Course note based on:

Medical Imaging Signals and Systems,

J.L. Prince and J.M. Links,
Pearson Prentice Hall, 2006.

Chapter 3

Imaging Quality

1
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Purpose of Medical Imaging

Primary purpose of a medical imaging system is to create

images of internal anatomical structures and physiological
functions of human body for :

• Diagnosing abnormal conditions

• Investigating causes of abnormal conditions
• Guiding therapeutic procedures
• Monitoring effectiveness of treatment

2
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Imaging Quality
Imaging quality can be measured by:

(1) Contrast: Difference between interested object and background

(2) Resolution: Details of object

(3) Noise: Random fluctuations in imaging intensity

(4) Artifacts: Invalid object or characteristics

(5) Distortion: Incorrect shape, size, position (geometric characteristics)

(6) Accuracy: Conformity to truth and clinical utility

How to evaluate these factors quantitatively and systematically?

3
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Modulation
Modulation mf of periodic signal f(x,y) is defined by:

( f max  f min ) / 2 f max  f min

mf   , (0  m f  1)
( f max  f min ) / 2 f max  f min

f(x,y)
Assuming f ( x, y )  A  B sin( 2u0 x),
(sinusoidal signal) fmax
B

B
mf 
fmin
A
A
x

Using a periodic signal and its modulation is an effective way to

quantify contrast. How it works?
4
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Modulation and Contrast
No contrast
B
mf  (fmax=fmin)
A
A

(0  m f  1)

f(x,y)
fmax
B
mf=0
fmin
A

x
Max contrast
(fmin=0)

Non-zero background intensity (fmin) reduces image contrast!

5
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Modulation Transfer Function (MTF)

LSI

Input δ(x,y) System S Output PSF, h(x,y)

• How PSF h(x,y) affects the contrast (modulation)?

• What is the relationship between mf of input f(x,y) and mg of output g(x,y)?

6
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Modulation of Output Signal
Assuming that f(x,y) is input signal to LSI system with circular symmetric PSF
h(x,y), and g(x,y) is the output signal.
f ( x, y )  A   Bk sin(2uk x  2vk y ),
k
 
g ( x, y )    f ( , )h( x   , y   )dd
  e j  e  j
sin  
    2j

    

A  
k
Bk sin( 2u k   2v k )  h( x   , y   )dd

 AH (0,0)   Bk H (uk , vk ) sin(2uk x  2vk y )
k

Modulation mg,k of output

signal g(x,y) at (uk, vk):
g max  g min
mg ,k 
g max  g min
Bk H (uk , vk )

AH (0,0)
H (uk , vk )
 m f ,k
H (0,0)
7
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 MTF and Contrast
MTF: Frequency response of the medical imaging system

Modulation transfer function (MTF):

mg H (u, v)
MTF (u, v)  
mf H (0,0)
For a typical medical imaging system,
amplitude of signals (spectrum)
decreases as frequency increases:

0  MTF (u, v)  MTF (0,0)  1

MTF also affects resolution!

8
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Local Contrast
Target signal intensity ft (object of interest, e.g., a tumor in the liver) need to be
identified from background signal intensity fb (e.g., the surrounding areas).

ft  fb
Local contrast is defined by: C
fb

Note: ft may be larger or smaller than fb.

9
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Resolution
PSF

Can two distinct objects be

distinguished in separate?

Full width at half maximum (FWHM)

= Width of the PSF at one-half of its
maximum value, which gives the minimum
distance where two points are imaged to
be separated.

10
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Resolution and PSF

Input Output

X-ray
Imaging
system

Image display Functional plot

Two pinholes

11
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Resolution and MTF
When f ( x, y )  B sin( 2ux), f(x,y) 1/u

g ( x, y )  B H (u ,0) sin( 2ux) x

 MTF (u ) H (0,0) B sin( 2ux)

If MTF(u) ≠ 0 at u ≤ uc, then separation g(x,y)

1/u
between two adjacent maxima of g(x,y) is
1/u as well. x

If MTF(u) = 0 at u ≥ uc, then g(x,y) =0.

MTF(u)
Thus, the resolution of the system is 1/uc,
where uc is the cut-off frequency of the system.

Spatial resolution and image contrast are tightly associated.

• PSF and LSF are measures of resolution. u
• MTF is a measure of contrast, which is obtained from uc
Fourier transform of PSF and LSF 12
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Line Spread Function
Assuming:
• An LSI system having isotropic PSF h(x,y) normalized to 1.
• Input signal is line impulse (vertical line l passing through origin): f(x,y)=δl(x,y)= δ(x)
 
g ( x, y )    h( , ) f ( x   , y   )dd
 

    h( , ) ( x   )d d

 

 
   

h( x)   h( ) ( x   )d

  h( x, )d


 ( x ) Line spread function (LSF)

Since h(x,y) is isotropic and normalized to1,
  

( x)dx   
 
h( x, )ddx  1
Then, L(u )  F1D {l ( x)}
  
   ( x )e  2ux
dx    h( x, )e  2ux ddx  H (u ,0)
  

LSF is adequate for determining the PSF of system.

13
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Line Spread Function (LSF)

LSF of the system

Line impulse

Imaging
system

Input Output

In practice, line-impulse phantoms are often

easier to fabricate than point-impulse phantoms.

14
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Resolution Tool (Phantom)

Line pairs per

millimeter

(Source: GE Healthcare Accessories and supplies)

15
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Typical Special Resolution

16
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Noise
Noise:
• Random fluctuation in image

Source:
• Depends on physics and instrumentation of imaging modalities
Ex.
• X-ray projection image: quantum mottle
• Nuclear medicine: random nature of radioactive
emissions (Poisson random variable)
• MRI: natural thermal vibration
• Ultrasound: scattering echo (speckle)

17
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Poisson Random Variable

When discrete random variable N is Poisson random variable, taking a

discrete values 0, 1, 2…, probability mass function (PMF), Pr[N=k], is
given by:

a k a a=1
Pr[ N  k ]  e
k!
for k= 0,1,2… Pr a=4

Mean of N: N  a a=10
Variance of N:  N2  a
k

Photon number counted in radiographic and nuclear medicine

imaging is Poisson random variable.
18
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Poisson Process in X-ray Imaging
In x-ray imaging, photons arrive at a detector, which is referred as Poisson process.

Probability mass function (PMF) of N(t), which is the number of photons arrive at
a detector in time t and Poisson random variable, is given by:
(  t ) k  t
Pr[ N (t )  k ]  e
k!
where  is the average arrival rate of the x-ray photons. Therefore,
Mean of N(t):  N  t
Variance of N(t):  N2  t

Example:
What is the probability that there is no photon detected in time t?
(  t ) 0  t
Pr[ N (t )  0]  e  e  t
0!
19
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Signal-to-Noise Ratio
Amplitude( f ) f: signal
Amplitude SNR: SNRa 
Amplitude( N ) N: noise

• Definition of SNR is case-dependent

• Need to specify what signals and noises are.

Decibel:
 Amplitude( f ) 
SNRa (dB)  20  log10  
 Amplitude( N ) 
 Power ( f ) 
SNR p (dB)  10  log10  
 Power ( N ) 

20
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 SNR in X-ray Image
Recall: the number of x-ray photons G detected by a detector follows Poisson
distributions.

In such a system, we can consider the signal to be the mean of G and noise to
be the standard deviation of G.

What is the amplitude SNR of this system?

For Poisson random variable G, the mean (μG) and variance (σG2) are μ.

G 
SNRa    
G 

Higher x-ray exposure improves the image quality.

But… more risk of radiogenic cancer.
21
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Artifacts
Invalid presentation of anatomical or functional objects

Mirror Artifact Speed of Sound Artifact Artifact due to refraction

Direct Echo
Liver

Fatty Lesion
Mirror Echo
Diaphragm

Mirror Image 5.8 % smaller

Ultrasound imaging
22
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Distortion
Inaccurate impression of shape, size and/or position of objects of interest

Size distortion due to magnification Shape distortion due to unequal

caused by different distances magnification caused by divergence
between the target and x-ray source of the x-ray beam

X-ray projection radiography

23
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Accuracy
Purposes of medical images:

• Diagnosis
• Prognosis
• Treatment planning
• Treatment monitoring
Accuracy:

• Conformity to truth
• Clinical utility
Quantitative accuracy
• Tumor dimensions from radiograph
• Glucose metabolic rate from nuclear medicine… etc.

Sources of error for quantitative accuracy

• Bias: systematic and reproducible error
• Imprecision: random and non-reproducible error
24
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)
 Summary
Image quality refers to the degree to which an image allows a medical
professional to accomplish the clinical goals. The following six factors
influence image quality.

• Contrast: Difference in image intensity of target and background.

• Resolution: Ability to distinguish two signals that differ in space, time or

frequency domain.

• Noise: Random fluctuation in an image.

• Artifacts: False signals that do not represent valid structural or functional signal
in the patient.

• Distortion: Geometrid inaccuracy in size or shape.

• Quantitative accuracy: Accuracy, comparison with the truth, of numerical values

obtained from images; diagnostic accuracy which is referred to accuracy of
interpretation and conclusions about disease obtained from images.
25
BIOM5200 / BMG5200 / SYSC5304 / ELG5127 - Biomedical Image Processing (Y. Ono, Carleton University, Ottawa, ON)