8/2/2016
Ke Li
Assistant Professor
Department of Medical Physics and Department of Radiology
School of Medicine and Public Health, University of Wisconsin-Madison


This work is partially supported by an NIH Grant No.
R01CA169331 and a GE Research Agreement
Special thanks to the following colleagues and collaborators for
their contributions to this presentation









Dr. Guang-Hong Chen
Dr. Jie Tang
Dr. Jiang Hsieh
John Garrett
Daniel Gomez-Cardona
Juan Pablo Cruz-Bastida
Yongshuai Ge
Adam Budde
Dr. Timothy Szczykutowicz

Quantitative image quality assessment of conventional
linear CT systems

Challenges in the quantitative image quality assessment
of MBIR-based nonlinear CT systems

Potential solutions
1
8/2/2016

Measures the degradation/distortion of the acquired
image (relative to an ideal image) using a quantitative
figure-of-merit

Advantages over qualitative image quality assessment:
 Reduces subjectivity associated with qualitative assessments
 Enables more meaningful comparison of CT systems across
institutions and vendors
 Usually has higher efficiency and lower cost

Frequency-independent metrics






Noise magnitude
CT number uniformity
Contrast
Contrast-to-noise ratio (CNR)
CNR normalized by dose (CNRD)
Frequency-dependent metrics
 Noise power spectrum (NPS)
 Modulation transfer function (MTF)
 Task-based detectability index

Quantitative CT image quality assessment does not
include:
 Visual determination of linear pair resolution (not quantitative)
 X-ray beam collimation accuracy (not an assessment of image
quality)
 Measurement of dose-length product (assessment of CT
dosimetry instead of image quality)
2
8/2/2016
Standard kernel
1.6
Std.
Bone plus
HD Ultra
1.2
MTF
Bone Plus kernel
0.8
0.4
0.0
0.0
0.5
1.0
1.5
2.0
Frequency (mm-1)
Hi-Res mode
with HD Ultra kernel
(in vivo results)
2 
Dose 
exp(uL)
Dose

exp(uL)
2
Dose (mGy)
J. P. Cruz-Bastida et al. Med. Phys. 43, 2399 (2016)
Patient Size L (mm)
Courtesy of Dr. Szczykutowicz
Szczykutowicz, T. P., Bour, R. K., Pozniak, M., & Ranallo, F. N. Journal of Applied Clinical Medical Physics, 16, 2 (2015)
1 mGy
10 mGy
60 mGy

Noise variance (σ2) or noise
standard deviation (σ)
10
10 mGy
1 N
 ( xi  x )
N  1 i 1
2 
1
Dose
4
60 mGy
2 (HU2)
1 mGy
2 
10
10
K. Li, J. Tang, G.-H. Chen, Med. Phys. 41, 041906 (2014)
3
2
1
2.5
5
10
20
40 60
Dose (mGy)
3
8/2/2016
0.625 mm
1.25 mm
1
 
z
2
2.5 mm
5.0 mm
2 (HU2)
6400
1600
1 mGy
5 mGy
20 mGy
60 mGy
400
100
25
0.625
1.25
2.5
5
Slice thickness (mm)

Contrast
C  xfeature  xbackground
W/L: 150/300
Contrast = 100 HU

Contrast = 225 HU
Contrast-to-noise ratio (CNR)
Contrast = 350 HU
CNR 
Contrast = 400 HU
C

W/L: 150/300
4
8/2/2016
4 mGy
8 mGy
CNR  Dose
25
16 mGy
CNR
CNR
20
12 mGy
15
10
R2 = 0.994
5
0
0
1
Standard kernel
Lung kernel
2
3
4
Dose
W/L: 300/200
Bone+ kernel
Highest CNR
Fracture least
visible
Edge kernel
Lowest CNR
Fracture most
visible
Signal present
Conventional CT
Signal absent
Differential phase contrast CT
absorption CT noise
DPC-CT noise
Same CNR
Different detection performance
AUC  1
AUC  0.88
K. Li et al., Proc. SPIE, 8313, 83131L(2012)
5
8/2/2016

Same CNR, different object detectability
Contrast = 30 HU
W/L: 50/16
CTDIvol = 16 mGy

NPS 
Noise power spectrum (NPS)
2
x y
DFT(ROI-ROI)
Nx N y
D. Gomez-Cardona et al., Med. Phys. 43, 4495 (2016)
The shape of the NPS is independent on radiation dose
NPS (FBP)
Normalized NPS (FBP)
12000
2.5
5 mAs
10000
25 mAs
8000
50 mAs
100 mAs
200 mAs
300 mAs
6000
4000
2000
0
0
0.2
0.4
0.6
0.8
Normalized NPS (A.U.)
NPS (HU2mm2)
12.5 mAs
2
1.5
1
0.5
0
0
0.2
fxy (mm-1)
0.4
0.6
0.8
fxy (mm-1)
K. Li, J. Tang, G.-H. Chen, Med. Phys. 41, 041906 (2014)
6
8/2/2016
NPS magnitude is independent on contrast level
Contrast = 100 HU
Contrast = 225 HU
Contrast = 350 HU
Contrast = 400 HU
W/L: 150/300

Point spread function (PSF) or modulation transfer function (MTF)
HD Std.
HD Bone plus
HD Edge
HD Lung
MTF
1.2
0.8
Standard
Lung
Bone plus
Edge
0.4
0.0
0.0
0.5
1.0
1.5
2.0
Frequency (mm-1)
J. P. Cruz-Bastida et al. Med. Phys. 43, 2399 (2016)
Spatial resolution is independent on dose and contrast
Dose independence
Contrast independence
0.25
16 mGy
12 mGy
8 mGy
4 mGy
0.2
Point Spread Function
Point Spread Function
0.25
30 HU
60 HU
100 HU
225 HU
0.2
Question: Are these assumptions still valid in MBIR?
0.15
0.1
0.05
0
-1.5
-1
-0.5
0
0.5
1
1.5
0.15
0.1
0.05
0
-1.5
-1
Position (mm)
-0.5
0
0.5
1
1.5
Position (mm)
Li, Garrett, Ge, Chen, Med. Phys. 41, 071911 (2014)
7
8/2/2016
final image
projection data
update the estimated image
forward
projection
estimated image
Noise magnitude is independent on contrast level
Li et al., RSNA 2014 Annual Meeting
Solomon and Samei, Proc SPIE 8668, 86684M (2013)
Solomon and Samei, Med. Phys. 41, 091908 (2014)
FBP image
FBP
MBIRnoise-only
reconstruction
MBIR noise-only image
Noise variance is inversely proportional to radiation dose
10
4
2 (HU2)
FBP
FBP (fitting)
MBIR
10
10
FBP:
2 
MBIR:
?
3
2
2
10
1
Dose
Dose
1
1
2.5
5
10
20
40
60
Dose (mGy)
8
8/2/2016
The shape of the NPS independent on radiation dose
Normalized
NPS
of(Veo)
MBIR
Normalized
NPS
6
700
512.5 mAs
Normalized NPS (A.U.)
5 mAs
600
500
2
2
NPS (HU mm )
NPS
of MBIR
NPS
(Veo)
800
400
300
200
100
0
0
0.2
100 mAs
200 mAs
300 mAs
3
2
1
0
0.6 0
0.4
5 mAs
12.5 mAs
25 mAs
50 mAs
100 mAs
200 mAs
300 mAs
25 mAs
450 mAs
0.80.2
fxy (mm-1)
0.4
0.6
0.8
fxy (mm-1)
K. Li, J. Tang, G.-H. Chen, Med. Phys. 41, 041906 (2014)
Noise variance is inversely proportional to slice thickness
0.625 mm
1.25 mm
2.5 mm
5.0 mm
10
2 
3
1
z
FBP
Veo
2 (HU2)
FBP
10
2
Veo
1
10
0.625
1.25
2.5
Slice thickness z (mm)
Matched W/L: 50/0; CTDIvol = 4 mGy
5
K. Li, J. Tang, G.-H. Chen, Med. Phys. 41, 041906 (2014)
FBP
0.625 mm
Veo
0.625 mm
FBP thin
FBP thick
Veo thin
Veo thick
FBP
5.0 mm
Veo
5.0 mm
HIPAA-compliant prospective human subject study, IRB-approved
9
8/2/2016
K. Li, J. Tang, G.-H. Chen, Med. Phys. 41, 041906 (2014)
Spatial resolution is independent on dose and contrast
contrast = 350 HU
contrast = 100 HU
0.25
16 mGy
12 mGy
8 mGy
4 mGy
0.2
0.15
0.1
0.05
Point Spread Function
Point Spread Function
0.25
0.2
0.15
0.1
0.05
0
0
-1.5
-1
-0.5
0
0.5
1
1.5
-1.5
-1
-0.5
0
0.5
1
1.5
Position (mm)
Position (mm)
Li, Garrett, Ge, Chen, Med. Phys. 41, 071911 (2014)
The Z resolution is independent on dose and contrast
Contrast = 270 HU
Contrast = 120 HU
0.15
25% dose (Veo)
50% dose (Veo)
75% dose (Veo)
100% dose (Veo)
100% dose (FBP)
0.1
0.05
0
-0.05
-2
-1
0
1
2
Slice sensitivity profile
Slice sensitivity profile
0.15
25% dose (Veo)
50% dose (Veo)
75% dose (Veo)
100% dose (Veo)
100% dose (FBP)
0.1
0.05
0
-0.05
-2
-1
z position (mm)
0
1
2
z position (mm)
Li, Garrett, Ge, Chen, Med. Phys. 41, 071911 (2014)
10
8/2/2016

Noise assessments
 Contrast dependence
 Modified dose dependence
▪ Magnitude
▪ NPS shape
 Modified slice thickness dependence
High contrast, high dose
▪ Magnitude
▪ NPS shape

Spatial resolution assessments
 Contrast dependence
 Dose dependence
 Spatial resolution assessments at low-contrast and low
dose conditions are extremely challenging
Low contrast, low dose
Task-based
spatial resolution
Repeated
scans
Task-based
detectability
Noise
ensemble
Task-based NPS
Task-based
“edge” noise
Object edge
Gradient image
Edge central line
Normal cuts
Edge profile at
arbitrary location
High
contrast
Low
contrast
•
Richard et al., “Towards task-based assessment of CT performance: System and objective MTF across different
reconstruction algorithm.” Med. Phys. 39, 4115 (2012)
•
Li et al., “Statistical Model Based Iterative Reconstruction (MBIR) in clinical CT systems: Part II. Experimental
assessment of spatial resolution performance.” Med. Phys. 41, 071911(2014)
•
Yu et al., “Technical Note: Measuring contrast-and noise-dependent spatial resolution of an iterative
reconstruction method in CT using ensemble averaging.” Med. Phys. 42, 2261 (2015)
11
8/2/2016
1
(mm)
(mm)
PSF
Width
(mm)
width
PSF of
Veo
0.8
0.8
0.7
FBP
0.6
0.6
0.5
0.4
0.4
100%
0.2
16
33
75%
62 99
224 346
50%
814 1710
25%
Dose
Contrast (HU)
Li, Garrett, Ge, Chen, Med. Phys. 41, 071911 (2014)
(mm)
PSF
Width
(mm)
width
PSF of
1

At each given exposure level,
(mm) the
Veo
spatial
resolution of Veo improves with
0.8
higher contrast level

At each givenFBP
contrast level, the spatial
0.6
resolution of Veo improves
with higher
0.5
dose level

0.4
At some intermediate contrast/dose
level, the spatial resolutions of Veo and
100%
FBP becomes equivalent
0.8
0.7
0.6
0.4
0.2
16
33
75%
62 99
224 346
50%
814 1710
Contrast (HU)
25%
Dose
NPS peak position f
peak
(mm-1)
Li, Garrett, Ge, Chen, Med. Phys. 41, 071911 (2014)
0.8
0.4
FBP
FBP (fitting)
Veo
Veo (fitting)
FBP: f peak  const.
0.2
1
MBIR: f peak   dose  5
0.1
0.05
1
2.5
5
10
20
40
60
Dose (mGy)
12
8/2/2016
10
3
2 (HU2)
FBP
Veo
10
FBP:  2 
1
z
2
2
MBIR:  
1
z
1
10
0.625
1.25
2.5
Slice thickness z (mm)
2 (HU2)
10
10
10
10
5
4
FBP
FBP (fitting)
Veo
Veo (fitting)
3
2
FBP:  
2
MBIR:  
2
1
Dose
1
Dose0.4
1
1
2.5
5
10
20
40
60
Dose (mGy)
0.4±0.1
1.0±0.1
Reduced
dose
Reference
dose
HIPAA-compliant and IRB-approved
D. Gomez-Cardona et al., 2015 AAPM Annual Meeting
13
8/2/2016
Noise = 15 HU
(FBP, 65 mGy)

Noise = 15 HU
(MBIR, 0.6 mGy)
Problem formulation
arg min Dose(kV,mA) s.t. d '(kV,mA)  d ℞'
kV,mA

Practical solution
Black lines: Iso-dose lines;
700
Iso-detectability
lines (FBP)
Iso-dose White
map lines: Iso-detectability
Iso-d’
linesmap
700
60
50
500
15
40
400
30
300
20
200
10
100
100
120
mA
mA
500
25
80
20
600
600
400
300
d ℞'
200
10
100
25
80
140
5
100
120
140
kV
kV
Li et al., Med. Phys. 42, 5209 (2015)
Routine dose scan
Reduced dose scan
120kV 650 mA
19.4 mGy
80 kV 275 mA
2.4 mGy
FBP
HIPAA-compliant and IRB-approved
FBP
MBIR
Li et al., Med. Phys. 42, 5209 (2015)
14
8/2/2016

Major challenges in quantitative image quality
assessment of MBIR
 Violation of several basic assumptions and relationships in CT
image quality assessment
 Nonlinearity of the reconstruction method determines that there
is no theoretical basis in deriving the modified relationships
▪ For example, how does the shape of the NPS quantitatively vary with
radiation dose

Potential solutions in address these challenges
 Empirical relationships between image quality and CT system
parameters have been found
 Try to use frequency-dependent metrics (e.g NPS) rather than
zero-frequency metrics (e.g. CNR)
 Be task-specific rather than task-generic
 Perform repeated acquisitions rather than shifting an ROI
 Use model observer detectability to combine noise and spatial
resolution metrics with task function
Thank You
45
15