Post on 24-Aug-2020
transcript
1 /
GE /
1
Jiang Hsieh, Jean-Baptiste Thibault, Jiahua Fan
GE Healthcare
Advanced Reconstruction Methods on GE CT Systems
2
GE Dose Reduction Technology
Sn
ap
-sh
ot
Pu
lse
AS
iR
MB
IR (
Ve
o)
He
lic
al
Sh
utt
er
He
lic
al
Sh
utt
le
Au
to-m
A
Sm
art
-mA
Co
llim
ato
r T
rac
kin
g
Sm
art
-Pre
p
Ne
uro
-Fil
ter
Dy
na
mic
Pro
jec
tio
n F
ilte
r
Z-s
mo
oth
ing
Fil
ter
EK
G-m
A-m
od
ula
tio
n
Co
lor
Co
de
fo
r K
ids
Co
lle
cti
on
Cu
p
Ba
ck
lit
Dio
de
1990 2010 single slice multi-slice volumetric
AS
iR-V
Org
an
-do
se-M
od
ula
tio
n
3
Strong collaborations feed innovation cycle
Clinical feedback Technology development
Iterative Reconstruction Technology
Veo is 510(k) pending. Not available for sale in the U.S.
2 /
GE /
4
FBP vs. Model-based Iterative Reconstruction
MBIR maximizes the probability that the reconstructed result matches the acquired data
according to an accurate model of the data acquisition process
compare
acquired synthesized
Closed-form Solution
FBP
MBIR: a probabilistic view of CT Reconstruction based on X-ray physics
MBIR: a statistical view of reconstruction
yxPxx
MAP |maxargˆ
Bayesian statistical estimation framework
xUAxyWAxyxT
xMAP
2
1minargˆ
MBIR statistical cost function (physics model):
Object model
System model
Noise model
Iterative optimization
Cost Function
nAxy Accounting for noise in the measurements with
The MBIR cost function defines the image quality
6 Page 6
Iterative Reconstruction Technology
Object
Optics
Physics
Noise
)(,minargˆ xGyAxLxx
MBIR method Models Cost Function Algorithm
FBP MBIR
Modeling Accuracy Leads to Better Image Quality
3 /
GE /
7
System Optics Model
Accurate description of system optics / data acquisition
process 3D discrete nature of scanned
object Realistic spatially-varying
model
Account for focal-spot and X-ray detector nonlinearities
Can include X-ray physics (BH,
bone, etc.)
“ideal” optics
“realistic” optics Higher spatial resolution
Artifact reduction
8
Noise Model Line
integrals
Variance map
Noise reduction Artifact
reduction
Based on x-ray physics (Poisson-Gaussian counting process)
Relative degree of confidence in each measurement as it relates to the image
Advanced modeling for low signal and electronic noise
FBP MBIR
9
Object Model Example: Typical prior model:
Markov Random Field (MRF)
Local behavior Probability distribution
Edge-preserving: local freq. info for adaptive processing
Controls image texture, noise / resolution / contrast
performance o Limited: global penalty model
over full 3D image volume
Local difference
between voxel and 26
neighbors in 3D
Ckj
kjqkjqxxb
qxU
,
1)(
Prior information -> medical images Stabilizes estimate for under-determined
reconstruction problem Probability distribution over the image
Convex parameterization for cost function optimization
Adjusts to statistics and local sampling
Nonlinearities -> linearize in clinical range
Allows flexibility in MBIR settings
Re-defines trade-offs between noise, resolution, and contrast
performance for CT imaging
4 /
GE /
10
Object Model
Min parameter: prior strength to balance data fit with image model
Extra parameters for edge-preserving behavior
User can adjust to fine-tune behavior for specific clinical applications
FBP MBIR Stnd MBIR RP05 MBIR RP20
Adjusts to clinical
task
11
Clinical Example Oncology abdomen/pelvis 0.61mSv*
Veo (MBIR) FBP 120kVp, 5mAs, 0.625mm
Courtesy of Dr Barrau, CCN, France
Liver Metastasis
Dose report
*Obtained by EUR-16262 EN, using an abdomen factor of 0.015*DLP and a pelvis factor of 0.019*DLP
12
FBP
Clinical Example Case showing blooming reduction in the stent
Veo
5 /
GE /
13 Page 13
Physics
Noise
Optics
Object
Physics
Reconstruction Algorithms
FBP ASiR ASiR-V Veo
Noise Reduction,
Dose Reduction, LCD Improvement
Noise Reduction, Dose
Reduction, LCD Improvement, Artifact Reduction
Noise Reduction, Dose
Reduction, LCD Improvement, Artifact Reduction, Spatial
Resolution Improvement
Noise Noise
Object Object Physics
Noise
Object Physics
Optics
Primary goal: dose reduction
Speed vs. performance tradeoff
From FBP to MBIR CTDIvol = 1.4 mGy. DLP = 61.8 mGy-cm. Effective dose 0.93 mSv
FBP ASiR 50% ASiR-V 50% Veo
15
Task-based Performance Evaluation
• Phantom: MITA LCD Phantom • Observer: CHO • Channel: D-DOG1
• Recon: FBP & GE ASiR-V • CT scanner: GE CT systems • Head mode: LROC Analysis • Body mode: ROC Analysis • Figure of merit: AUC • Variance estimation: MRMC Methods
– One-Shot2
Reference: 1. C.K. Abbey and H.H. Barrett, “Human- and model-observer performance in ramp-spectrum
noise: effects of regularization and object variability,” Vol. 18, No. 3/March 200/JOSA. 2. B. Gallas, “One-Shot Estimate of MRMC Variance: AUC,” Acad Radiol 2006; 13:353-362.
Task: Signal detection (and localization)
Observer: Channelized Hotelling observer
as a model of human observers
Figure of merit: Area under the ROC or
LROC curve
6 /
GE /
16
0.5
0.6
0.7
0.8
0.9
1
Body Mode ROC:5mm 7HU
200mAs-FBP
40mAs-FBP
40mAs-IR
Body ROC example: 5mm 7HU
Sample Results
0
0.1
0.2
0.3
0.4
0.5
0.6
Head Mode LROC:10mm 3HU
120mAs-FBP
30mAs-FBP
30mAs-IR
Head LROC example: 10mm, 3HU
17
Example of phantom images
FBP ASiR-V
18
Example of clinical Images
FBP ASiR-V
7 /
GE /
19
Summary
• Dose reduction has been one of the key CT technology
drivers for the past two decades.
• The continued development of iterative reconstruction
technology will likely fundamentally change the
operation of a CT scanner (ASiR: thousands of global sites with
millions patients, Veo: dose reduction and image quality
improvement, ASiR-V: latest technology)
• Dose reduction is a journey and requires the
participation from CT manufactures, CT physicists, CT
operators, radiologists, professional organizations, and
government agencies.