1. Evaluation of clinical Full Field Digital
Mammography with the Task-Specific-
System-Model-Based Fourier (TSMF)
SNR
Haimo Liu
Fischell Department of Bioengineering
University of Maryland-College Park
FDA CDRH OSEL/DIAM
Haimo Liu: haimo.liu@fda.hhs.gov
2. Outline
Task specific evaluation methodology
Development
Benefits of using phantom for FFDM image
quality assessment
Application on clinical systems
GE Senographe DS
Comparison between two image acquisition modes
Validation of the method
Hologic Selenia
Conclusion
2 Haimo Liu: haimo.liu@fda.hhs.gov
3. Background: current FFDM evaluation
FFDM system
performance
Clinical trials Mean pixel value, Fourier based Image based
standard deviation method method
Expensive,
time consuming
Detector based:
Diff. signal
MTF, NPS
Human observers: Pixel SNR Cov. matrix
radiologists
Limitations: No system
Suffer from model
background DQE
reader’s variability
variability, signal Hotelling SNR
shape and size Does not account
for scatter, focal Large number
spot unsharpness of images
3 Haimo Liu: haimo.liu@fda.hhs.gov
4. Motivation
Develop Fourier based
X-ray tube Compression paddle
evaluation methodology
for clinical FFDM systems
Entire FFDM image
acquisition
With phantom that models
breast attenuation
Objective, task based
Practical, collection of
limited number of images
Empirical model of the
system, results not limited
to specific settings
Detector Computer console
4 Haimo Liu: haimo.liu@fda.hhs.gov
5. Assumptions
Linear system
No non-linear/adaptive image processing
Cyclostationary system
Stationary noise
Shift-invariant system transfer function
Infinite detector area
5 Haimo Liu: haimo.liu@fda.hhs.gov
6. Phantom assembly
Inspired by the CDMAM
phantom
Same uniform background
Same HVL
Simulated signals
Uniform PMMA plates
Four 1 cm plates
Aluminum plate
0.5 mm thick
Model the Aluminum base
(where signals are
attached to) of the
CDMAM phantom
6 Haimo Liu: haimo.liu@fda.hhs.gov
7. Clinical systems
System GE Senograhe Hologic Selenia
DS
Detector Indirect Direct
Pixel size 0.1mm ×0.1mm 0.07mm ×0.07mm
Kvp 30 30
Target/Filt Mo/Mo Mo/Mo
er
Image A:Fine View Mode Phantom Mode
acquisition B:Standard Mode
mode
Grid Linear Cross-hatch grid
7 Haimo Liu: haimo.liu@fda.hhs.gov
9. Generalized Modulation Transfer Function (GMTF)
GMTF
The modulus of the Fourier transformation of the line
response function measured within the breast phantom
Five images of a copper edge placed between the PMMA
plates
Three detector entrance exposures
Convert to mean glandular doses
2D GMTF
Spline interpolation between 1D profiles along the two
axes
X direction: parallel to chest wall
Y direction: perpendicular to chest wall
9 Haimo Liu: haimo.liu@fda.hhs.gov
10. Evaluation of scatter removal method (GE)
MTF (x direction)
100 mAs, x-direction
Without phantom
System without grid gives
better performance
GMTF(x direction)
100 mAs, x-direction
With phantom placed in
the FOV
System with grid gives
better performance
10 Haimo Liu: haimo.liu@fda.hhs.gov
11. GMTF (x direction) for two clinical systems
GE
11 Haimo Liu: haimo.liu@fda.hhs.gov
12. GMTF along the two axes
12 Haimo Liu: haimo.liu@fda.hhs.gov
13. Generalized Normalized Noise Power Spectrum (GNNPS)
GNNPS
The square of the Fourier transformation of the system
noise measured at the center of the breast phantom
Five images of the background phantom
Three detector entrance exposures
Convert to mean glandular doses
13 Haimo Liu: haimo.liu@fda.hhs.gov
14. GNNPS (x direction) for two clinical systems
14 Haimo Liu: haimo.liu@fda.hhs.gov
16. Methodology validation – comparison between methods
Fourier based method with GMTF 2 FT[ΔSs ]2
SNR =
simulated signals (Ss): GNNPS
Phantom Image based method
(image based signal: SI):
SNR = ΔS I T K -1ΔS I
Fourier based method with FT[ΔSI ]2
phantom image based signals, SNR =
GNNPS ROI size 256×256: GNNPS256 256
Fourier based method with
FT[ΔSI ]2
phantom image based signals, SNR =
GNNPS ROI size 19×19: GNNPS19 19
16 Haimo Liu: haimo.liu@fda.hhs.gov
18. Contrast-Detail curve
Contrast-Detail
(CD) curve
Four alternative
forced choice (4
AFC) task
Link SNR to detection
probability
62.5% detection
probability as the
threshold
Threshold thickness
Generate CD curve
18 Haimo Liu: haimo.liu@fda.hhs.gov
19. Contrast-Detail curve: system performance prediction
CD curve at unit dose
SNR2 is linearly proportional
to dose
Normalize SNR by dose
Normalize CD curve by dose
Can be used to predict
system performance
Two clinical systems
Different noise and
deterministic properties
Identical CD curves
19 Haimo Liu: haimo.liu@fda.hhs.gov
20. Contrast-Detail curve: human performance prediction
Human performance
prediction (GE Senographe
DS)
SNR 2
Human
FHuman 2
30% 5%
SNR TSMF
TSMF CD curve at 51 μGy
TSMF CD curve (adjusted by
human efficiency) at 51 μGy
Human CD curve at 70 μGy (GE
Senographe 2000D)
Human CD curve at 140 μGy (GE
Senographe 2000D)
20 Haimo Liu: haimo.liu@fda.hhs.gov
21. Conclusions
Fourier based evaluation methodology for clinical FFDM
Provide more information of the system using phantom
Scatter from the phantom
Focal spot unsharpness
Magnification
Create a model of the system
Not limited to specific system settings
Predict system performance
Get closer to link image quality to diagnostic performance of the
system
Clinical application
GE Senographe
Hologic Selenia
21 Haimo Liu: haimo.liu@fda.hhs.gov
Editor's Notes
Divide into chapters: 1. benefits of using phantom 2. comparison between two modes for GE 3. validation 4. application on HologicWhat I proposed to go in the paper, I need your opinion of how I should present the dataIn the end: title, coauthor list, how to present, which journalYour opinions and suggestions are welcome, as well as how should I present the data
Clinical trial: Sensitivity and specificity of the technology Pissano: 2008-10, FFDM Guidance More information, create model of system, predict performance for other settingsLink between image quality and diagnostic performanceIncluding phantom- extra information, getting steps closer to link image quality to diagonostic…Limited to specific settings- gagne’s & channelsThis figure shows the current FFDM evaluation methods. First, FFDM performance can be evaluated by running clinical trials, they are ……the clinical trials give the diagnostic performance of the system. However, it has been shown in the literature that there is a link between the diagnostic performance of the system and the image quality, therefore ppl developed methods for evaluating……to represent the diagnostic performance of the system.
Same order of the legendPut it in the beginningMake a link between MTF and SNR, mtf itself is not enough for image quality assessmentBetter explain the plots
Change yellow to sth else, change curve style
NPS does averaging
This plot shows the threshold detectability, or the threshold of detectable discs with specific radius and thickness for a given system setting. For example, if we pick up “this dashed red line”, it shows the threshold detectability when using image acquisition mode B for a given dose: 0.65 mGy. All the signals below this line can not be detected, and all the signals above this line can be detected. Here, 75% detection probability was used as the threshold, which is b between 50% and 100%, where 50% is ….different colors indicate three different exposures or doses, dashed lines are the mode B and solide lines are the mode A. notice that, for all three exposures, the mode a has better performance then mode B because the threshold value is lower. However, even though this difference is consistent, it is still not clear if this difference is signifancant. Error bars and confidance intervals will be estimated in the future research to determine if the difference is significant statistically.Empty squares and circles