MIT OpenCourseWare

http://ocw.mit.edu

HST.582J / 6.555J / 16.456J Biomedical Signal and Image Processing

Spring 2007

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

Harvard-MIT Division of Health Sciences and Technology

HST.582J: Biomedical Signal and Image Processing, Spring 2007
Course Director: Dr. Julie Greenberg

HST-582J/6.555J/16.456J Biomedical Signal and Image Processing Spring 2007

Problem Set Y
QUIZ 2 will take place on Thursday, May 10, from 9:30-11 am in 56-154
(usual time and place).
The quiz will be closed book. Two 8 12 11 inch sheet of notes (both
sides) will be allowed.
The quiz will cover material presented in lectures from March 13 (PDFs)
thru April 26 (Random Signals II). Note that there is some overlap with
the image processing topics covered on Quiz 1.
Coverage of topics on the quiz will be somewhat representative of the
amount of time spent on each topic in lectures, labs and problem sets.
You are not responsible for material in the course notes that was not
covered elsewhere in the course. Please see the annotated outline for
listings of specic topics.
This ungraded problem set, which includes questions from with Quiz
2 from 2005 and 2006, will help you prepare for the quiz in two ways:
First, it gives you a chance to work with recently presented material that
was not covered by previous problem sets. Second, it illustrates the type
of questions asked on quizzes in prior years.
This problem set will not be collected or graded. Solutions will be posted
on the course website on Monday, May 7.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 1
x[n] is a stationary random signal. Each sample value of x[n] is independent and uniformly
distributed between +1 and 1. y[n] is the result of taking the rst dierence of x[n], that
is,
1
y[n] = x[n] x[n 1].
2
(a) Compute and sketch the autocorrelation functions Rx [k] and Ry [k].

(b)
y[n] is Gaussian

true / false (circle one)

y[n] is stationary

true / false (circle one)

y[n] is zero-mean

true / false (circle one)

Sy (0) = 2
(Sy (f ) denotes the power spectrum of y[n].)

true / false (circle one)

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 2

(A)

(B)

(C)

(D)

Identify the gure that corresponds to each description below:

(a) original image

A/B/C/D

(b) image after homomorphic ltering

(c) result of edge detection

(circle one)
A/B/C/D

A/B/C/D

(d) result of histogram equalization

(circle one)

(circle one)

A/B/C/D

(circle one)

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 3

2
3

(A)

1
(B)

2
3

(C)

1
(D)

(a) Indicate whether or not each gure above represents a valid probability density function
(PDF).
(A) valid / invalid

(circle one)

if invalid, state reason why:

(B) valid / invalid

(circle one)

if invalid, state reason why:

(C) valid / invalid

(circle one)

if invalid, state reason why:

(D) valid / invalid

(circle one)

if invalid, state reason why:

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

1 2
x
3

2
x
3

1
3

(b) Recall that the cumulative distribution function (CDF) of the random variable X is
dened as the probability that X x. The gure above is the cumulative distribution
function corresponding to which of the PDFs in part (a)?
A/B/C/D

(circle one)

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 4
(a) Consider the discrete-time random process, z[n], with
E {z[n]} = 0

E z[n]2

= z2

E {z[10]z[11]} = (1 1/4)z2
E {z[12]z[13]} = (1 1/4)z2
Which of the following is a correct statement: (circle one)
z [n] is a wide sense stationary process.
z [n] is not a wide sense stationary process.
z [n] might be a wide sense stationary process.
(b) Consider the discrete-time random process, y[n]. The probability density function (PDF)
of y[n] is Gaussian with zero mean and variance, 2 , for n = , , . Additionally, the
joint PDF of y[10] and y[11] is jointly Gaussian with zero mean and covariance:

E{ [y[10] y[11]]

[y[10] y[11]]} =

2 0
0 2

while the PDF of y[12] and y[13] is jointly Gaussian with zero mean and covariance:

E{ [y[12] y[13]]

[y[12] y[13]]} =

2
2
2

2
2
2

Which of the following is a correct statement: (circle one)

y [n] is a strict sense stationary process.
y [n] is not a strict sense stationary process.
y [n] might be a strict sense stationary process.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

(c) Consider the discrete-time random process, x[n], where samples are drawn alternately
from either a Gaussian probability density function or a uniform probability density function,
so that x[n] is sampled from a Gaussian PDF when n is even and x[n] is sampled from a
uniform PDF when n is odd. Both the Gaussian and uniform PDFs have zero mean and
variance, 2 .
Which of the following are correct statements: (circle all that apply)
x[n] is a strict sense stationary process.
x[n] is not a strict sense stationary process.
x[n] might be a strict sense stationary process.
x[n] is a wide sense stationary process.
x[n] is not a wide sense stationary process.
x[n] might be a wide sense stationary process.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 5 Problem 1
Figure 1 depicts a system block diagram in which
H1 (f ) and H2 (f ) are stable linear shift invariant lters,
Both H1 (f ) > 0 and H2 (f ) > 0 for all f ,
x[n] is a zero-mean white noise sequence with < x[n]2 >= 3,
w[n] is a zero-mean white noise sequence with < w[n]2 >= 2,
x[n] and w[n] are uncorrelated with each other, and
< z[n]2 >= 8.
Note that you have already worked out some details of this specic diagram from the handout
of last years Quiz 2.
a) Give expressions for the following (in terms of the signals and lters in the diagram):
Rxy [m] =
Ry [m] =
Rxz [m] =
Rz [m] =
Sxy (f ) =
Syx (f ) =
Sxz (f ) =
Sy (f ) =

Sz (f ) =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Question 6

Figure 2: Iso-probability contours

It has been discovered that an individuals blood contains one (and only one) of ve antibody
types denoted T0 , T1 , T2 , T3 , and T4 . We cannot measure the presence of these directly, but
we have two blood tests, denoted X1 and X2 , which yield numerical values, x1 and x2 ,
according to the conditional probability density functions:

pX1 ,X2 (x1 , x2 |T = Ti ) = N

x1
x2

; Mi , i

where N (x; Mi , i ) is a 2-dimensional Gaussian density function with mean, Mi , and co

variance, i . The class-conditional means and covariances, denoted by their index are:

0 = 1 = 2 = 3 = 4 =

M0 =

1.5
1.5

M1 =

4
1.5

M2 =

1.5
4

1 0
0 1

M3 =

4
4

M4 =

6.5
4

Iso-probability contours for each of the conditional densities are shown in Figure 4 for your
reference. This gure will be repeated for each part of this question in the event that you
wish to mark it.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

In the problems that follow we will consider subgroups (i.e. classes) of the population at

large which contain individuals with dierent antibody types. In all cases, the prior

probability of each class is 0.5.

Figure 3: Iso-probability contours

(a) AB-Classier: Consider the population where each individual has either the T0 , T1 ,
T2 , or T3 antibody, with equal probability. In other words, 25% of the population has T0 ,
25% has T1 , 25% has T2 , 25% has T3 , and no one has T4 . We divide this population into
two classes. Class A individuals have either the T0 antibody or the T1 antibody, while class
B individuals have either the T2 antibody or the T3 antibody. Which tests are necessary
to achieve a minimum classication error rate when distinguishing class A from class B?
(circle one)
X1 only
X2 only
both X1 and X2
neither test is necessary

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Figure 4: Iso-probability contours

(b) CD-Classier: Again, consider the population where each individual has either the T0 ,
T1 , T2 , or T3 antibody, with equal probability. In other words, 25% of the population has
T0 , 25% has T1 , 25% has T2 , 25% has T3 , and no one has T4 . We divide this population into
two classes. Class C individuals have either the T1 antibody or the T2 antibody, while class
D individuals have either the T0 antibody or the T3 antibody. Which tests are necessary
to achieve a minimum classication error rate when distinguishing class C from class D?
(circle one)
X1 only
X2 only
both X1 and X2
neither test is necessary

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

Figure 5: Iso-probability contours

(c) AE-Classier part 1: Now consider the population where each individual may have
any of the antibodies, T0 through T4 , but with unequal probabilities: 25% of the population
has T0 , 25% has T1 , 20% has T2 , 20% has T3 , and 10% has T4 . We divide this population into
two classes. Class A individuals have either the T0 or T1 antibody, as before, while class E
individuals have either the T2 , T3 , or T4 antibody. (Note that the prior probabilites of class
A and class E are each 0.5.) We use the classication rule for the AB-classier (from part
(a) of this problem) to classify individuals as class A or class E, so that those individuals
which would be labeled as class B will now be labeled as class E). This classier will result
in a classication error rate which is: (circle one)
higher than the AB-classication error rate.
lower than the AB-classication error rate.
the same as the AB-classication error rate.
(d) AE-Classier part 2: Using the classes A and E as dened in part (c), the minimum
achievable classication error rate for classes A or class E as dened above is: (circle one)
higher than the AB-classication error rate.
lower than the AB-classication error rate.
the same as the AB-classication error rate.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

NAME:

HST-582J/6.555J/16.456J

Biomedical Signal and Image Processing

Spring 2006

Quiz 2 May 11, 2006

There are 4 questions on 9 pages.
Please write all of your answers and explanations directly on this quiz
form. NO credit will be given for anything written on scratch paper.
Explanations should be brief and to the point. Unnecessarily long or
wordy justications may be penalized.

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

p1 (x)

p2 (x)

1
2

1
2

1
(a)

1
(b)

Figure 1: PDFs for Question 1. Note that gures are not necessarily to scale.

Question 1 (20 %)
Figures 1(a) and 1(b) depict probability density functions.(Your answers to parts (b)-(d) must
be consistent with your answer to part(a))
(a) Determine constants and such that p1 (x) and p2 (x) are valid probability density
functions.
=

(b) The PDF of the random variable z is:

1
1
p(z) = p1 (z) + p2 (z)
2
2
Compute the expected value of z.
E {z} =

(c) Let y = 12 x1 + 12 x2 , where

x1 p1 (x) x2 p2 (x)

and

p(x1 , x2 ) = p1 (x1 )p2 (x2 )

Compute the expected value of Y .

E {y} =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

(d) Sketch the cumulative distribution function corresponding to the probability density
function, p2 (x), over the range 12 < x < 23 . Clearly, label important details in your plot.

P2 (x)

12

1
2

3
2

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

p0 (x)

p1 (x)

1
4

1
2

3
4

1
4

(a)

1
2

3
4

(b)

Figure 2: PDFs for Question 2. Note that gures are not necessarily to scale.

Question 2 (30 %)
Figures 2(a) and 2(b) depict class-conditional probability density functions.
This question concerns the binary hypothesis test where:
H0 : x p0 (x)

H1 : x p1 (x)

The prior probabilities of hypotheses H0 and H1 are denoted P0 and P1 respectively.

For parts (c)-(e) dene:
PF A = Pr {choosing H1 |H0 is true}

PD = Pr {choosing H1 |H1 is true}

(a) Determine the minimum value of P0 such that the decision rule would be to always assign
a measurement x to H0 .
min P0 =
(b) Determine the minimum value of P1 such that the decision rule would be to always assign
a measurement x to H1 .
min P1 =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

(c) Given P1 = 12 , a 1-nearest neighbor classier rule based on a set of labeled samples results
in the following decision rule:

decision rule
choose H1
choose H0
1
4

1
2

3
4

For this decision rule compute:

PF A =

PD =

PE =

(d) Consider the decision rule that is the reverse of part (c). That is, it chooses H0 when the
decision rule of part (c) chooses H1 and chooses H1 when the decision rule of part (c)
chooses H0 , compute:
PF A =

PD =

PE =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

(e) On the plot provided, sketch the optimal decision rule. Assume P1 =
1

. Be sure to clearly
2

label all points.

decision rule
choose H1
choose H0
1
For the optimal decision rule compute:
PF A =

PD =

PE =

(f) On the plot provided, sketch the posterior probability of H0 as a function of x. Assume
P1 = 1

. Be sure to clearly label all points.

2

P r {H0 |x}

1
1
2

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

y1 [n]

y2 [n]

w1 [n]

w2 [n]

x[n]

yN [n]

wN [n]

Figure 3: System Block Diagram for Question 3.

Question 3 (20 %)
Figures 3 depicts a system block diagram for the set of random processes for which:
yi [n] = x[n] +

wk [n]

k=1

where
x[n] is a zero-mean WSS random process with autocorrelation function Rx [m].
wi [n]s are zero-mean white noise sequences with < wi [n]2 >= 12 .
wi [n]s are uncorrelated with each other and with x[n].
Using the information given, nd expressions for the following:

Rxy2 [m] =
RxyN [m] =
Ry2 [m] =
RyN [m] =

(y10 [n]

y5 [n])2 =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

y[n]

L(f )

r[n]

G(f )

v[n]

H(f )

x[n]

K(f )

v[n] r[n]

z[n]

M (f )

t[n]

w[n]

y[n] z[n]

Figure 4: System Block Diagram for Question 4.

Question 4 (30 %)
All parts of Question 4 refer to the system block diagram of Figure 4.
H(f ), G(f ) are stable linear shift-invariant lters.
H(f ), G(f ) are real-valued with H(f ), G(f ) > 0 for all f .
H(f ) = H(f ), G(f ) = G(f ).
y[n] is a zero-mean WSS random process with autocorrelation function Ry [m].
w[n] is a zero-mean white noise sequence with < w[n]2 >= 2.
y[n] and w[n] are uncorrelated with each other.
(a) Using the information given, determine L(f ) such that < (r[n] v[n])2 > is minimized.

L(f ) =

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].

(b) Using the information given, determine K(f ) such that < (y[n] z[n])2 > is minimized.

K(f ) =

(c) Given the optimal K(f ) from part (b), suppose you wish to design M (f ) to minimize
< (v[n] t[n])2 >. Will M (f ) differ from L(f ) designed in part (a)?

(circle one)

YES

NO

Explain your answer:

Cite as: Julie Greenberg, Bertrand Delgutte, William (Sandy) Wells, John Fisher, and Gari Clifford. Course materials
for HST.582J / 6.555J / 16.456J, Biomedical Signal and Image Processing, Spring 2007. MIT OpenCourseWare
(http://ocw.mit.edu), Massachusetts Institute of Technology. Downloaded on [DD Month YYYY].