PosterPDF Available

Abstract and Figures

Accuracy and robustness of MR studies depend directly on the quality of the acquired imaging data that relies on imaging hardware, software, work processes, and research objectives. This is especially relevant for multi-center studies and data sharing projects that have to deal with a variety of different image properties. To simplify handling of image quality measures we developed a quality rating and combination scheme.
Content may be subject to copyright.
INTRODUCTION
Accuracy an d robustness of MR stud ies depend directly on
the qua lity of the acquired i maging dat a (1-4). Fur thermore,
the images quality depends on imaging hardware (1-5),
software ( 2-5), work process es (6) and speci c res earch ob-
jectives (7-10). This is especially relevant for multi-center
studies and data sharing projects that have to deal with a
variety of di erent properties (7-10).
In (5) we showed how to measure the images quality pa-
rameters such as noise, inhomogeneity, and resolution in
the preprocessi ng process. To simplify handlin g we devel-
oped a single rating (IQM ) based on di erent quality mea-
sures (QM ). We have evalu ated this ratin g for synthetic a nd
real data sh owing that it is more accurate compared to the
single QMs and that it allows an objective, optimized and
accelerated quality assurance even for non-experts.
METHODS
First, all QMs were sc aled as simple school marks (Table 1).
The scaling was performed on the basis of theoretical as-
pects and the analysis of sy nthetic and real MRI data (Table
2). Finally, the QMs were merged into one single rating
by using a weighted root mean square (RMS) equation to
control the impact of each QM and to weight a poor mea-
surement stronger than a good one.
The relation between the QM and the preprocessing qual-
ity (described by the average Kappa) was modeled as a
linear equation “Kappa = QM * w”, where w describes the
in u ence or w eight ing o f eac h QM . Th en th e eq uatio n was
solved by a least square approach for w for the BWP data-
set. Thus, the results worked well for the BWP, but not for
the real lif e da ta. Ther efore , th e sim ple weigh ting of w=[0. 5
0.0 0.5 ]’ f or “ IQM =QM * w ” wi th “ QM=[ NCR ICR RES] ” (N oise
Cont ras t Ra tio, Inh omo gene ity Co ntras t, RES olut ion (5)) was
used as far as the inhomogeneity had marginal e ects for
mos t pr epro cess ing meth ods that al lowe d a near ly loss- les s
corr ecti on.
Quality Assurance in Structural MRI
Robert Dahnke1, Gabriel Ziegler1,2, Julian Grosskreutz3, Christian Gaser1,
1 - Structural Brain Mapping Group, Department of Psychiatry / Neurology, Jena University Hospital, Germany
2 - Wellcome Trust Centre for Neuroimaging, Institute of Neurology, University College London, United Kingdom
3 - Department of Neurology, Jena University Hospital, Germany
E-mail: robert.dahnke@uni-jena.de
PDF of the poster is available at: http://dbm.neuro.uni-jena.de/HBM2015/Dahnke.pdf
The average Kappa of 5 di erent tissue segmentations of
the BWP (Brain Web Phantom) dataset with 300 images
with varying noises, inhomogeneity and resolutions were
used to create 3 quality levels for testing the average re-
constructed quality against IQM and the single QMs.
Furthermore, a human expert cl assi ed images of 5 di er-
ent sites into 3 quality levels - no, slight, and strong inter-
ferences (rem oving necessary) - for grou p wise evaluation.
Then t he IXI dat aset was u sed to te st for age and sex e ects,
whereas the OASIS and ADHD200 datasets were used to
test for disease e ects. Finally, the INDI dataset was used
to compare the images qualities regardin g di erent scan-
ning sites.
To show the correlation between image quality and the
n al analysis parameter, we estimated the relati ve GM vol-
ume as global measurements.
RESULTS
The BWP Kappa groups clearly showed that the IQM
allows a better prediction of image and preprocessing
qualities than a single QM and that lower image qual-
ity comes along with a slightly underestimated relative
GM volume. The measurements show a similar behav-
ior for the real expert groups like the BWP. However,
one major difference is that all real dataset have nearly
the same resolution around 1 mm. As a result the noise
measurements NCR allow a similar good rating like the
IQM. Therefore, the strong reduction of the relative GM
volumes for lower image qualities is more interesting (all
groups have a similar age range and health status).
The IXI dataset clearly shows that there are no age or
sex effects for the IQM. As expected a clear relative GM
volume reduction in aging is visible. There are small dif-
ferences between healthy and non-healthy subjects and
also outliers with low image quality have often less GM.
The comparison of the INDI center shows great visual
and measurable differences in image quality. Because
the mean age varies strongly for each site, a comparison
to the relative GM volume is not applicable here.
CONCLUSIONS
The detection of WMHs is an important task in MR image
processing. Especially severe WMHs require corrections fo r
correct preprocessing. Our method was able to improve
the results of the segmentation and normalization on the
visual and the numeric level.
ACKNOWLEDGEMENTS
Julian Grosskreutz is supported by the SOPHIA project.
REFERENCES
(1) Blaimer M., Breuer F., Mueller M., Heidemann R.M.,
Griswold M.A., and Jakob P.M., ‘SMASH, SENSE, PILS,
GRAPPA: how to choose the optimal method’, Top
Magn Reson Imaging, 2004 vol. 15 (4) pp. 223-236
(2) Collins D.L., Zijdenbos A.P., Kollokian V., Sled J.G., Kabani N.J.,
Holmes C.J., and Evans A.C., ‘Design and construction of a
realistic digital brain phantom’, IEEE transactions on medical
imaging, 1998 vol. 17 (3) pp. 463-468
(3) Sled J.G., Zijdenbos A.P., and Evans A.C., ‘A Nonparametric
Method for Automatic Correctionof Intensity Nonuniformity
in MRI Data’, Medical Imaging, 1998 pp. 1-11
(4) Ashburner J. and Friston K.J., ‘Voxel-based morphometry -
the methods’, NeuroImage, 2000 vol. 11 (6 Pt 1) pp. 805-821
(5) Dahnke R., Ziegler G., Großkreutz J., Gaser C., ‘’, HBM 2012
(6) Koller C.J., ‘A survey of MRI quality assurance programmes’,
British Journal of Radiology, 2006 vol. 79 (943) pp. 592-596
(7) Van Horn J.D. and Toga A.W., ‘Is it time to re-priori-
tize neuroimaging databases and digital reposito-
ries?’, NeuroImage, 2009 vol. 47 (4) pp. 1720-1734
(8) Poline J.B., Breeze J.L., Ghosh S., Gorgolewski K., Halchenko
Y.O., Hanke M., Haselgrove C., Helmer K.G., Keator D.B.,
Marcus D.S., Poldrack R.A, Schwartz Y, Ashburner J.,
and Kennedy D.N., ‘Data sharing in neuroimaging re-
search’, Frontiers in Neuroinformatics, 2012 vol. 6 p. 9
(9) Mortamet B., Bernstein M.A., Jack Jr. C.R., Gunter J.L., Ward C.,
Britson P.J., Meuli R., Thiran J.P., Krueger G., and Alzheimer’s
Disease Neuroimaging Initiative. ‘Automatic quality as-
sessment in structural brain magnetic resonance imag-
ing’, Magn Reson Med, 2009 vol. 62 (2) pp. 365-372
(10) Kruggel F., Turner J., Muftuler L.T., and Alzheimer’s Disease Neu-
roimaging Initiative. ‘Impact of scanner hardware and imaging
protocol on image quality and compartment volume precision
in the ADNI cohort’, NeuroImage, 2010 vol. 49 (3) pp. 2123-2133
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
!"#$%&'()
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
!"#$%&'()
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
!"#$%&'()
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
!"#$%&'()
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
)*$
+*$
#*$
&*$
(*$
!"#$%&'()
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
!*$
2!*$
3!*$
%!*$
'!*$
)!!*$
*(&+,+-'(./0/)
,-.'$
/01$
!"#$
!"%$
!"&$
!"'$
!"($
)"!$
)$44$
+$44$
#$44$
&$44$
($44$
1%&.%&/"#23$#(-)
,-.'$
/01$
1 3 5 7 9 mm0% 1% 3% 5% 7% 9%
0% 20% 40% 60% 80% 100%
1.0
0.9
0.8
0.7
0.6
0.5
1.0
0.9
0.8
0.7
0.6
0.5
1.0
0.9
0.8
0.7
0.6
0.5
B Noise Inhomogeneity Slice resolution
rn0_rf0_st1mm rn0_rf-100C_st1mm rn3_rf+20A_st3mmrn9_rf0_st1mm
A Original Noise Inhomogeneity Slice resolution
kappa
kappa
kappa
noise inhomogeneity slice resolution
VBM8 FSL4
Fig 1: Examples of the BrainWebPhantom (BWP) without image interference, with 9% noise, with 100%
inhomogeneity, and lower slice thickness (A). Low image quality reduced classication accuracy (B). The
loss of accuracy depends on the method and the strength of the respective interference (B). Quality
measures can identify and describe such systematic dierences to exclude outliers.
Fig 1: Examples of the BrainWebPhantom (BWP) without image
interference, with 9% noise, with 100% inhomogeneity, and lower
slice thickness (A). Low image quality reduced classifi cation
accuracy (B). The loss of accuracy depends on the method and
the strength of the respective interference (B). Quality measures
can identify and describe such systematic differences to exclude
outliers.
Fig. 2: The VBM12 school mark rating system (A) is used to describe dierent quality measures (QM) in
a common space. Dierent QMs can be simply averaged with a weigthing w to controll the inuence
of each QM (B). Table 3 (C) gives an overview of the used datasets.
Mark Meaning
1.0 - 1.5 perfect
1.5 - 2.5 good
2.5 - 3.5 average (light interferences)
3.5 - 4.5 poor (noticable interferences)
4.5 - 5.5
critical (low res. or problematic interferences)
>5.5
unacceptable (low res. or severe interferences)
Dataset Images Reference Datasource
ADHD 948 - http://fcon_1000.projects.nitrc.org/indi/adhd200
BWP 145 Collins (1998) http://brainweb.bic.mni.mcgill.ca/brainweb
INDI 1179 - http://fcon_1000.projects.nitrc.org
IXI 555 - http://www.brain-development.org
OASIS 436 Marcus (2007) http://www.oasis-brains.org
QM 1.0 6.0 w
NCR 0.05 0.40 0.5
ICR 0.05 1.00 0.0
RES 0.50 3.00 0.5
NCR (Noise Contrast Ratio)
ICR (Inhomogeneity Contrast Ratio)
RES (resolution)
A Table 1: VBM12 school mark rating system B Table 2: QM scaling
C Table 3: Datasets and their sources
Fig. 2: The VBM12 school mark rating system (A) is used to
describe different quality measures (QM) in a common space.
Different QMs can be simply averaged with a weigthing w to control
the infl uence of each QM (B). Table 3 (C) gives an overview of
the used datasets.
Fig. 3: The BWP data was preprocessed by 5 dierent segmentations (SPM8, SPM12, VBM8, VBM12,
FSL4). The reconstruction quality (described by Kappa) was used to create 3 quality groups (group 1:
Kappa>=0.9 with 31 images, group 2: Kappa 0.9-0.8 with 71 images, group 3: K appa<0.8 with 43
images). (A) shows the image quality rating of dierent QM (left, NCR (Noise Contrast Ratio), ICR
(Inhomogeneity contrast ration, RES (RMS resolution)) and the relative GM volume. To test real data, an
expert classied images of 5 dierent sites into 3 quality levels: no (61 images), slight (42 images), and
strong interferences (31 images). Examples of the no and strong artifact groups are shown in (C).
Lower image quality correlatets with reduced relative GM volume in simulated and real images.
NCR1 NCR2 NCR3 ICR1 ICR2 ICR3 RES1 RES2 RES3 IQM1 IQM2 IQM3PQM1
1
2
3
4
5
6
7
8
QM − expert quality groups
mark
QMs by expert quality group (real data)
GMV1 GMV2 GMV3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
expert quality groups
relative GMV
rel. GMV of expert group
B
NCR1 NCR2 NCR3 ICR1 ICR2 ICR3 RES1 RES2 RES3 IQM1 IQM2 IQM3
1
2
3
4
5
6
7
8
QM − Kappa group
mark
QMs by groups based on the median Kappa of all methods (phantom data)
GMV1 GMV2 GMV3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
Kappa quality groups
relative GMV
rel. GMV of Kappa groups
A
C
group 1 (good)
group 3 (poor)
site 1 site 2 site 3 site 4 site 5
Fig. 3: The BWP data was preprocessed by 5 different
segmentations (SPM8, SPM12, VBM8, VBM12, FSL4). The
reconstruction quality (described by Kappa) was used to create
3 quality groups (group 1: Kappa>=0.9 with 31 images, group
2: Kappa 0.9-0.8 with 71 images, group 3: Kappa<0.8 with 43
images). (A) shows the image quality rating of different QM
(left, NCR (Noise Contrast Ratio), ICR (Inhomogeneity contrast
ration, RES (RMS resolution)) and the relative GM volume. To
test real data, an expert classifi ed images of 5 different sites into
3 quality levels: no (61 images), slight (42 images), and strong
interferences (31 images). Examples of the no and strong artifact
groups are shown in (C). Lower image quality correlates with
reduced relative GM volume in simulated and real images.
ICQM
INDI center
DaKWYM L J PN F OS I GC Z U EQ T HV RX A B
1
2
3
4
ABD
segmentation T1
excellent image quality,
good segmentation
protocols with low image quality,
but acceptable segmentation
20−30 30−40 40−50 50−60 60−70 70−80
1
2
3
4
5
6
age (in years)
IQM
IQM versus age in IXI
20−30 30−40 40−50 50−60 60−70 70−80
0.3
0.4
0.5
0.6
age (in years)
relative GM volume
relative GM volume by age in IXI
male female
1
2
3
4
5
6
sex
IQM
IQM by sex in IXI
male female
0.3
0.4
0.5
0.6
sex
relative GM volume
rel. GM vol. by sex in IXI
HC ADHD HC AD
1
2
3
4
5
6
disease and project
IQM
IQM versus disease
HC ADHD HC AD
0.3
0.4
0.5
0.6
disease and project
relative GM volume
relative GM volume
A
D
IQM for dierent sites of the INDI data sharing project
B C
Fig. 4: For our nal QM IQM (top row) we test for inuences of age (A), sex (B), and health status (C) in
contrast to the relative GM volume (bottom row). Especially data sharing projects like INDI have to deal
with strongly varying MR protocols with very high, but also marginal image quality. (D) show the results
for the renamed sites and visual examples of the worst (site A,B) and best center (site D).
ADHD OASIS
ADHD OASIS
Fig. 4: For our fi nal QM IQM (top row) we test for infl uences of
age (A), sex (B), and health status (C) in contrast to the relative
GM volume (bottom row). Especially data sharing projects like
INDI have to deal with strongly varying MR protocols with very
high, but also marginal image quality. (D) show the results for the
renamed sites and visual examples of the worst (site A,B) and
best center (site D).
ResearchGate has not been able to resolve any citations for this publication.
Article
After conception and implementation of any new medical image processing algorithm, validation is an important step to ensure that the procedure fulfills all requirements set forth at the initial design stage. Although the algorithm must be evaluated on real data, a comprehensive validation requires the additional use of simulated data since it is impossible to establish ground truth with in vivo data. Experiments with simulated data permit controlled evaluation over a wide range of conditions (e.g., different levels of noise, contrast, intensity artefacts, or geometric distortion). Such considerations have become increasingly important with the rapid growth of neuroimaging, i.e., computational analysis of brain structure and function using brain scanning methods such as positron emission tomography and magnetic resonance imaging. Since simple objects such as ellipsoids or parallelepipedes do not reflect the complexity of natural brain anatomy, we present the design and creation of a realistic, high-resolution, digital, volumetric phantom of the human brain. This three-dimensional digital brain phantom is made up of ten volumetric data sets that define the spatial distribution for different tissues (e.g., grey matter, white matter, muscle, skin, etc.), where voxel intensity is proportional to the fraction of tissue within the voxel. The digital brain phantom can be used to simulate tomographic images of the head. Since the contribution of each tissue type to each voxel in the brain phantom is known, it can be used as the gold standard to test analysis algorithms such as classification procedures which seek to identify the tissue "type" of each image voxel. Furthermore, since the same anatomical phantom may be used to drive simulators for different modalities, it is the ideal tool to test intermodality registration algorithms. The brain phantom and simulated MR images have been made publicly available on the Internet (http://www.bic.mni.mcgill.ca/brainweb).
  • M Blaimer
  • F Breuer
  • M Mueller
  • R M Heidemann
  • M A Griswold
  • Jakob P M Smash
  • Sense
  • Pils
  • Grappa
Blaimer M., Breuer F., Mueller M., Heidemann R.M., Griswold M.A., and Jakob P.M., 'SMASH, SENSE, PILS, GRAPPA: how to choose the optimal method', Top Magn Reson Imaging, 2004 vol. 15 (4) pp. 223-236