+ All Categories
Home > Documents > Statistics and Image Quality Evaluation III

Statistics and Image Quality Evaluation III

Date post: 30-Dec-2015
Category:
Upload: daquan-eaton
View: 29 times
Download: 2 times
Share this document with a friend
Description:
Statistics and Image Quality Evaluation III. Oleh Tretiak Medical Imaging Systems Fall, 2002. This Lecture. Mean square error as quality measure Shortcomings How to do ROC by hand Visual ROC experiment This and other files for todays lecture at - PowerPoint PPT Presentation
Popular Tags:
26
1 Statistics and Image Quality Evaluation III Oleh Tretiak Medical Imaging Systems Fall, 2002
Transcript
Page 1: Statistics and Image Quality Evaluation III

1

Statistics and Image Quality Evaluation III

Oleh Tretiak

Medical Imaging Systems

Fall, 2002

Page 2: Statistics and Image Quality Evaluation III

2

This Lecture

• Mean square error as quality measure– Shortcomings

• How to do ROC by hand

• Visual ROC experiment

• This and other files for todays lecture at– http://www.ece.drexel.edu/courses/ECE-S684/Notes/IQ3/

Page 3: Statistics and Image Quality Evaluation III

3

Analytic Image Fidelity

• Mean square difference between imagesu(m, n) - perfect image, u’(m, n) - defective

image

σ a2 =

1MN

u(m,n) − ′ u (m,n)( )2

n=1

N

∑m=1

M

SNR =10 log10σ 2

σ a2

PSNR =10 log10(peak − to − peak value in reference image)2

σ a2

Page 4: Statistics and Image Quality Evaluation III

4

‘White Noise’ Pattern

Page 5: Statistics and Image Quality Evaluation III

5

Noise Patterns

• White, low frequency, and high frequency noises. All have same standard deviation

Page 6: Statistics and Image Quality Evaluation III

6

Effect of noise on image quaity: UL ~ original 8-bit image; UR ~ white noise; LL ~ low pass noise; LR ~ high pass noise. Noise standard deviation is equal to 8, so that PTP signal to rms noise ratio is 32.

Page 7: Statistics and Image Quality Evaluation III

7

Effect of noise on image quaity: UL ~ original 8-bit image; UR ~ white noise; LL ~ low pass noise; LR ~ high pass noise. Noise standard deviation is equal to 32, so that PTP signal to rms noise ratio is 8.

Page 8: Statistics and Image Quality Evaluation III

8

Conclusions

• Same SNR does not produce the same image degradation when noises are different

• ‘Low frequency’ noise is most visible

• ‘High frequency’ noise is most damaging

Page 9: Statistics and Image Quality Evaluation III

9

• “The mean square error criterion is not without limitations, especially when used as a global measure of image fidelity. The prime justification for its common use is the relative ease with which it can be handled mathematically…” Anil Jain, in Fundamentals of Digital Image Processing.

Page 10: Statistics and Image Quality Evaluation III

10

Method

• Rating on a 1 to 5 scale (5 is best)

• Rating performed by 21 subject

• Statistics:– Average, maximum, minimum, standard

deviation, standard error for Case A, Case B– Difference per viewer between Case A and

Case B, and above statistics on difference

Page 11: Statistics and Image Quality Evaluation III

11

Two Kinds of Errors

• In a decision task with two alternatives, there are two kinds of errors

• Suppose the alternatives are ‘healthy’ and ‘sick’– Type I error: say healthy if sick– Type II error: say sick if healthy

Page 12: Statistics and Image Quality Evaluation III

12

• X - observation, t - threshold

= Pr[X > t | H0] (Type I error)

= Pr[X < t | H1] (Type II error)

Choosing t, we can trade off between the two types of errors

H0

H1

a

b

Decision threshold

Page 13: Statistics and Image Quality Evaluation III

13

ROC Terminology

• ROC — receiver operating characteristic

• H0 — friend, negative; H1 — enemy, positive

• Pr[X > t | H0] = probability of false alarm = probability of false positive = PFP =

• Pr[X > t | H1] = probability of detection = probability of true positive = PTP =

Page 14: Statistics and Image Quality Evaluation III

14

The ROC

• The ROC shows the tradeoff between PFP

and PTP as the threshold is varied

H0

H1

Decision threshold

PFP

PTP

PTP

PFP

0 10

1

AZ

Page 15: Statistics and Image Quality Evaluation III

15

Binormal Model

• Negative: Normal, mean = 0, st. dev. = 1

• Negative: Normal, mean = a, st. dev. = b

fN(x)=12π

exp(−x2 / 2)

fP(x) =12π

exp[−(x−a)2 /2b2]€

PFP (t) =1−Φ(t)

PTP (t) =1−Φt − a

b

⎛ ⎝ ⎜

⎞ ⎠ ⎟

Page 16: Statistics and Image Quality Evaluation III

16

Page 17: Statistics and Image Quality Evaluation III

17

ROC from Experimental Data

-0.96 N-1.04 N-1.75 N1.85 N1.64 N0.70 N0.02 N

-0.49 N0.45 N

-1.75 N 0.888-1.04 N 0.777-0.96 N 0.666-0.49 N 0.5550.02 N 0.4440.45 N 0.3330.70 N 0.2221.64 N 0.1111.85 N 0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-2.00 -1.50 -1.00 -0.50 0.00 0.50 1.00 1.50 2.00 2.50

Strarting data: 9 samples of N(0, 1) pseudorandom variable, simulating negative population. Middle: data sorted, with cumulative distribution. Right: plot of middle.

Page 18: Statistics and Image Quality Evaluation III

18

ROC from Experimental Data 2

3.77 P1.18 P

-0.82 P3.05 P4.06 P

-1.01 P

-1.01 P 0.833-0.82 P 0.6661.18 P 0.4993.05 P 0.3323.77 P 0.1654.06 P 0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

-2.00 -1.00 0.00 1.00 2.00 3.00 4.00 5.00

Strarting data: 7 samples of N(2, 1.5) pseudorandom variable, simulating positive population. Middle: data sorted, with cumulative distribution. Right: plot of middle.

Page 19: Statistics and Image Quality Evaluation III

19

ROC from Experimental Data 3

-1.746 N 0.888-1.035 N 0.777-1.011 P 0.833-0.965 N 0.666-0.817 P 0.666-0.494 N 0.5550.025 N 0.4440.448 N 0.3330.701 N 0.2221.183 P 0.4991.642 N 0.1111.848 N 03.049 P 0.3323.767 P 0.1654.059 P 0

FPF TPF1 1

0.888 10.777 10.777 0.8330.666 0.8330.666 0.6660.555 0.6660.444 0.6660.333 0.6660.222 0.6660.222 0.4990.111 0.499

0 0.4990 0.3320 0.1650 0

Formula in FPF column=IF(N4="N",O4,P3)

Formula in TPF column=IF(N4="P",O4,Q3)

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1 1.2

Merged sorted data and ‘Empirical ROC’ plot.

Page 20: Statistics and Image Quality Evaluation III

20

Comparison of Finite and Infinite Sample

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1 1.2

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1 1.2

Empirial ROC9 negative and 7 positive cases

Theoretical ROCWhat would have been obtained from a

very large sample

Page 21: Statistics and Image Quality Evaluation III

21

Metz

• University of Chicago ROC project:

• http://www-radiology.uchicago.edu/krl/toppage11.htm

• Software for estimating Az, also sample st. dev. And confidence intervals.

• Versatile

Page 22: Statistics and Image Quality Evaluation III

22

Ordinal Dominance Theory

• Donald Bamber, Area above the Ordinal Dominance Graph and the Area below the Receiver Operating Characteristic Graph, J. of Math. Psych. 12: 387-415 (1975).

• Sample Az is an unbiased, efficient, asymptotically normal estimate of population Az

• Bamber gives a formula for variance of Az

• Worst-case estimate (overbound) of variance

σmax

2

=

A ( 1 − A )

min( NX

, NY

)

Page 23: Statistics and Image Quality Evaluation III

23

Visual ROC Target

• Target is 9x16 pixels, Gaussian noise plus 4x4 dark rectangle on one side

• Protocol: Answer L, R if you think target is on left or right.

• In this example, target is on left

Page 24: Statistics and Image Quality Evaluation III

24

Experiment Goal

• Test human (your) visual system processing power

• Target is easy to see in absence of noise• This sort of stimulus can be analyzed, and

theoretical best ROC found• Use paper form for your answers, enter your

results on computer & mail to me.• Alternate method: Download file ROC-test-

form.xls, enter data on form, and send to me by e-mail.

Page 25: Statistics and Image Quality Evaluation III

25

Group 1

Page 26: Statistics and Image Quality Evaluation III

26

Group 2


Recommended