image quality – quantifying quality? · 2015-10-06 · image quality –quantifyingquality?...
TRANSCRIPT
Image Quality – Quantifying Quality?
Stephan Scheidegger,2015
Image Quality – Quantifying Quality?
Contents
Motivation Mathematical representation
of image and imaging process From quality to quantity? Low contrast detectability:
Contrast, SNR, CNR, Noise characteristics: NPS High contrast detectability:
MTF CDC, DDC Model observers – observer
models Phantoms
RO
ENTG
ENTE
CH
NIK
STR
AHLE
NBI
OLO
GIE
GR
UN
DLA
GEN
RA
DIO
LOG
IE
STR
AH
LEN
PH
YSI
K
Why are we trying to measure Image Quality?
Different reasons – different tasks:
Performance: system suitable for clinical rasks?Clinical relevant quality (Model Observer)
Optimisation: system working at optimal point (ALARA)?Comparison of quality with applied dose (e.g. DDC with CTDI)
Quality control: Change in systems performance?Comparison of a measure representig systems performancewith the base line (e.g. DQE, NPS, MTF etc.)
Mathematical Image Representation
Different concepts (models):
2Dim.‐Functions (continuous models) Matrices (discrete representation – pixel‐model) Vector‐representation (Dim.=k x l x 1 resp. k x l x 3 for RGB)
( , ) ( , , )
...( , ) ( , )( , )
I x y t O x y z t
I x y F P x yS u v FFT F FFT P
A ( , , )
...
( , ) ( , )
H
kl
kl k lA
I O x y z
I P x y W x x y y dx dy
I S N
M
From Quality to Quantity?
technicalclinicalObserverimpression
numbers
MTF, NNPSDQE, SNRCNR
CDC / DDC
AlternativeForce Choice AFC
Visual GradingAnalysis VGA
Receiver OperatingCharacteristics ROC
Line pair TO
Image Quality – Image Characteristics
Qualities to quantities
Scharpness (spatial resolution, esp. High‐contrast resolution)
Gray level (distribution), dynamic range
Contrast Noise … others like uniformity, lag &
ghosting
Image Signal – Gray Levels, Blackening & Co.
Image signal representation
Blackening in film‐screen systems: optical density
Gray level displayed on a monitor (8 or 10 bit resolution, 256 – 1024 levels)
Gray level stored in a imagefile (12 bit – 4096 levels)
CT numbers, HU (12 bit, extended 12 bit)
Image Signal – Gray Levels, Blackening & Co.
Gray level – optical density
By definition adapted to visualprocess (and range!)
II
D refopt log
0.5
1.0
2.0
3.0
Image Signal – Visual Contrast Detection
Visual range & «contrastresolution»
Visual range approx. 10 bit(env. 900 gray scales)
Thresholds!
(1) (2)
max min
max min
(Michelson)
opt opt
nm kl
m
C D D
C I II ICI I
Image Signal – Visual Contrast Detection
LC detectabilityContrast C
Contrast & Dose: Gradation Curve
11000
optDDGyS
Contrast & Dose: Gradation Curve
usablerange
Gradation Curve
☼ flat☼ steep
Contrast & Dose: Gradation Curve
Image range=usablerange?
Gray Scale to Dose: CR System
0
0.0002
0.0004
0.0006
0.0008
0.001
0.0012
0.0014
0.0016
0.0018
0.002
2000 2500 3000 3500 4000
x (12bit)
c / m
Gy/
x(12
bit)
cf (meas)cf (calc)
Good contrast?
Image Quality and Optimisation
1
1
, ,...( ; ,...)( ; )
PMMA PVCHU HUSNR HU kVCNR HU kV
1 1( ), ( )c kV w kV
Parameter Set 11kV Parameter Set 2
2kV
2
2
, ,...( ; ,...)( ; )
PMMA PVCHU HUSNR HU kVCNR HU kV
1
1
( ; ,...)( ; )
SNR I kVCNR I kV
2
2
( ; ,...)( ; )
SNR I kVCNR I kV
2
2
( ),( )
c kVw kV
Low‐Contrast‐Detectability: Signal – to – Noise Ratio SNR
Sources of noise:
Quantum noise Detector noise More clinical: «decision
noise!» Often used: additive noise
model: IN(x,y) = I(x,y) + N(x,y)
1 1
1 ( , )
( ) ( )
N M
n mn m
nm nm
HU x y HUN MSNRs HU s HU
Noise level 20%, r = 2 pixel
Noise level 5%, r = 2 pixel
Noise level 10%, r = 2 pixel
Dose-dependence of noise(CR-system)
1 1
1 ( , )N M
N M n mn m
R I x ys N M
0
20
40
60
80
100
120
0 0.5 1 1.5 2 2.5
Dose / mGy
R(10
1x10
1)
73 kV90 / 125 kV
(125 kV)
CT: Tube current and noise
Low‐Contrast‐Detectability: Contrast – to – Noise Ratio CNR
Simple approach: Difference of SNR in two compared ROI’s
Usefull for relative signaldetection with threshold?
2 112
, 2 , 1( ) ( )Pos Pos
nm Pos nm Pos
HU HUCNR SNR
s HU s HU
Low‐Contrast‐Detectability: Contrast – to – Noise Ratio CNR
CNR = ‐0.007
CNR = 0.03
Is CNR a usefull quantity?
upper threshold
lower threshold
signal
upper threshold
lower threshold
signal
Increasing noise
+ Noise Stochastik Resonance!
Nosie Characteristics?
Noise Characteristics CT
Miéville et al (2012): Effects of computing parameters and measurement locations…Phys Med
Marshall N: The diagnostic Xray perspective, UZ Leuven
Noise Characteristics: NPS
High‐Contrast‐Resolution
Observer‐based vs. calculated:
Line pairs (lp / mm) Modulation transfer
function MTF
HC resolution(lp / mm)
a) 73kV 32mAs b) 90kV 8mAs
FT
Modulation Transfer Function MTF
Modulation Transfer Function MTF
Results without Noise• ROI with high contrast edge:
no filter (1); r = 0.5 p (2); r = 1 p (3); r = 2 p (4); r = 4 p (6); r = 12 p (8)
MTF high C
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
lp / pixel
MTF
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
Gaussian filter r = 12 pixel
initial
Gaussian filter r = 4 pixels
-50
0
50
100
150
200
250
300
-8 2 12 22 32
Position / pixel
Gre
y Va
lue
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
MTF high C
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
lp / pixel
MTF
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
Gaussian filter r = 12 pixel
initial
Gaussian filter r = 4 pixels
-50
0
50
100
150
200
250
300
-8 2 12 22 32
Position / pixel
Gre
y Va
lue
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
MTF high C
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
lp / pixel
MTF
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
MTF high C
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
lp / pixel
MTF
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
MTF(50)
MTF high C
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
lp / pixel
MTF
Reihe1Reihe2Reihe3Reihe4Reihe5Reihe6Reihe7Reihe8
MTF(50)
Modulation Transfer Function MTF
Results with Noise• Gaussian filter with r = 0.5 pixel, no
noise (blue), 10% Gaussian noise (yellow), 20% Gaussian noise (pink)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
Results with Noise• ImageJ: MTF with no noise vs.
MTF with noise, at high contrast level: (a) Gaussian filter with r = 4 pixels, noise level 5%; (b) Gaussian filter with r = 0.5 pixel, noise level 20%; (c) Gaussian filter with r = 0.5 pixel, noise level 10%
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1
00.10.20.30.40.50.60.70.80.9
1
0 0.5 10
0.10.20.30.40.50.60.70.80.9
1
0 0.5 1
Optimisation of CR-Systems
b) 73kV 2mAsa) 73kV 32mAs c) 90kV 8mAs d) 125kV 4mAs
Optimisation of CR Systems
Modulations-Transfer-Funktion
0
0.5
1
1.5
2
2.5
3
3.5
4
0 0.5 1 1.5 2 2.5
Dose / mGy
MTF
/ lp
/mm
MTF50MTF80
1.0
0.5
0.3
0.8
5 10 15 2000.0
Bildfrequenz / (lp/mm)
MTF
2( ) ( ) ixMTF L x e dx
Modulation Transfer Function MTF: Iterative Reconstruction total kk
MTF MTF
MTF 50Abdomen (Siemens Force Abdomen 2.0 Br36 3) 0.227Abdomen (Siemens Force Abdomen 2.0 Br36 4) 0.239Abdomen (Siemens Force Abdomen 2.0 Br36 5) 0.2435
Richard et al. (2012): Towards task‐based assessment of CT performance. Med Phys
Alternative Approach
Images:[1]
Contrast‐detail curve [1, 2, 3]
Alternative Approach
Images:[1]
Alternative Approach
Thorax Protocol @ 120 kV B70s
92.5mA/7.62CTDI
69mA/5.83CTDI
58.75mA/4.93CTDI
No Window Window C40 W300
Window C300 W1500
Images using different windows
Pelvis Protocol@
80 kV
100 kV
No Window Window C40 W300 Window C300 W1500
Aproach for obtaining CDC and DDCRegistration
CT ImageTemplate Image
Approach for CDC and DDCPaired point matching using ICP
Affine TransformationT
Model Observers – Observer Models
Aims
Standardized observer To mimic psychophysiological aspects of recognition To cover image quality close to the clinical need
1 0
0 1
0
1
2 2
: H: H( )
; 1 12 2
b
b s
H Htt
H H
HH
t tt SNR
I I NI I I N
T I
Choosing the «Right Phantom» … ?