Document 178266

Physics for Dummies:
How to optimise your kit
XRM, DBT and CEDM
Professor Ken Young
National Co-ordinating Centre for the Physics of Mammography,
Guildford, UK
& University of Surrey, Guildford
www.nccpm.org
OPTIMISING X-ray Mammography
 Digital Mammography
 Digital Breast Tomosynthesis
 Contrast Enhanced Digital Mammography
We have been working on how to optimise
DM and DBT systems for 5 years in the
OPTIMAM research project – and have
just started OPTIMAM2
Our Experimental Approach
Model the image
formation process
Real and simulated
digital images
of breasts
Real and simulated
features of cancers
CANCER
DETECTION
OBSERVER
STUDIES
comparing
different
factors
Use results
to optimise
design,
choice
and use
of new
technology
in
cancer
screening
What have we learned in OPTIMAM
Effects on cancer detection of: Dose
 Detector type
 Image processing
 Link to image quality standards
Observer studies in OPTIMAM
1. Calcification
2. Mass & calc
processing
3. Mass & calc
4. Mass & calc
detection v image quality
detection v image
detection v image quality
detection: 2D v Tomo
L. Warren et al, Medical Physics 2012
AFROC Curves for IQ levels
(fitted to average of 7 observers)
Lesion localisation fraction (LLF)
1.0
0.8
Normal dose DR
Half dose DR
0.6
0.4
Quarter dose DR
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Fraction of normal images with at least one NL
L. Warren et al, Medical Physics 2012
1.0
AFROC Curves for IQ levels
(fitted to average of 7 observers)
Lesion localisation fraction (LLF)
1.0
0.8
Normal dose DR (Agfa)
Half dose DR
0.6
Normal dose CR
Half dose CR
0.4
Quarter dose DR
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Fraction of normal images with at least one NL
L. Warren et al, Medical Physics 2012
1.0
Calcification cluster sensitivity at 0.1 FP per image
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
72%
47%
42%
30%
27%
Normal
Dose DR
Half dose
DR
L. Warren et al, Medical Physics 2012
Quarter
dose DR
Normal
dose CR
Half dose
CR
Reader-averaged JAFROC FoM
Does CDMAM test object predict calcification detection?
1.0
0.8
better
0.6
R2 = 0.94
p = 0.04
0.4
0.2
0.1
DR
CR
better
0.2
0.3
0.4
Threshold gold thickness (µm) - 0.25mm disc diameter
L. Warren et al, Medical Physics 2012
Reader-averaged JAFROC FoM
Does CDMAM test object predict calcification detection?
achievable
1.0
minimum
0.8
0.6
R2 = 0.94
p = 0.04
0.4
0.2
0.1
DR
CR
0.2
0.3
0.4
Threshold gold thickness (µm) - 0.25mm disc diameter
L. Warren et al, Medical Physics 2012
Observer studies in OPTIMAM
1. Calcification
2. Mass & calc
processing
3. Mass & calc
4. Mass & calc
detection v image quality
detection v image
detection v image quality
detection: 2D v Tomo
L. Warren et al, Medical Physics 2012
Image processing comparison
Hologic software
Standard version
Low contrast version
L. Warren et al, American Journal of Radiology, 2014
Simulated screen-film
masses
1.0
0.8
0.6
0.4
Standard version
Low contrast version
Simulated screen-film
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Fraction of lesions localised
Fraction of lesions localised
Image processing comparison
calcifications
1.0
0.8
0.6
0.4
Standard version
Low contrast version
0.2
Simulated screen-film
0.0
0.0
False positive fraction
L. Warren et al, American Journal of Radiology, 2014
0.2
0.4
0.6
0.8
False positive fraction
1.0
masses
1.0
0.8
0.6
0.4
Standard version
Low contrast version
Simulated screen-film
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Fraction of lesions localised
Fraction of lesions localised
Image processing comparison
calcifications
1.0
0.8
0.6
0.4
Standard version
Low contrast version
0.2
Simulated screen-film
0.0
0.0
False positive fraction
L. Warren et al, American Journal of Radiology, 2014
0.2
0.4
0.6
0.8
False positive fraction
1.0
masses
1.0
0.8
0.6
0.4
Standard version
Low contrast version
Simulated screen-film
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Fraction of lesions localised
Fraction of lesions localised
Image processing comparison
1.0
0.8
0.6
0.4
Standard
Low contrast
Simulated screen-film
Standard version
Low contrast version
0.2
Simulated screen-film
0.0
0.0
False positive fraction
Image processing
calcifications
1.0
0.2
0.4
0.6
0.8
False positive fraction
Figure of merit
Masses
Calcifications
0.73
0.72
0.72
L. Warren et al, American Journal of Radiology, 2014
0.65
0.63 (p = 0.04)
0.61 (p = 0.0005)
1.0
Observer studies in OPTIMAM
1. Calcification
2. Mass & calc
processing
3. Mass & calc
4. Mass & calc
detection v image quality
detection v image
detection v image quality
detection: 2D v Tomo
A. Mackenzie et al, SPIE Medical Imaging 2014
Types of image detector compared
• Starting image set was converted to:
–
–
–
–
Arm 1: ‘a-Se’ detector
Arm 2: ‘CsI’ detector
Arm 3: ‘NIP CR’ detector
Arm 4: ‘Powder phosphor CR’ detector
AGD = 1.08 mGy for 50 to 60 mm breast thickness
• Image processing
– Agfa Musica2
– Suitable for a wide range of image qualities
A. Mackenzie et al, SPIE Medical Imaging 2014
Calcification cluster detection and recall
Would you recall on the basis of this lesion?
CsI v. NIP
CsI v. CR
NIP v. CR
0.0001
<0.0001
0.048
Lesion localisation fraction
IQVery
pairs
No:
confidentp-value
No:
Moderately
a-Se
v. CsI confident
0.27
No: Slightly confident
a-Se
v. NIP
<0.0001
Yes:
Slightly
confident
Yes:
Moderately
a-Se
v. CR confident
<0.0001
Yes: Very confident
1.0
0.8
a-Se
0.6
CsI
NIP
0.4
CR
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Non-localisation fraction
A. Mackenzie et al, SPIE Medical Imaging 2014
1.0
Calcification cluster detection and recall
Fraction of lesions
0.8
Marked
0.6
0.4
0.2
0.0
Arm 1
'a-Se'
Arm 2
'CsI'
Arm 3
'NIP CR'
Arm 4
'Powder
CR'
A. Mackenzie et al, SPIE Medical Imaging 2014
Calcification cluster detection and recall
Fraction of lesions
0.8
Marked
Recalled
0.6
0.4
0.2
0.0
Arm 1
'a-Se'
Arm 2
'CsI'
Arm 3
'NIP CR'
Arm 4
'Powder
CR'
A. Mackenzie et al, SPIE Medical Imaging 2014
Compare
with a-Se
NIP
CR
Change
in recall
-28%
-44%
Non-Calcification detection and recall
No: Very confident
IQ pairs
p-value
No: Moderately confident
a-Se
v. CsIconfident
0.93
No:
Slightly
Yes:
Slightly
a-Se
v. NIPconfident
0.002
Yes: Moderately confident
a-Se
v. confident
CR 0.0007
Yes:
Very
CsI v. NIP
CsI v. CR
NIP v. CR
0.002
0.0009
0.77
Lesion localisation fraction
Would you recall on the basis of this lesion?
1.0
0.8
CsI
a-Se
0.6
CR
0.4
NIP
0.2
0.0
0.0
0.2
0.4
0.6
0.8
Non-localisation fraction
A. Mackenzie et al, SPIE Medical Imaging 2014
1.0
Non-Calcification detection and recall
Fraction of lesions
0.8
Marked
0.6
0.4
0.2
0.0
Arm 1
'a-Se'
Arm 2
'CsI'
Arm 3
'NIP CR'
Arm 4
'Powder
CR'
A. Mackenzie et al, SPIE Medical Imaging 2014
Non-Calcification detection and recall
Fraction of lesions
0.8
Marked
Recalled
0.6
0.4
0.2
0.0
Arm 1
'a-Se'
Arm 2
'CsI'
Arm 3
'NIP CR'
Arm 4
'Powder
CR'
A. Mackenzie et al, SPIE Medical Imaging 2014
Non-Calcification detection and recall
Fraction of lesions
0.8
Marked
Recalled
0.6
Compare
with a-Se
NIP
CR
0.4
0.2
0.0
Arm 1
'a-Se'
Arm 2
'CsI'
Arm 3
'NIP CR'
Arm 4
'Powder
CR'
A. Mackenzie et al, SPIE Medical Imaging 2014
Change
in recall
-6.5%
-9.1%
Relationship between
physical image quality &
cancer detection?
Calcification clusters
Reader averaged
AFROC AUC
0.9
0.8
Calcifications v 0.25 mm
Achievable Acceptable
a-Se
Clear relationship
between calcification
detection and CDMAM
results
0.7
CsI
CR
NIP
0.6
R2 = 0.98
0.2
0.3
Threshold gold threshold
thickness (m)
0.4
Is there a need to
review image quality
guidelines?
Non-calcification lesions
Reader averaged
AFROC AUC
0.9
Relationship is less
strong than for
calcifications
Non-calcifications v 1 m m
Achievable Acceptable
0.8
0.7
0.6
a-Se
CsI
CR
2
R = 0.73
0.04
0.06
NIP
0.08
0.10
0.12
Threshold gold threshold
thickness (m)
How well are systems in
NHSBSP optimised?
Plot dose v image quality
from data for 318 systems
Hologic systems in NHSBSP
OPTIMAL
Review of QC measurements, Young, NHSBSP 2014
Siemens systems in NHSBSP
OPTIMAL
Review of QC measurements, Young, NHSBSP 2014
GE systems in NHSBSP
OPTIMAL
Review of QC measurements, Young, NHSBSP 2014
Philips/Sectra systems in NHSBSP
OPTIMAL
Review of QC measurements, Young, NHSBSP 2014
Fuji systems in NHSBSP
Review of QC measurements, Young, NHSBSP 2014
Proportion exceeding achievable IQ
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
82%
59%
83%
59%
50%
25%
0%
0%
28%
5%
30
0D
CR
L
0
i
j
0
a
2
Fu
ctr
e
GE
S
o)
tto
ion
M
t
o
(
i
a
G
ia
pir
n
s
e
l
In
Se
D
GE
Digital system
Review of QC measurements, Young, NHSBSP 2014
S
l
ns
W)
tia
(
o
n
i
s
se
n ia
en
e
l
Es
m
Se
Di
How do we check image
quality at other
thicknesses?
Use contrast-to-noise ratio
with blocks of PMMA
Measurement of contrast-to-noise ratio
0.2 mm Al
6 cm
20 mm perspex
breast support table
+ add Perspex to simulate thicker breasts
Review of QC measurements, Young, NHSBSP 2014
AEC performance for Hologic Dimensions
IQ measurement
Review of QC measurements, Young, NHSBSP 2014
AEC performance for Siemens Inspiration
IQ measurement
Review of QC measurements, Young, NHSBSP 2014
How will we optimise
DBT and CEDM?
OPTIMAM2 MAIN OBJECTIVES
X-ray Tube
 Optimise the
introduction of
tomosynthesis into
breast screening
Movement
Compression
paddle
Breast
No Grid
 What DBT system design?
 What IQ standard?...Phantoms?
 What dose?
10/06/2014
Centre of
Rotation
Flat Panel
Detector
Observer studies in OPTIMAM
1. Calcification
2. Mass & calc
processing
3. Mass & calc
4. Mass & calc
P. Elangovan, RSNA 2014
detection v image quality
detection v image
detection v image quality
detection: 2D v Tomo
DM v Tomo for detection of spheres and masses
P. Elangovan et al, RSNA 2014
Conclusions






IQ measured using the CDMAM correlates with
radiologists’ performance in DM
Good IQ is important for calcification detection (and
to a smaller extent masses) and systems should
exceed achievable levels in EU protocol. Acceptable
level is too low
Image processing affects calcification detection but
so far not masses – needs further research
Important to choose a good DM system and use
sufficient dose
AECs need better design to give more dose to
thicker breasts
DBT much better than DM for mass detection, but
may be worse for calcifications
Main Colleagues in OPTIMAM
NCCPM,
Guildford
CVSSP,
University of
Surrey
University
of Leuven
Addenbrooke St
’s Hospital,
George’s
Cambridge
Hospital
Jarvis
Centre,
Guildford
David
Dance
Kevin
Wells
Hilde
Bosmans
Mathew
Wallis
Ros
GivenWilson
Julie
Cooke
Alistair
Mackenzie
Prem
Elangovan
Emmy
Shaheen
Paula
Wilsher
Charul
Patel
Lucy
Warren
Oliver
Diaz
Nick
Marshall
Padraig
Looney
Alaleh
Rashidnasab
Lesley
Cockmartin
Mark HallingBrown
+ Students