Professional Documents
Culture Documents
6.1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6.2
6.2
6.4
6.4
6.6
6.6
6.7
6.7
6.7
6.9
6.10
6.10
6.11
6.11
6.2
Introduction
The previous two chapters considered spatial resolution variations as a result of the source and detector.
The other primary determinants of image quality are noise, artifacts, and contrast.
Noise
Noise means random phenomena. In X-ray radiography there are several random processes.
The number of photons that leave the source (Poisson)
The number of those photons that pass unaffected through the object (Binomial)
The number of those photons that are captured by the detector (Binomial)
The number of light photons generated per captured X-ray photon (Binomial, approximately, with E-dependent mean)
The response of film grains to light photons (Binomial)
Because Binomial selection of a Poisson process yields another Poisson process (to be shown), the overall noise distribution is
approximately Poisson in X-ray imaging.
In X-ray imaging, each quanta carries much more energy h = hc/ than in visible light, so there are many fewer quanta, and
these quanta are the main source of variability between experiments.
Noise is most meaningful if it is put in perspective by comparing to signal strength. In X-ray radiography, the relevant comparison
is to contrast.
Artifacts
It is important to distinguish random noise effects from image artifacts. Although artifacts can appear random, such as the
smearing from the tails of the PSF in ultrasound beamforming, usually the term artifacts refers to undesirable image features that
occur essentially the same way even if a given scan is repeated (under identical conditions).
Contrast
The absolute contrast due to a small structure of interest is the change in intensity caused by that structure (i.e., the change
compared to if the structure had the same attenuation coefficient as the surrounding material).
We will be most interested in the relative contrast C defined by
C,
I
,
I
where I is the average background intensity (in the vicinity of the structure of interest) and I is the change in intensity caused by
the lesion or region of interest.
6.3
Object Contrast
x
The definition of contrast is not universal. Here are several reasonable interpretations of the concept of contrast in a particular situation.
0.1/cm
w = 2cm
Cobject =
0.13/cm
0.13 0.1
=
= 0.3,
background
0.1
Lineintegral Contrast
Clineintegral =
L(x)
0.66
0.6
l
0.66 0.6
0.06
=
=
= 0.1.
lbackground
0.6
0.6
Note that the line-integral contrast is lower than the original object
contrast, due to collapsing of a 2D object onto 1D function.
For intensity contrast:
Cintensity = CI =
I
Ibackground
I0 e0.6 I0 e0.66
= 1e0.06 .
I0 e0.6
I(x) / I_0
Z
R
C = 1 e .
I
x
R
The final approximation is reasonable when is small. This is
part of the small signal approximation, and is a reasonable simplifying assumption for system design, because big structures are easyit
is the small features that are challenging to image.
For structures that are small relative to the system spatial resolution,
contrast can be degraded, as illustrated.
6.4
SNR
The ability to detect an aberrant object (lesion) in a radiograph is related to the ratio of the differential intensity to the ambient
noise level. This ratio is called the absolute contrast to noise ratio, or the image signal to noise ratio:
SNR ,
I
I
=C ,
I
I
where I denotes the standard deviation of the background intensity due to noise processes.
To analyze I in the SNR expression above, we must consider noise sources and statistics. Conventional thermal noise processes
typical for communications systems apply only to radiography systems based on digital detectors with electronic readout noise.
Systems are designed to minimize electronic noise, so we focus on the quantum noise effects here.
Poisson statistics for quantum noise
During a specified exposure interval, suppose an X-ray source transmits N photons of a certain energy along a ray passing through
an object towards a specified pixel on the detector. We assume N is a Poisson random variable with mean N0 , where N0 thus
includes solid angle (pixel area /(4d)2 ), exposure time, and source strength at that energy. The PMF of a Poisson r.v. is:
P{N = n} =
1 N0 n
N0 ,
e
n!
n = 0, 1, . . . .
N
1
N0
c.o.v. ,
= ,
=
E[N ]
N0
N0
which decreases as N0 increases. So although the absolute variance of N increases with N0 , the relative variability decreases. So
more photons is better. Note that the SNR formula involves E[N ] /N , the reciprocal of the c.o.v.
Now consider the number of photons M that pass unaffected through the object along the given ray. Each of the N transmitted
photons may either pass unaffected (survive passage) or may interact with the object. These are Bernoulli trials because the
photons interact independently. From Beers law we know
n
P{M = m | N = n} =
pm (1 p)nm , m = 0, . . . , n
m
where p = e
(z) dz
1 N0 p
e
(N0 p)m ,
m!
P{M = m | N = n} P{N = n} =
X
n
nm 1 N0
pm (1 p)
N0n
e
m
n!
n=m
m = 0, 1, . . . .
Therefore the distribution of photons that survive passage is also Poisson! In particular
E[M ] = N0 p = N0 e
(z) dz
6.5
Example. Suppose that a structure might be present in the object. In terms of the signal M , the absolute contrast due to this
structure is:
R
R
M = E[M | no lesion] E[M | lesion] = N0 p N0 p e dz = 1 e dz N0 p = CM .
The SNR of M is therefore:
p
CM
M
= = C M.
M
M
The important thing here is not the final formula, but rather how we got from left equality to final equality.
SNR =
The following figure shows disks of different sizes, different contrasts, and hence different SNR values.
N = 100 photons/pixel
20
40
60
80
100
120
20
40 60 80 100 120
Range: [0, 147]
SNR
60
40
6.25
6.25
6.25
6.25
7.5
7.5
7.5
7.5
7.5
8.75
8.75
8.75
8.75
8.75
20 6.25
0
20
40
60
50
50
6.6
Picture:
Source
E[N ] = N0
Object
Bernoulli p = e
or e
Incident
Capture
E[M ] = N0 p
Detected
Bernoulli E[Y ] = N0 p
So the number of detected X-ray photons (where here we mean detected if captured) is
Y Poisson{N0 p} .
Absolute contrast for Y :
Y = E[Y | object only] E[Y | object+lesion] = N0 p N0 p e
R
= 1 e N0 p = CN0 p.
Thus
SNR =
where N1 , N0 p = N0 e
Y
Yobject
p
p
CN0 p
=
= C N0 p = C N1
N0 p
Clearly higher improves SNR, unless that higher comes at a price of reduced spatial resolution that decreases C.
General expression
SNR = C
where N is the (mean) number of photons at the stage of interest.
6.7
R
R
p
N0 e N0 e (+) p
N
R
p
=
=
= N 1 e = C N
R
N
N 0 e
R
C = 1 e
R
and where g = is the change in the line integral due to the structure of interest.
When the change in the line integral is small, called the small signal approximation, we have by the Taylor series (1 ex x)
that
Z
p
SNRN N |g| .
C = |g| ,
so
SNR of line integral (Treatment in [1] is not mathematically rigorous.)
We are often more interested in the line integral than the raw detected intensity (especially in tomography).
The method-of-moments estimate of g is
N0
g = log
= log N0 log N.
N
Note that if the data is noiseless, i.e., if N = N , then g = g, so the line integral is perfectly determined.
But in general N 6= N , so g is a random variable, and we would like to look at the SNR associated with g:
SNRg =
| E[
g ]|
.
g
To compute SNRg , we need the first two moments of g. Because g is a function of the random variable N , in principle one
could apply the standard method for transformations (functions) of a random variable to determine the distribution of g, and then
calculate the moments from that distribution. This is messy for the logarithm of a Poisson random variable. Thus, instead we settle
for approximations to the mean and variance of g. (These are in fact asymptotic approximations that improve as N0 increases.)
Note that g = f (N ) where
f (x) = log N0 log x.
By Taylor series expanded about x = N :
1
f (x) = f (N ) + f(N )(x N ) + f(x N )2 + . . . .
2
6.8
1
f(x) = ,
x
1
f(x) = 2 .
x
and
1
2
E[f (N )] f (N ) + f(N )N
2
f (N ) +
1
N
2
2
1 1
1
2
N
=g+
2 N
2N
2
N
=
1
.
N
p
| E[
g ]|
|g|
N |g| .
q
=
g
1/N
Thus the SNR of the estimated line integral g is approximately the same as the SNR of the number of X-ray photons N . So
improving the SNR by any of the methods we have discussed (e.g., improving contrast, detector efficiency, source strength, scan
time, etc.) will help both the raw image as well as the calculated line integrals.
1 However,
this approximation is not accurate in PET or SPECT transmission scans, where the photons are counted one by one, see [4].
6.9
SNR for recorders with scintillators or cascade processes (e.g., screen-film system)
So far we have considered only the SNR of captured X-ray photons, which is not the final signal that we observe.
When an X-ray photon is captured by a scintillator, it produces several light photons. The number of light photons is a random
variable, so we must account for this source of variability.
It is reasonable to assume that each captured X-ray photon produces a Poisson number of light photons. Suppose M X-ray photons
are captured, and let Xm denote the number of light photons produced by the interaction (that are emitted towards the film). We
assume that
the Xm values are independent
the Xm values are identically distributed Poisson random variables:
Xm Poisson{g1 }
where g1 500 is a gain factor. (Misleading term because although there is a gain in quanta, there is a loss in SNR.)
Now we look at the SNR of the total light signal:
Y ,
M
X
Xm .
m=1
This is a random sum because M is a random variable. We need the moments of Y to compute SNR.
How to evaluate the moments of Y ? By iterated expectation:
E[Y ] = E[E[Y |M ]] = E[M E[X]] = E[M ] E[X] = M g1
E Y2
X
= E E Y 2 |M =
E Y 2 |M = k P{M = k}
k=0
k=0
kE X
2
+k(k 1) E2 [X] P{M = k}
= E[M ] E X 2 +(E M 2 E[M ]) E2 [X]
2
= E[M ] X
+ E M 2 E2 [X] .
Var{Y }
=
=
=
=
E Y 2 E2 [Y ]
2
+ E M 2 E2 [X] (E[X] E[M ])2
E[M ] X
2
2
E[M ] X
+ E2 [X] M
M (g1 + g12 ) = M g12 (1 + 1/g1 ).
E[Y ]
g1 M
C M
SNR = C
= Cq
=p
.
Y
1 + 1/g1
M g12 (1 + 1/g1 )
As g1 , SNR C M , so for large gain g1 , the SNR is dominated by the X-ray photon statistics. But for small g1 , the SNR
of the light photon signal could be considerably worse than the SNR of the incident or captured X-ray photons.
For screen-film systems, g1 is upper bounded (by energy conservation) by the ratio of the X-ray photon energy to visible photon
energy, which is about 20,000. But this conversion is only about 5% efficient, so only about 1000 visible light photons produced
per captured X-ray photon. And about half of those photons do not reach the film, so g1 500.
Thus there is negligible loss in SNR at this stage, which is good for patient dose, because reduced SNR would mean that one must
increase scan time or source intensity to get adequate SNR.
6.10
Film grains
The scintillation photons are not observed directly. Instead, these photons cause silver grains in the film to be developed. Let Zi
be the number of developed film grains produced by the ith light photon. We assume the Zi values are iid random variables with
mean g2 .
Similar considerations arise in electronic detectors where the X-ray photons produce electron-hole pairs directly, or where the
scintillator light produces photoelectrons in a photodiode.
Because g2 1, the Poisson model is debatable. A more realistic model might be that Zi is a Bernoulli random variable with
p = g2 . The mean of Zi would be g2 and the variance would be 1g2 (1 g2 ) g2 , so the mean and variance are approximately the
same. Because the only Poisson property that we use is equivalence of the mean and variance, the Poisson model and the Bernoulli
model with small p lead to the same final formula.
The total number of film grains developed is:
W =
Y
X
Zi
i=1
E[W ]
C M
SNR = C
= q
W
1 + g11 +
1
g1 g2
E[W ]
C M
.
SNR = C
= q
W
1 + 1 + 1 + 1 +
g1
g1 g2
g1 g2 g3
So the cumulative gain products for each stage must be much greater than unity to have the SNR dominated by X-ray photon
statistics rather than variability in detector processes.
Unfortunately, for film G2 1/200, i.e., roughly only 1 in every 200 light photons develops a silver grain or on average 200 light
photons are needed to develop a silver grain.
Presumably different types of films (speeds) have different g2 values; what is the corresponding tradeoff (graininess)?
p
So g1 g2 500/200 = 2.5, so 1 + 1/2.5 1.2, so there is a 20% reduction in SNR relative to an ideal detector. So one must
effectively scan about 40% longer to get the same SNR as one would get with an ideal detector.
What is meant here by ideal detector? One that directly records every X-ray photon.
Fluoroscopy
X-ray film does not permit real-time X-ray imaging. For some types of studies, such as observing the flow of Iodine through
arteries or interventional procedures involving catheters etc., real-time imaging is essential.
In early systems, the visible light photons emitted from a fluorescent screen were observed directly by the unaided human eye.
p
Capture efficiency of the eye [1, p. 83] is only g2 105 , so g1 g2 103 105 = 102 (often even worse), so 1 + 1/(g1 g2 ) 10.
So SNR of image on retina is degraded by factor 10 relative to SNR of captured X-ray photons. Would require a factor of 100
increase in source strength to make up this loss.
Not great images!
6.11
Image intensifier
Patented in 1934 by GE. See SNR analysis on [1, p. 86].
Picture: fluorescent screen, photocathode, electrostatic lens, anode, output fluorescent screen, viewed by eye or TV camera.
g1 103 light photons / X-ray photon
g2 101 electrons / light photon
g3 103 light photons / accelerated electron
g4 105 human eye
p
g1 g2 g3 g4 1 (which dominates), so 1 + 1/g1 + . . . 2 so about 40% loss in SNR. (Better than unaided viewing of fluorescent screen though.)
With a TV camera:
g4 4 102 loss due to lens
g5 101 electrons per light photon in photoemitter
Now g1 g2 . . . g5 4 102 , so 1 + . . . 1!
Scatter
Loss of contrast: detected photons is transmitted plus scattered:
I d = I t + I s.
I t
I t /I t
C
I d
=
=
=
,
Id
It + Is
1 + I s /I t
1 + I s /I t
p
I t I t
C It
I d
It
=
= Cp
=q
SNR =
.
Id
I t Id
It + Is
1 + I s /I t
The SNR is reduced by the square root of the factor by which the contrast is reduced.
Scattered photons differ from unscattered photons in two ways. These differences offer possible scatter reduction methods.
Lower energy:
use an thin Pb sheet in front of detector
discriminate based on deposited energy in an energy-sensitive detector (e.g., in SPECT, PET)
Different direction:
Increase object-detector distance
Use a collimator
Bibliography
[1] A. Macovski. Medical imaging systems. Prentice-Hall, New Jersey, 1983.
[2] J. L. Prince and J. M. Links. Medical imaging signals and systems. Prentice-Hall, 2005.
[3] J. A. Fessler. Mean and variance of implicitly defined biased estimators (such as penalized maximum likelihood): Applications
to tomography. IEEE Trans. Im. Proc., 5(3):493506, March 1996.
[4] J. A. Fessler. Hybrid Poisson/polynomial objective functions for tomographic image reconstruction from transmission scans.
IEEE Trans. Im. Proc., 4(10):143950, October 1995.