectral masking effect.
By changing the truncation window, the spectral resolution is sacrificed for a smaller spectral
masking effect.
The spectral resolution of the Modified Periodogram Method depends of the truncation window.
The resolutions for the five windows are shown in Table 1. The spectral resolution is in the form:

[ ]
Res PModified (e j ) =
2
N
, where is a window-dependent constant.

Generalizing the truncation window in the Periodogram Method,

[ j
]1 j j
lim E PModified ( e ) = lim E W ( e ) Px (e ) .
N N
2
1
As long as the generalized truncation window in the frequency domain W (e j ) converges to an
2
impulse as N , the desirable condition

[ ]
lim E PModified ( e j ) = P(e j )
N

is satisfied and thus the Modified Periodogram Method is asymptotically unbiased.


The variance of the power spectrum estimated by the Modified Periodogram Method is
approximately the same as that for the Periodogram Method:

[ ]
Var PModified (e j ) P 2 ( e j )

In other words, the variance generally does not vanish even as the data length approaches infinity.

- 6 -
The complexity can be analyzed by inspecting the loop structure of the implementation. The
Modified Periodogram Method can be implemented as follows:
1) Compute the estimated autocorrelation
1 N 1
w(n) x(n)w(n + )x (n + ) which is an O(n 2 ) operation.
rModified ( ) =
N n= 0
2) Take the DFT of the estimated autocorrelation
2
1 N 1 j u
PModified (u ) = rx ( ) e N which is an O( n 2 ) operation.
N =0
Thus, the overall complexity of the Modified Periodogram Method is O( n 2 ) + O( n 2 ) = O(n 2 ) .

1.3 Bartletts Method


Both the Periodogram Method and Modified Periodogram Method do not give zero variances as the
data length approaches infinity. One way to enforce zero variance is by averaging. The Bartletts
Method divides the signal of length N into K segments, with each segment having length
L = N / K . The Periodogram Method is then applied to each of the K segments. The average of
the resulting estimated power spectra is taken as the estimated power spectrum of the Bartletts
Method. Mathematically, the Bartletts Method can be expressed as:
2
1 K 1 1 L1
PBartlett (e ) = x (n + iL ) e
j jn

K i= 0 L n= 0
144424443
periodogram of segment i

2
K 1 L 1
1
PBartlett(e j ) =
N x(n + iL )e
i = 0 n =0
jn

The spectral resolution of the estimated power spectrum PBartlett (e j ) is determined by the
spectral resolution of each segment which is of length L . Thus, by replacing N for L in the
expression in Table 1, the spectral resolution of the Bartletts Method is:

[ ]
Res PBartlett( e j ) = 0.89
2
L
, but since L = N / K

[ ]
Res PBartlett( e j ) = 0.89 K
2
N
.

Each of the K segments is simply the Periodogram estimate which is asymptotically unbiased.
Consequently, the estimated power spectrum PBartlett(e j ) , which is the average of the Periodogram
estimates of the K segments, is also asymptotically unbiased.
The variance of the estimated power spectrum PBartlett (e j ) is:

[ ]
1
[ 1
]
Var PBartlett (e j ) = Var P peri ( e j ) Px2 (e j ) .
K K

- 7 -
[ ]
Thus, as K , the variance Var PBartlett( e j ) approaches zero.
In summary, the value of K can be used to design the trade-off between spectral resolution and
variance. With a larger number of segments, the variance is reduced but at the expense of spectral
resolution. Conversely, with a smaller number of segments, the spectral resolution (i.e. the
bandwidth of the main lobe) is reduced at expense of larger variance.
The complexity can be analyzed by inspecting the loop structure of the implementation. The
Bartletts Method can be implemented as follows:
3) Estimate the power spectra of the K segments using the Periodogram Method. Since the
Periodogram Method is an O( n 2 ) operation (as described in Section 1.1.4) and is repeated
K times, the complexity of this step is n O( n 2 ) = O( n 3 ) .
4) Compute the average of the estimates of all the segments which is an O (n) operation.

Thus, the overall complexity of the Bartletts Method is O( n 3 ) + O(n ) = O( n 3 ) .

1.4 Welchs Method


The Welchs Method eliminates the trade off between spectral resolution and variance in the
Bartletts Method by allowing the segments to overlap. Furthermore, the truncation window can
also vary. Essentially, the Modified Periodogram Method (instead of the Periodogram Method) is
applied to each of the overlapping segments. Mathematically, the estimated power spectrum
PWelch (e j ) can be expressed as:
2
1 K 1 L 1
PWelch (e j ) =
KLU i= 0 n =0
w(n ) x (n + iD )e jn , where

K is the number of segments,


L is the length of each segment,
1 L1

2
U = w( n ) ,
L n= 0
D is the offset between two consecutive segments;
L D is the number of overlapped points.
The spectral resolution of the estimated power spectrum PWelch (e j ) is determined by the spectral
resolution of each segment which is of length L . It is window dependent. By replacing L for N
in the expression in Table 1,

[ ]
Res PWelch( e j ) =
2
L
, but since the total data length is N = L + D ( K 1) ,

[ ]
Res PWelch( e j ) =
2
N D( K 1)
, where is a window-dependent constant.

- 8 -
Each of the K segments is simply the Modified Periodogram estimate which is asymptotically
unbiased. Consequently, the estimated power spectrum PWelch (e j ) , which is the average of the
Periodogram estimates of the K segments, is also asymptotically unbiased.
The variance of the estimated power spectrum PWelch (e j ) is more difficult to compute since there
is correlation between overlapping segments. However, it is shown in [1] that with 50% overlap
( D 12 L ), the variance of the estimated power spectrum PWelch (e j ) is about:

[
Var PWelch (e j )
16 N
]
9 L 2 j
Px (e ) .

[ ]
Thus, as N , the variance Var PBartlett( e j ) approaches zero.
In summary, the length of each segment, L = N D ( K 1) , can be used to improve both the
spectral resolution and the variance. By allowing more overlapping (larger D ), more segments
(larger K) or longer segments (larger L ) can be obtained to improve both the spectral resolution
and the variance. However, overlapping introduces correlations between segments. In practice, the
amount of overlapping is typically between 50% to 75% [1].
The complexity can be analyzed by inspecting the loop structure of the implementation. The
Welchs Method can be implemented as follows:
1) Estimate the power spectra of the K segments using the Modified Periodogram Method. Since
the Modified Periodogram Method is an O( n 2 ) operation (as described in Section 1.2) and
is repeated K times, the complexity of this step is n O( n 2 ) = O( n 3 ) .
2) Compute the average of the estimates of all the segments which is an O (n) operation.

Thus, the overall complexity of the Welchs Method is O( n 3 ) + O(n ) = O( n 3 ) .

1.5 Blackman-Tukey Method


Recall that the estimated autocorrelation is:
N 1

x (n )x (n + ) .
1
rx ( ) =
N n =0

The estimate is worst when is largest because the average is computed for the smallest number of
terms. For example, consider the best and the worst case when = 0 and = N 1 ,
N 1
1 1 1 2 1 2
rx (0) =
N
x(n) x(n) = N x 2
(0) +
N
x (1) + K + x ( N 1)
N
n =0

1 0 1
rx ( N 1) =
N n =0
x ( n) x ( n + N 1) = x( 0)x( N 1) .
N
In the best case when = 0 , the autocorrelation is estimated based on the average of N terms. In
the worst case when = N 1 , the autocorrelation is estimated based on the average of only one
term. Thus, regardless how large N is, the variance of the estimated autocorrelation rx ( ) ,
contributed mainly from the estimates where is large (e.g. N 1 ), still exists. The Blackman-

- 9 -
Tukey Method tries to improve the estimated power spectrum by applying a window wBT ( ) to
rx ( ) before taking the DTFT. The purpose of the window wBT ( ) is to give smaller weights to
the contributions by the terms where is large (e.g. N 1 ).
Mathematically, the estimated power spectrum PBT ( e j ) can be expressed as:
M
j
PBT ( e ) = w BT ( ) rx ( )e
jk
,
k=M

where [ M , M ] is the support interval of the window wBT ( ) with M 12 N . A common and
simple window would be a rectangular window, wBT ( ) = wM ( ) , supported on interval [ M , M ] ,
where M 12 N . With this rectangular window wM ( ) , estimates for rx ( ) where > M are
ignored.
The spectral resolution of the estimated power spectrum PBT ( e j ) is window dependent. By
replacing M ( M ) = 2 M for N in the expression in Table 1,

[ ]
Res PBT ( e j ) =
2
2M
,

[ ]
Res PBT ( e j ) =
2
2 M
, where is a window-dependent constant and M 12 N .

[
Since M 12 N , Res PBT ( e j ) ]
2
N
[ ]
= Res PModified ( e j ) , the spectral resolution is worse than
the power spectrum estimated with the Modified Periodogram Method. The higher the M value, the
smaller the truncation of the estimated autocorrelation, and the better (smaller) the resolution is.
By taking the expected value and let N ,

[ ] 1
lim E PBT ( e j ) = lim E W BT (e j ) Px (e j )
N N
2
1 j j j
= lim E W BT ( e ) W ( e ) Px (e )
N
2
1
= lim E W BT ( e j ) Px ( e j ) , where W BT ( e j ) = W ( e j ) W ( e j )
N
2

As long as
1
2
[
W BT ( e j ) converges to an impulse as N , lim E PBT (e j ) = P( e j ) ,
N
]
allowing the power spectrum estimate to be asymptotically unbiased.
It is shown in [1] that the variance of the estimated power spectrum PBT ( e j ) is

[ ]
M
j 1
j
Var PBT ( e ) Px ( e )
2

N
w 2
BT (k ) , provided that N >> M >> 1 .
k = M

Thus, the higher the M value, the larger the summation term is, giving larger variance.

- 10 -
In summary, there is a trade off between spectral resolution and variance. By eliminating more
components of the estimated autocorrelation (i.e. smaller M ) a smaller variance is achieved at the
expensive of higher bandwidth (resolution).
The complexity can be analyzed by inspecting the loop structure of the implementation. The
Blackman-Tukey Method can be implemented as follows:
1) Compute the estimated autocorrelation
1 N 1
rBT ( ) = x (n)x (n + ) which is an O(n 2 ) operation.
N n= 0
2) Multiply the estimated autocorrelation with the window wBT ( ) , which is an O (n) operation.
3) Take the DFT of the estimated autocorrelation
N 1 2
1 j
rx ( )e
un
PBT (u ) = N
which is an O( n 2 ) operation.
N n =0

Thus, the overall complexity of the Blackman-Tukey Method is O( n 2 ) + O( n) + O( n 2 ) = O(n 2 ) .

- 11 -
2 Feature Extraction
A feature extraction can be thought as a mapping function that maps a region of interest (ROI) in
the image space to a numerical feature value in the feature space. Inspired by [3] and [4], the
linearized power spectrum of a one-dimensional scan-line of an image is used to extract two
features, namely the slope and the y-intercept, as described in details below.
2.1 Experiment Outline
Figure 2 shows an ultrasound image that was provided by the University of Western Ontario. It
shows the axial view of a prostate. Figure 3 shows the outline and the cancerous parts of the
prostate, identified preliminarily by Dr. Fenster at the University of Western Ontario. (Note: this
cancer identification is to be confirmed with a radiologist; it is used as preliminary input to the
algorithms discussed in this paper). The outlined and identified image was divided into 119 scan-
lines. These scan-lines are referred as regions of interest (ROIs) in this paper. The ROIs that are
completely contained in the cancerous part of the prostate are referred as cancerous ROIs. The
ROIs that are completely contained in the non-cancerous part of the prostate are referred as benign
ROIs in this paper. Other ROIs, such as those that are partly outside the outlined prostate and
those that span over both cancerous and benign parts of the prostate, are not included in the
analysis.

Figure 2 Prostate Ultrasound Image

Figure 3 Outlined and Identified Prostate Ultrasound Images

- 12 -
The power spectrum of each ROI was estimated with all five methods described in this paper. The
power spectrum was linearized by taking the logarithm: 20 log P ( e j ) . The linearized power
spectrum was approximated by a straight line with the minimum mean squared error (MMSE), as
shown in Figure 4. The slope and the y-intercept of the straight line were taken as two feature
values for the ROI.
4
x 10
2.5 50

0
2

-50
1.5

-100

1
-150

0.5
-200

0 -250
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 4 Linearization of Power Spectrum. Left: P ( e j ) . Right: 20 log P ( e j )

By applying this feature extraction to each pixel (and its 119 neighbourhood) of an ultrasound
image and normalizing the resulting feature values, a feature image can be constructed for each
feature. Figure 5 and Figure 6 show the feature images of the two features. Pixels with the highest
feature values (normalized to 1) in a feature image are coloured in white, while pixels with the lowest
feature values (normalized to 0) are coloured in black. Intermediate features values are coloured in
different grey levels accordingly. The feature images shown have been equalized.

Figure 5 Feature Image for Slope Feature

- 13 -
Figure 6 Feature Image for Y-Intercept Feature

2.2 Comparison of PSE Methods


There were 73 cancerous ROIs and 520 benign ROIs in the test image. Each ROI was one
dimensional and was 19 pixels long. The 73+520=593 ROIs were mapped to 593 feature values
using the Periodogram Method. The feature values were normalized to the interval (0,1). The
median feature value of the feature values of each class of each feature was used as the expected
feature value for the class. Table 2 shows the expected feature values of the two features for the
two classes.
Table 2 Expected Feature Values
Sl ope Y -Inter ce pt
Cancer ou s 0.1267 0.3211
Benign 0.1428 0.3721
With the expected feature values, the mean-squared-error (MSE) can be calculated for the different
PSE methods. The MSE of each method for each feature was calculated as a weighted average of
the MSE of the two classes, so that the MSE of cancerous class (i.e. undetected cancer) and the
MSE of the benign class (i.e. false alarm) contribute to the overall MSE equally. In other words,
although the cancerous part is normally only a small part of the prostate, the error of undetected
cancer and the error of false alarm have the same impact to the overall MSE.
1

2
MSEcancerous = actualVF ( x, y ) expectedVFcancerous ( x, y )
Cancerous ( x, y )Cancerous
1
actualVF ( x , y ) expectedVFbenign ( x, y)
2
MSEbenign =
Benign ( x , y )Benign
Area benign Area cancerous
MSE = MSE cancerous + MSEbenign
Area total Areatotal
where
( x, y) is a pixel in the feature image;
Cancerous is the set of pixels in the identified cancerous part of the prostate;

- 14 -
Benign is the set of pixels in the identified benign part of the prostate;

Cancerous is the number of pixels in the set Cancerous ;

Benign is the number of pixels in the set Benign ;


actualVF ( x, y) is the feature value of the pixel ( x, y) in the feature image;
expectedVF( x , y ) is the expected feature value shown in Table 2.
The first term of the expression above represents the weighted MSE of undetected cancer. The
second term represents the weighted MSE of falsely detected cancer.
Table 3 and Table 4 show the MSE for the different PSE methods of the two features. Table 5
shows the interpretation of the two tables, pointing that the Welchs Method can be thought as a
generalization of the Periodogram Method (no overlap, rectangular window), the Modified
Periodogram Method (no overlap, non-rectangular window), and the Bartletts Method (non-zero
overlap, rectangular window).
Table 3 MSE for Slope Feature
% Over l ap : 0% 50%
Segment Le ngt h ( % D ata Le ngt h) : 100% 50% 25% 50% 25%
Wel ch Recta ngul ar 0.0149 0.0155 0.0320 0.0142 0.0457
Tr iangul ar 0.0152 0.0154 0.0344 0.0305 0.0385
Han ning 0.0281 0.0233 0.0276 0.0230 0.0251
Hammi ng 0.0265 0.0255 0.0284 0.0293 0.0324
Bl ackman 0.0264 0.0225 0.0326 0.0225 0.0343
Bl ackman -Tuk ey 100% 0.0260
75% 0.0227
50% 0.0256
Table 4 MSE for Y-Intercept Feature
% Over l ap : 0% 50%
Segment Le ngt h ( % D ata Le ngt h) : 100% 50% 25% 50% 25%
Wel ch Recta ngul ar 0.0906 0.0911 0.1019 0.0929 0.1036
Tr iangul ar 0.0908 0.0913 0.1003 0.0982 0.1002
Han ning 0.1003 0.0988 0.0969 0.0987 0.0972
Hammi ng 0.0994 0.0995 0.0991 0.0994 0.0980
Bl ackman 0.0985 0.0959 0.1004 0.0965 0.0996
Bl ackman -Tuk ey 100% 0.1001
75% 0.1016
50% 0.1017

Table 5 Interpretation of Table 3 and Table 4


% Over l ap : 0% 50%
Segment Le ngt h ( % D ata Le ngt h) : 100% 50% 25% 50% 25%
Wel ch Recta ngul ar Periodogram Bartlett Bartlett Welch Welch
Tr iangul ar Modified Per. Welch Welch Welch Welch
Han ning Modified Per. Welch Welch Welch Welch
Hammi ng Modified Per. Welch Welch Welch Welch
Bl ackman Modified Per. Welch Welch Welch Welch
Bl ackman -Tuk ey 100% Blackman-T.
75% Blackman-T.
50% Blackman-T.

- 15 -
The lowest MSE of each method, shown in bold in Table 3 and Table 4, is shown again in Table 6
for comparison. The Blackman-Tukey Method provides the largest MSEs. The MSEs of all other
four methods are within 5% from each other. Thus, with the exception of the Blackman-Tukey
Method which actually makes the estimate worse, the different estimation methods do not affect the
performance of the feature extraction much. This is expected because this particular feature
extraction depends mainly on the overall shape of the power spectrum, described by the slope and
the y-intercept, not on the local characteristics of the power spectrum such as spectral resolution
and local fluctuations (or variance).
Table 6 MSE Comparisons
Per iodo. Modified Per iodo . Bar tl ett Wel ch Bl ackman -Tuk ey
Sl ope MSE 0.0149 0.0152 0.0155 0.0142 0.0227
% Diff. fr om Per iodo. 0.00% 1.82% 3.64% -4.24% 47.27%
Y -Int MSE 0.0906 0.0908 0.0911 0.0913 0.1001
% Diff. fr om Per iodo. 0.00% 0.22% 0.54% 0.75% 10.24%

Since the four methods do not affect the performance of the feature extraction much, their
complexities would be the critical factor in choosing the best method. The Periodogram Method
and the Modified Periodogram Method both have complexity O( n 2 ) . The Bartletts Method and
the Welchs Method both have complexity O( n 3 ) . Since the rectangular window of the
Periodogram Method is simply a constant 1, it is less complex than the Modified Periodogram
Method in general. Furthermore, as shown in Table 3 and Table 4, the Periodogram Method
provides the slightly better estimates (with smaller MSEs) than the Modified Periodogram Method.
Consequently, the Periodogram Method is the best method for this type of feature extraction
because of its accuracy and its simplicity (i.e. efficiency).
2.3 Cancer Detection
In this section, the two features, namely the slope and the y-intercept of the linearized power
spectrum, were used to detect the cancerous part of the prostate.
All methods discussed in this paper estimate the power spectrum by first estimating the
autocorrelation which is a second-order-statistic. Hence, there is a relationship between the power
spectrum and the image visual characteristics. In particular, the slope and the y-intercept of the
linearized power spectrum can be related to the texture complexity and the brightness of a ROI.
Since the power spectrum is proportional to the squared of the DTFT of a signal (see Section 1.1), a
ROI with higher frequency components tends to make the slope smaller. In other words, a ROI
with high texture complexity tends to have smaller feature value for the slope feature. Furthermore,
the estimated power spectrum is direction dependent. For example, a horizontal ROI with small
feature value for the slope suggests that there is high texture complexity in the horizontal direction.
Since the y-intercept is proportional to the DC component of a signal, a ROI with lower overall
grey-level tends to make the y-intercept smaller. In other words, a ROI that is darker overall tends
to have smaller feature value for the y-intercept feature.
A careful inspection of the seven ultrasound images obtained from the University of Western
Ontario reveals that there are three characteristics of the cancerous part of a prostate:
1) The cancerous part is relatively darker than most benign part of the prostate. Thus, a smaller
feature value is expected for the slope feature.

- 16 -
2) There are relatively more granule patterns in the cancerous part of the prostate. Thus, a smaller
feature value is expected for the y-intercept feature.
3) Cancer is more likely to be found in the part of the prostate that is closer to the anus (bottom of
the image). In this part of the prostate, the granule patterns look more vertical than
horizontal, making the texture complexity higher in the horizontal direction than in the
vertical direction. Thus, it is expected that the y-intercept feature in the horizontal direction
can help detecting cancer better.
Thus, in order to detect cancer effectively, the slope feature and the y-intercept (for the horizontal
direction) feature must be used jointly. This requires transforming the feature values so that the
ideal cancerous feature value is logic-TRUE and the ideal benign feature value is logic-FALSE, then
applying a logic-AND operator to the two features. The numerical values 1 and 0 can be used for
logic-TRUE and logic-FALSE. The numerical multiplication can be used to simulate the logic-
AND operator. A simple, but effective, approach was applied using Photoshop for two ultrasound
images. Figure 9 and Figure 12 show the resulting images, which were processed as follows:
1) Blur the two feature images with a Gaussian blur (with variance of 4 pixels). This is justified by
the assumption that no single pixel is cancerous by itself.
2) Equalize the two blurred images so there is a high contrast.
3) Invert the two equalized images so that cancerous feature value is about 1 (as opposed to 0).
4) Multiply the two inverted images together to simulate logic-AND.
5) Equalize the resulting image so there is a high contrast.
6) Invert the image back so the cancerous part is shown as black.
7) Apply thresholding with a threshold value that is experimentally determined.

Although the results (Figure 9 and Figure 12 ) do not look identical to the expected images (Figure 7
and Figure 10), they provide good indication of the existence and approximate location of the
cancerous parts of the prostates.

- 17 -
Figure 7 Cancerous part identified by Dr. Fenster

Figure 8 Feature Images. Left: slope; Right: y-intercept

Figure 9 Processed Images. Left: before thresholding; Right: after thresholding (33/255)

- 18 -
Figure 10 Cancerous part identified by Dr. Fenster

Figure 11 Feature Images. Left: slope; Right: y-intercept

Figure 12 Processed Images. Left: before thresholding; Right: after thresholding (95/255)

- 19 -
3 Conclusions
Five non-parametric power spectrum estimation methods were analyzed in this paper. They all have
pros and cons, with trade-offs between spectral resolution, variance, and complexity.
The feature extraction discussed in this paper depends mainly on the overall shape, described by the
slope and the y-intercept, of the linearized power spectrum. Since it does not closely depend on the
local characteristics of the power spectrum, such as spectral resolution and variance, the complexity
becomes the critical factor in choosing the best estimation method for this feature extraction. It was
found that the Periodogram Method is the simplest and one of the most accurate estimation method
for this feature extraction.
Although the different PSE methods do not affect the results much in this particular feature
extraction using linearized power spectrum, the methods can still make great difference in
applications that are sensitive to spectral resolution and variance, such as estimations of signal and
noise variances for the Wiener Filter.
The texture and the overall brightness of a ROI are reflected on the slope and the y-intercept of the
linearized power spectrum. By using these two features jointly, the cancerous part of the prostate
can be detected quite successfully.

- 20 -
References
[1] M. H. Hayes, Statistical Digital Signal Processing and Modeling, John Wiley & Sons Inc.,
1996.
[2] F. Attivissimo, M. Savino, A. Trotta, Power Spectral Density Estimation via Overlapping
Nonlinear Averaging, IEEE Transcations on Instrumentation and Measurement, Vol. 50,
No. 5, October 2001.
[3] F. Lizzi, "Ultrasonic Scatterer-Property Images of the Eye and Prostate", IEEE Ultrasonics
Symposium, pages 1109 to 1117.
[4] E. Feleppa, A. Kalisz, J. Sokil-Melgar, L. Lizzi, T. Liu, A. Rosado, M. Shao, W. Fair, Y. Wang,
M. Cookson, V. Reuter, W. Heston, Typing of Prostate Tissue by Ultrasonic Spectrum
Analysis, IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, Vol.
43, No. 4, July 1996.
[5] Power Spectrum Estimation - Lecture 15, http://www.cis.rit.edu/class/simg713/
Lectures/Lecture713-15.pdf, Spring 2002.
[6] Lecture 21 - Spectrum Estimation,
http://robotics.eecs.berkeley.edu/~elghaoui/ee225a/lect21.pdf, Department of Electrical
Engineering and Computer Sciences, University of California at Berkeley, Spring 2000.
[7] A. Schiffler, "Power Spectrum Analysis using Spectral Estimators",
http://unser.ferzkopp.net/Personal/Thesis/node26.html, October 1996.

- 21 -

Vous aimerez peut-être aussi