8/2/2012
Assessment
of Image Quality
in The New CT
Jeff Siewerdsen, PhD
Biomedical Engineering
Johns Hopkins University
Jeff Fessler, PhD
Electrical Engineering and Computer Science
University of Michigan
Kyle Myers, PhD
Center for Devices and Radiological Health
Food and Drug Administration
Acknowledgments
I-STAR Laboratory
Imaging for Surgery, Therapy, and Radiology
www.jhu.edu/istar
JW Stayman, A Muhit, Y Otake, S Schafer, W Zbijewski
G Gang, H Dang, Y Ding, P Prakash, J Xu, S Arora
S Nithiananthan, D Mirota, A Uneri, S Reaungamornrat
Hopkins Collaborators
J Carrino, K Taguchi, M Mahesh (Radiology)
J Prince, J Lee (RadOnc)
Academic-Industry Partnership
J Yorkston (Carestream Health)
R Graumann, G Kleinszig, T Mertelmeier (Siemens XP)
Disclosures and Support
Advisory Board, Carestream and Siemens
Elekta Oncology Systems
National Institutes of Health 2R01-CA-112163
Part 1: Overview
The “New CT”
New scanner configurations (including CBCT)
New reconstruction methods (including statistical / iterative)
Basic Technical Assessment
Radiation dose and imaging performance
Phantoms and standardization
Measurement and Modeling of Performance
Noise, spatial resolution, and detectability
Application to new technology development
Extensions and Challenges in “The New CT”
Assumptions and limitations
Dual-energy CT, Phase contrast CT, etc.
Iterative / statistical reconstruction
1
8/2/2012
Basic Technical Assessment
not-so- Basic
Technical
Assessment
Trade names removed.
Basic Technical Assessment
Radiation Dose
Radiation Dose
Farmer chamber + 16 cm cylinder
Short-scan protocols
Farmer chamber + 16 cm cylinder
Short-scan protocols
Quantitative Accuracy
Quantitative Accuracy
Electron density inserts
Comparison to MDCT
Electron density inserts
Comparison to MDCT
Contrast Resolution
Contrast Resolution
Low-contrast tissue inserts
SDNR versus kVp, mAs
Low-contrast tissue inserts
SDNR versus kVp, mAs
Spatial Resolution
Spatial Resolution
Line-pair pattern (subjective)
Modulation transfer function (MTF)
Line-pair pattern (subjective)
Modulation transfer function (MTF)
“Clinical” Image Quality
“Clinical” Image Quality
Anthropomorphic phantoms
Expert readers
Anthropomorphic phantoms
Expert readers
J Xu et al. Med. Phys. (in press - August)
J Xu et al. Med. Phys. (in press - August)
2
8/2/2012
Basic Technical Assessment
Basic Technical Assessment
Basic Technical Assessment
Radiation Dose
Radiation Dose
Radiation Dose
Farmer chamber + 16 cm cylinder
Short-scan protocols
Farmer chamber + 16 cm cylinder
Short-scan protocols
Farmer chamber + 16 cm cylinder
Short-scan protocols
Quantitative Accuracy
Quantitative Accuracy
Quantitative Accuracy
Electron density inserts
Comparison to MDCT
Electron density inserts
Comparison to MDCT
Electron density inserts
Comparison to MDCT
Contrast Resolution
Contrast Resolution
Contrast Resolution
Low-contrast tissue inserts
SDNR versus kVp, mAs
Low-contrast tissue inserts
SDNR versus kVp, mAs
Low-contrast tissue inserts
SDNR versus kVp, mAs
Spatial Resolution
Spatial Resolution
Spatial Resolution
Line-pair pattern (subjective)
Modulation transfer function (MTF)
Line-pair pattern (subjective)
Modulation transfer function (MTF)
Line-pair pattern (subjective)
Modulation transfer function (MTF)
“Clinical” Image Quality
“Clinical” Image Quality
“Clinical” Image Quality
Anthropomorphic phantoms
Expert readers
Anthropomorphic phantoms
Expert readers
Anthropomorphic phantoms
Expert readers (sanity check)
J Xu et al. Med. Phys. (in press - August)
J Xu et al. Med. Phys. (in press - August)
J Xu et al. Med. Phys. (in press - August)
3
8/2/2012
Checking for Pulse…
Why don’t we use a 10 cm pencil ionization
chamber to measure dose in cone-beam CT?
30%
13%
7%
20%
30%
1.
2.
3.
4.
5.
The dose is too high.
The dose is too low.
The field is longer than the chamber.
CBCTDI is a clumsy acronym.
We do.
Checking for Pulse…
Why don’t we use a 10 cm pencil ionization
chamber to measure dose in cone-beam CT?
1. The dose is too high.
2. The dose is too low.
3. The field is longer than the chamber.
Measurement & Modeling
Noise
Spatial Resolution
Detectability
4. CBCTDI is a clumsy acronym.
5. We do.
AAPM Task Group 111
www.aapm.org/pubs/reports/ (Feb 2010)
4
8/2/2012
Measuring the Noise
Measuring the Noise
Noise-Power Spectrum
Noise-Power Spectrum
Measuring the Noise
Noise-Power Spectrum
Axial Plane (x,y)
Axial NPS
S(fx, fy)
30
Noise (CT#)
25
20
15
10
5
0
0
0.5 1.0 1.5 2.0 2.5 3.0
Dose (mGy)
Barrett, Gordon, and Hershel (1976)
5
8/2/2012
Measuring the Noise
Measuring the Noise
Noise-Power Spectrum
Noise-Power Spectrum
Sagittal Plane (x,z)
Sagittal NPS
S(fx, fz)
Axial domain (fx,fy)
“Filtered-ramp”
Mid-Pass
Sanity Check
What is wrong with analyzing the local NPS
from a single axial slice in cone-beam CT?
27%
Longitudinal domain (fz)
“Band-limited”
Low-Pass
13%
Low-frequency NPS
NPS(fx,fy) α f
(dNPS / df ) α NEQ(0)
NPS(0,0,0) ≠ 0 (aliasing)
20%
13%
27%
1.
2.
3.
4.
5.
The magnitude is wrong.
The units are wrong.
Ignores correlation in the z direction.
Would overestimate the NEQ.
All of the above.
Units
[signal2] [Π(domain)]
[µ]2 [x] [y] [z]
→ (HU2)(mm3)
→ (/mm2)(mm3)
1. Hanson. Med Phys 1979
2. Kijewski and Judy. Phys Med Biol 1987
3. Siewerdsen, Cunningham, and Jaffray Med Phys 2001
6
8/2/2012
Sanity Check
2D PROJECTION DATA
What is wrong with analyzing the local NPS
from a single axial slice in cone-beam CT?
Projection
4. Would overestimate the NEQ.
5. All of the above.
Siewerdsen, Jaffray, and Cunningham
Med Phys 29(11) (2002)
CASCADED SYSTEMS ANALYSIS
Incident Quanta q0(E)Logarithm
8 Interpolation
Scintillator
Detection
Gain, Blur
9 Ramp Filter
Flat-Panel Detector
Conversion
Aperture
10 Apodization
Electronics
Sampling
Backprojection
Noise
3D Filtered Backprojection
3. Ignores correlation in the z direction.
Modeling the Noise
Task Function
1. The magnitude is wrong.
2. The units are wrong.
Modeling the Noise
T
T8
T9
T
Siewerdsen et al. (1997)
Zhao et al. (1997)
Sampling
III12 3D… and
others
kVp, Dose
Scintillator, Detector
Pixel Aperture, 2D Sampling
Reconstruction Filter
Backprojection
Voxel Size
3D Sampling
T10NEQ
Generalized
fx
fz
Quantum Noise
Background Noise
Focal Spot,11
Scatter
Σ
III
D Tward (Med Phys 12
2009)
G Gang (Med Phys 2011)
System
Geometry
Anatomical
Background
κ
T
Σ11 et al. (1994)
Cunningham
Quantum
Noise
Task Type
Object
size
Projection
Object contrast
S(fx, fy, fz)
f
Focal Spot Size
Scatter-to-Primary Ratio
Magnification
β
NEQ
Observer
Model
Imaging
Task
PW, PWE
NPW, NPWE
Wtask = Ftask · Ctask
Detectability
Index
7
8/2/2012
Example
Flat-Panel
Detector
0.194 mm
pixels
2
Magnification
1x1 Readout
A Dedicated Musculoskeletal Extremity Scanner
X-ray source
1.8
Detectability Index (d’)
Quantum
High-Frequency
Task
Noise
(Edge Detection)
kVp, Dose
Scintillator, Detector
Pixel Aperture, 2D Sampling
Reconstruction Filter
Backprojection
Voxel Size
3D Sampling
1.6
1.4
β
Magnification
2x2 Binning
f 0.1
2
0.388 mm
pixels
1.8
For Example: Stationarity
2
1.8
Axial µ(x,y)
Noise (stdev) σµ(x,y)
1.6
1.4
Focal Spot Size
Scatter-to-Primary Ratio
0.8
Magnification
0.2
0.3
0.4
Pixel Size (mm)
Low-Frequency Task
(Muscle-Fat)NEQ
1.7
1.6
1.5
Observer
1.6
Model
Imaging1.4
Task
1.3
1.4
PW, PWE
Wtask = Ftask · C1.2
task
NPW, NPWE
1.1
1.2
Zbijewski et al. Med Phys 2011
Prakash et al. Med Phys 2011
Assumptions and Limitations
For Example: Stationarity
System1.2
Geometry
1
Anatomical
Background
1.2
κ
Assumptions and Limitations
0.1
Detectability
0.2Index0.3
Pixel Size (mm)
1
0.4
Angel Pineda et al.
Med. Phys. 39(6) (2012)
8
8/2/2012
Modeling Noise Stationarity
Example
CSA model for NPS
H2O cylinder (16 cm)
No bowtie filter
Polyenergetic beam
90 kVp
1 mGy
360 views
360o orbit
FBP
[ 0,80, 0]
x 10
-0.6
-7
2
For example: Penalized Likelihood
-7
[56,56,0]
1.8
x 10
-0.6
-0.4
2
1.6
Discretized
Object Volume
Forward model:
1. 8
1.4
-0.2
-0.4
1. 6
1.2
0
1. 4
-0.2
1
0
1
0. 8
0.2
0.4
Measurements
0. 6
0.2
0.4
0.6
-0.6
-0.4
-0.2
0
[0,40,0]
0.2
0.4
0.6
0. 4
0
-7
x 10
-0.6
[28,28,0]
x 10
0.6
-0.6
2
-0.6
1. 8
0. 2
-7
-0.4
2
-0.2
0
0.2
0.4
0.6
Log-likelihood estimator:
0
Projection
Operator
Number of
photons
0
1. 6
1.6
1. 4 -0.2
1.4
1. 2
1.2
0
1
0. 6
0.6
0. 4 0.4
0.4
0. 2
0.2
0.6
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
[0,0,0]
0.6
0
-0.6
x 10
-0.4
-0.2
-7
0
0.2
0.4
0.6
[40,0,0]
2
-0.4
-0.2
1. 6
1. 4 -0.2
1. 4
1.2
1. 2
1. 2
1
0. 8
0.2
0
0.2
0.4
0.6
Noise and spatial resolution are object-dependent and spatially variant
However, local covariance properties can still be estimated:
Measurement Covariance
0. 6
0. 4
0. 2
0. 2
-0.6
-0.4
-0.2
0
0.2
0.4
-0.4
0.6
0
Regularization
Operator
0.6
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0
12
-0.2
8
10
0
6
8
4
0.2
0.4
6
2
4
0.4
0.6
-0. 4
-0.2
[87, 0138]
0.2
0.4
0.6
2
-14
x 10
-14
[102, 174]
x 10
0.6
16 -0.6
-0. 6
14
12
Fessler et al. IEEE-TIP (1996)
J Web Stayman et al. AAPM (2010)
8
0.2
0.4
0.6
10
0
8
6
6
0.2
4
2
0.6
-0.4
-0.2
0
0.2
0.4
4
0.4
2
0.6
0.6
-0.6
-0.4
-0.2
-14
0
0. 2
0. 4
0.6
-14
[138, 189]
x 10
16-0.6
-0.6
-0.4
8
6
8
6
0.2
0.4
0.6
8
6
4
0.4
2
0.6
0
10
0
0.2
4
0.4
-0.2
12
10
0
0.2
2
-0. 4
14
-0.2
4
0.6
16
12
-0.2
0.4
x 10
14 -0.4
-0.4
12
0.2
-14
[138, 238]
x 10
-0.6
16
14
-0. 6
0
12
10
0
0.4
0
jth Unit Vector
(at voxel j)
-0.2
-0.2
10
Weighted ProjectionBackprojection Operator
-0. 416
14
-0.4
0.2
-0.6
14
10
-0.2
0. 8
0. 4 0.4
0
Regularization
Term
1
0. 6
0.4 0.4
0.2
0.6
-0.2
0
0.2
0.6
0.6
LogLikelihood
-0.4
1. 6
0
-0.4
1. 8
1.4 -0.2
0.8
-0.6
2
1. 8
-0.4
1
0.4
x 10
-0.6
1.6
0
0.2
-7
[80,0,0]
x 10
2
1.8
0
-7
-0.6
-0.6
Objective
Function
x 10
16
12
-0.2
0.8
0.2
0.4
-14
[67, 209]
-0.6
14
-0.4
1
0. 8
0.2
16
-0.2
-0. 6
-0.4
-0.2
x 10
-0.4
-0.6
1.8
-0.4
-14
[38, 138]
-0.6
Estimator model for NPS
H2O ellipse (32x16 cm)
No bowtie filter
Mono-energetic beam
µH2O ~0.018mm-1
1 mGy
360 views
360o orbit
PL reconstruction
Quadratic penalty
I0 = 5x105
β = 5x107
0.2
0.6
0.4
Example
0
1. 2
0.8
0.2
Modeling Noise Stationarity
Extensions of the Models
Non-Linear Reconstruction Algorithms
-0. 6
2
0.6
-0. 4
-0.2
0
0.2
0.4
0.6
-0. 6
-0. 4
-0.2
0
0.2
0.4
0.6
9
8/2/2012
Modeling Noise Stationarity
[90, 138]
x 10
-0.3
Example
4.5
-0.2
Estimator model for NPS
H2O ellipse (32x16 cm)
No bowtie filter
Mono-energetic beam
µH2O ~0.018mm-1
1 mGy
360 views
360o orbit
PL reconstruction
Quadratic penalty
I0 = 5x105
β = 5x107
4
3
0
2.5
2
x 10
-0.3
12
-0.2
10
0.1
-13
[115, 138]
x 10
0.5
12
-0.3
0.3
-0.3
-0.2
-0.1
0
0.1
0.2
8
6
-0.1
8
4
0
6
2
0.1
4
-0.3
7
-0.2
6
0.1
2
0.2
[138, 99]
1
2 -13
x 10
-13
x 10
-0.3
0.3
14
-0.3
-0.2
[138,-0.3
177]
-0.2
-0.1
0
-0.2
-0.1
0
0.1
0.2
10
8
0
0
8
8
6
0.16
0.1
4
0.2
0.2
6
4
4
2
2
0.3
0
0.1
0.2
0.3
due to variation in Nphotons at the detector.
due to a finite number of projections.
due to the cone-beam effect.
but we can still model the local NPS.
All of the above.
12
10
-0.1
0.2
-0.3
20%
1.
2.
3.
4.
5.
-0.212
0.1
0.3
0.3
14
0.3
12
0
-13
0.1
0.2
x 10
-0.3
14
-0.1
10
-0.1
17%
0.3
0.2
0.3
Tang et al.
0.2
-0.2
-0.1
13%
0
3
-0.3
-0.2
23%
5
-0.1
10
-0.2
0.3
-0.3
CT image noise is non-stationary:
27%
-13
4
0.2
0.1
x 10
-13
0.3
0
0
Waiter, Check Please…
1
0.2
0.1
-0.1
[100, 195]
1.5
-0.1
-0.2
DE Radiography
Cone-Beam CT
Phase Contrast
Tomosynthesis
DE CBCT
3.5
-0.1
[100, 81]
-0.3
Extensions of the Models
2D, 3D, Dual-Energy, Phase Contrast, …
-13
5
2
0.3
-0.2
-0.1
0
0.1
0.2
0.3
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
Richard et al.
Ducote et al.
Zhao et al.
Glick et al.
Tward, Gang et al.
Gang et al.
Fredenberg et al.
Tang et al, Chen et al.
10
8/2/2012
Waiter, Check Please…
CT Imaging Performance
The System Design Perspective
CT image noise is non-stationary:
Technical Assessment
4. but we can still model the NPS.
Must account for complexities in scanner configuration
For example, Cone-Beam CT:
- Dose measurement
- Fully 3D spatial resolution and noise characteristics
Must account for complexities in reconstruction methods
For example, statistical / iterative reconstruction
- Nonlinearity: spatial resolution dependent on signal
- Nonstationarity (may be better or worse than FBP)
Must acknowledge assumptions and limitations of the metrics
For example: LOCALITY
5. All of the above.
Technology Development
1. due to variation in Nphotons at the detector.
2. due to a finite number of projections.
3. due to the cone-beam effect.
Strengthened by a foundation in imaging physics
Accelerates translation to clinical application
Extension to New Techniques
Baek et al. Med Phys 37(5) (2010)
Pineda et al. Med Phys 39(6) (2012)
New modalities (PCXD, PCCT, etc.) and algorithms (model-based)
New challenges for modeling and measurement standards
11
8/2/2012
Fundamental Image Science
Kg
Jθ
Statistical decision theory
Intricacies of the human visual system
NPS
MTF
DQE
NEQ
Imaging Physics & Engineering
Statistical descriptions of signal and noise
Measurement and modeling
Measuring the Noise
Basic Technical Assessment
Noise-Equivalent Quanta (NEQ)
Radiation Dose
Farmer chamber + 16 cm cylinder
Short-scan protocols
Quantitative Accuracy
The 3D NEQ
Effective number of quanta used at
each spatial frequency
(Efficiency x Fluence)
Electron density inserts
Comparison to MDCT
Technology Development
ROC
Design and optimization
Accelerate translation / pre-clinical testing
d’, Az
Technical Assessment and QA
d'
Physical measurements. Practicality
Protocol optimization
Clinical Assessment
SDNR
Contrast-Detail
Dose
Diagnostic performance
Complex scenes and imaging tasks
Sensitivity, specificity, cost-benefit
Contrast Resolution
Low-contrast tissue inserts
SDNR versus kVp, mAs
NEQ = θ tot f
MTF 2
NPS
Axial NEQ
fy (mm-1)
… andPerformance
the Metrics
Perspectives on Imaging
Spatial Resolution
Line-pair pattern (subjective)
Modulation transfer function (MTF)
“Clinical” Image Quality
Anthropomorphic phantoms
Expert readers
J Xu et al. Med. Phys. (in press - August)
Observations:
NEQ(f) / mq0 ~ “DQE(f)”
NEQ(0) / mq0 ~ Projection DQE(0)
3D NEQ(f) dependent on reconstruction parameters
Units (fluence): [photons / mm 2]
fx (mm-1)
12
8/2/2012
Measuring the Noise
Validity of the Models
Extensions of the Models
Noise-Equivalent Quanta (NEQ)
Comparison to Human Observers / Simple Tasks
Non-Linear Reconstruction Algorithms
For example: Penalized Likelihood
Uniform Background
Sphere Detection
Axial NEQ
PW, PWEi
PWEi, NPWE
1.0
fy (mm-1)
1.0
NPWEi
0.9
Human
Observer
Az
B
Az
0.7
0.6
fx (mm-1)
0.5
NPWE
0.9
0.8
B
fz (mm-1)
PW, PWEi, NPW, NPWE
fx (mm-1)
These results fairly old.
Replace with new nonstat
cov, nps for PL
0.8
0.7
0.6
0
60
120
θtot
180
0.5 0
0
NPWNPW
Local NPS
Sagittal NEQ
Cluttered Background
Large Sphere
Sphere
Small
NPWEi
NPWEi
Human
Human
Observer
Observer
60
Measurement Covariance
120
Regularization
Operator
180
180
θtot
G Gang et al. Med. Phys. 2011
Weighted ProjectionBackprojection Operator
jth Unit Vector
(at voxel j)
Fessler et al. IEEE-TIP (1996)
J Web Stayman et al. AAPM (2010)
13