EUROGRAPHICS 2015/ H.-C. Hege and T. Ropinski Dirk Bartz Prize
Guided Analysis of Cardiac 4D PC-MRI Blood Flow Data
Benjamin Köhler1, Uta Preim2, Matthias Grothoff3, Matthias Gutberlet3, Katharina Fischbach4, Bernhard Preim1
1Department of Simulation and Graphics, Otto-von-Guericke University, Magdeburg, Germany
2Department of Diagnostic Radiology, Hospital Olvenstedt, Magdeburg, Germany
3Department of Diagnostics and Interventional Radiology, Heart Center, Leipzig, Germany
4Department of Radiology and Nuclear Medicine, University Hospital, Magdeburg, Germany
(a) Graph cut-assisted vessel segmentation. (b) Semi-automatic vortex extraction. (c) Cardiac function assessment.
Figure 1: Screenshots of the presented software Bloodline for cardiac 4D PC-MRI data evaluation.
Four-dimensional phase-contrast magnetic resonance imaging (4D PC-MRI) allows the non-invasive acquisition
of temporally resolved, three-dimensional blood ﬂow information. Quantitative and qualitative data analysis help
to assess the cardiac function, severity of diseases and ﬁnd indications of different cardiovascular pathologies.
However, various steps are necessary to achieve expressive visualizations and reliable results. This comprises the
correction of special MR-related artifacts, the segmentation of vessels, ﬂow integration with feature extraction and
the robust quantiﬁcation of clinically important measures. A fast and easy-to-use processing pipeline is essential
since the target user group are physicians. We present a system that offers such a guided workﬂow for cardiac 4D
PC-MRI data. The aorta and pulmonary artery can be analyzed within ten minutes including vortex extraction and
robust determination of the stroke volume as well as the percentaged backﬂow. 64 datasets of healthy volunteers
and of patients with variable diseases such as aneurysms, coarctations and insufﬁciencies were processed so far.
Categories and Subject Descriptors (according to ACM CCS): I.4.9 [Computing Methodologies]: Image Processing
and Computer Vision—Applications
Cardiovascular diseases (CVDs) are the most frequent cause
of death in the world. Understanding their origin and evo-
lution may improve diagnosis and the choice of appropriate
treatments. Vortex ﬂow has been determined as an atypical
ﬂow pattern that is caused by, e.g., heart valve defects or
an altered vessel morphology. Quantitative measures such as
stroke volumes facilitate the assessment of the present car-
diac function and the tracking of disease progression by eval-
uating follow-up examinations.
Four-dimensional phase-contrast magnetic resonance
imaging (4D PC-MRI) gained increasing importance in the
last decade. Its 2D equivalent measures the ﬂow in only one
preangulated slice. Unsatisfactory results make a new acqui-
sition necessary, which is stressful for the patient. In con-
trast, 4D PC-MRI datasets contain the full spatio-temporal
blood ﬂow information and thus allow a more ﬂexible anal-
ysis. Recent advances greatly reduced acquisition times to
levels that are feasible for the clinical routine. Although
it has the potential to replace 2D PC-MRI, 4D ﬂow scans
are mainly performed for research purposes at the moment.
This points out the need for standardized and guided tech-
niques to analyze these highly complex data. Software solu-
The Eurographics Association 2015.
Köhler et al. / Cardiac 4D PC-MRI Data Analysis
tions that integrate such methods into easy-to-use workﬂows
are of equal importance. We present a tool that facilitates
data analysis within ten minutes. This includes an automated
data preprocessing, a graph cut-assisted vessel segmentation,
semi-automatic vortex ﬂow extraction and analysis of the
stroke volume as well as percentaged backﬂow. Resulting
visualizations can easily be saved and shared using the pro-
vided one-click solutions for videos of the animated ﬂow and
screenshots of the 3D view or GUI.
2. Previous and Related Work
Markl et al. [MFK∗12] provide an overview about 4D PC-
MRI acquisitions, Calkoen et al. [CRvdG∗14] document its
high ﬂexibility by describing recent applications. Line predi-
cates were used by Born et al. [BPM∗13] to extract ﬂow fea-
tures and employed in a previous work to extract vortex ﬂow
[KGP∗13]. Relevant quantitative measures such as stroke
volumes are described by Hope et al. [HSD13]. We devel-
oped a method to robustly determine stroke volumes and
percentaged backﬂow (regurgitation fractions) [KPG∗14].
MeVisFlow by Hennemuth et al. [HFS∗11] and FourFlow
by Heiberg et al. [HSU∗10] are softwares that provide a
pipeline including preprocessing, segmentation and interac-
tive data exploration. The Siemens Flow Demonstrator is
a similar prototype [SCG∗14]. The Quantitative Flow Ex-
plorer by van Pelt et al. [vPBB∗10] encompasses interac-
tive, illustrative visualizations for data exploration. Ensight
and GyroTools GTFlow are commercial tools that, however,
do not focus on cardiac blood ﬂow.
3. Medical Background
A goal of current medical research papers is to correlate car-
diovascular diseases with speciﬁc ﬂow behaviors, i.e. vor-
tex ﬂow patterns. For instance, Hope et al. [HHM∗10] found
systolic vortex ﬂow in 75 % of their patients with bicuspid
aortic valves – a defect where the aortic valve consists of
only two instead of three leaﬂets. Altered vessel morphology
can be another important factor that promotes the formation
of vortices. Slight dilations are called ectasia, severe dila-
tions are referred to as aneurysm. Pathological narrowings
are called stenosis or, in case of the aortic arch, coarctation.
The stroke volume is the amount of pumped blood per
heart beat in ml and can be determined as ﬂow that passes
a plane, usually located above the valve, orthogonally. It is
calculated as integral of the time-dependent ﬂow rate which
is given in ml/s throughout the cardiac cycle. The cardiac
output is the stroke volume multiplied with the heart rate and
thus describes the heart’s pumping capacity in l/min. These
measures help to assess the cardiac function. Regurgitation
fraction denotes the percentaged amount of blood that ﬂows
back into the ventricle during systole due to improperly clos-
ing valves. It is below 5% in a healthy person. High values
of 20% and more can indicate a valve replacement surgery
if the patient shows severe symptoms.
3.1. Data Acquisition
Most of our datasets were acquired with a 3 T Magnetom
Verio (Siemens Healthcare, Erlangen, Germany). A 4D PC-
MRI dataset consists of each three (x-, y-, z-dimension)
magnitude and ﬂow images that represent the ﬂow strength
and direction, respectively, per voxel. The grid size is 132×
192 in the plane with 15 to 23 and between 14 and 21 tempo-
ral positions. The spatio-temporal resolution is 1.77 mm ×
1.77 mm ×3.5 mm ×50 ms. The velocity encoding (VENC )
– an a-priori MR parameter that describes the maximum ex-
pected velocity – was set to 1.5 m/s, which is a common
choice for aortic blood ﬂow [MFK∗12].
4. Requirement Analysis
The use of (semi-)automatic methods is essential to estab-
lish a ﬂuent workﬂow. Exploitation of the GPU’s computa-
tional power is desirable to speed up the data processing.
Required input should ﬁt into the mental model of our tar-
get user group with a medical background. Thus, employed
algorithms should allow to make use of physicians’ expert
knowledge. Reasonable default parameters have to be pro-
vided for everything that is unintuitive from their perspec-
tive. It is necessary that results of evaluated datasets can eas-
ily be shared via screenshots or videos.
In this section, we describe our developed software named
Bloodline, shown in Figure 1. It is written in C++ and uses
OpenGL for rendering, embedded in a Qt/QML-based GUI.
5.1. Data Import
The raw data – one ﬁle per slice per temporal position – are
converted to 4D images using information from the DICOM
headers. An eddy current correction is then applied to the
ﬂow image using the method by Walker et al. [WCS∗93]
with their suggested default parameters.
5.2. Vessel Segmentation
A temporal maximum intensity projection (tMIP) of the
magnitude images is performed, which yields a high-
contrast 3D image. Graph cuts require the speciﬁcation of
regions in- and outside the target structure as input. The user
provides these information by drawing on the tMIP slices
(see Figure 1a). The better the image quality, the less input
is necessary to achieve satisfactory results. Though, detail
corrections can be performed if the segmentation includes
unwanted or excludes wanted parts. The employed 3D graph
cut with a 26-neighbourhood per voxel allows that the user
The Eurographics Association 2015.
Köhler et al. / Cardiac 4D PC-MRI Data Analysis
does not have to provide input in every slice. Edge weights
in the graph are set to exp(−α· ||∇I||), where Iare the
[0,1]-scaled intensities from the tMIP and αis a tolerance
parameter with 1000 as experimentally determined default
value. Noise in the resulting segmentation is reduced with a
3×3×3 morphological opening and closing.
Phase wraps occur when the measured velocity exceeds
the VENC. In this case, values ﬂip to the other end of the do-
main. We correct phase wraps within the obtained segmen-
tations according to Dìaz et al. [DR04]. The rest of the ﬂow
image is not processed to save time.
5.3. Surface Mesh and Centerline Extraction
Marching cubes is employed to automatically extract the
vessel surface from the segmentation. We apply a low-pass
ﬁlter [TZG96] and reduce the mesh via quadric decimation
[Hop99]. The user marks a start and end point on the vessel
surface for the subsequent centerline extraction [AEIR03].
Multiple end points are allowed to create centerlines in
branching vessels such as the pulmonary artery. An aorta is,
on average, represented by 2500-3000 triangles with a mean
edge length of 4.9 mm. For comparison, a voxel diagonal is
4.3 mm long. This mesh resolution is sufﬁcient because of
the non-complex shape of the aorta and pulmonary artery.
An adapted graph cut enables the semi-automatic 4D
vessel segmentation from time-resolved anatomical images,
e.g., a SSFP cine sequence. An explicit moving surface is au-
tomatically extracted for visualization purposes and optional
quantiﬁcation with increased accuracy [KPG∗15a].
5.4. Qualitative Analysis
Flow Integration: Runge-Kutta-4 with adaptive step size is
implemented on the GPU to integrate the full set of path-
lines. Velocity vectors ~u∈R3in the 4D ﬂow ﬁeld Vat
the spatio-temporal position ~x= (x,y,z,t)Tare obtained via
quadrilinear interpolation. The temporally adjacent vectors
~ubtc=V(x,y,z,btc)and ~udte=V(x,y,z,dte), both obtained
via hardware-accelerated trilinear interpolation, are used to
perform a last linear interpolation manually. We ensure that
each voxel of the segmentation is visited at least once in ev-
ery temporal positions. For this purpose, we seed one path-
line at a random position inside each segmentation voxel at
the ﬁrst temporal position. For each remaining time step, in
succession, we determine the voxels that were not visited,
create new seeds and integrate the pathlines.
Vortex Extraction: During the full ﬂow integration, we
also calculate the λ2vortex criterion for each pathline point.
To alleviate the impact of the low data resolution and noise,
we smooth the values along each pathline using a 1D bi-
nomial ﬁlter with kernel size 3. Contrary to our previous
work [KGP∗13], we do not crop away parts of the pathlines.
Instead, we provide the option to ﬂexibly hide all non-vortex
parts using a slider that adjusts the λ2threshold (see Figure
1b). Filtering ﬂow velocities is possible in the same way. For
the aorta, a circular 2D plot can be generated as overview of
present vortices [KMP∗15].
Visualization: The vessel front is culled and only hinted
at with a ghosted viewing [GNKP10]. The back faces are
rendered with Phong illumination. Pathlines with halos are
created in the geometry shader as view-aligned quads. Il-
luminated streamlines are implemented in the subsequent
fragment shader. In the animation mode, cone-shaped par-
ticles with trails are drawn on every position where the cur-
rent animation time matches a pathline’s temporal compo-
nent. Order-independent transparency ensures correct alpha
blending. The default line width, particle width and particle
length are set according to the dataset’s voxel diagonal. The
standard trail length depends on the temporal resolution and
number of time steps. Real-time adjustment of all visualiza-
tion parameters is possible via sliders.
Media: Results can easily be shared by taking a high-
resolution screenshot of the GUI or render window. The an-
imated ﬂow can be exported to a 1080p video with a sin-
gle click. Patient and dataset information are automatically
added to the top left corner.
5.5. Quantitative Analysis
Measuring planes are automatically oriented orthogonal to
the centerline and their size is automatically determined so
that they ﬁt the vessel (see Figure 1c). The user can drag a
plane along the centerline or adjust the angulation, i.e. ro-
tate it. A diagram shows the time-dependent ﬂow rate deter-
mined for this plane conﬁguration. Additionally, the stroke
volume, cardiac output, regurgitation fraction, mean as well
as peak velocity and the vessel diameter are provided. Un-
fortunately, the calculations are highly sensitive towards the
plane’s angulation. Therefore, a robust stroke volume and
regurgitation fraction analysis can be performed [KPG∗14].
This work received an honorable mention and was invited to
be submitted in an extended version to the Computer Graph-
ics Forum [KPG∗15b]. Another quantiﬁable measure on the
vessel surface is the vectorial wall shear stress.
Bloodline is used by the Heart Center in Leipzig, Germany,
who also use a prototype by Siemens, and the university hos-
pital in Magdeburg, Germany. 64 datasets were evaluated for
research purposes so far in close collaboration with radiol-
ogists specialized on the cardiovascular system. Besides 36
healthy volunteers, the following pathologies were present:
1 aneurysm in the left subclavian artery, 3 aortic insufﬁcien-
cies, 3 ectasias / aneurysms in the ascending aorta, 15 bi-
cuspid aortic valves, some of them with ectatic ascending
aortas, 1 tetralogy of Fallot with pulmonary insufﬁciency, 3
vascular prostheses and 2 coarctations.
The Eurographics Association 2015.
Köhler et al. / Cardiac 4D PC-MRI Data Analysis
After familiarization, physicians are able to perform a
standard evaluation, i.e. vortex ﬂow extraction and stroke
volume as well as regurgitation fraction analysis, in less than
ten minutes, which was rated as feasible for the clinical rou-
tine. The robust quantiﬁcation is most expensive and takes
about 30 s using an Intel i7-3930K and a GeForce GTX 680.
Other costly computations such as the full ﬂow integration
including vortex extraction are each performed within 10 s.
The graph cut-assisted segmentation shows high accep-
tance due to the exploitation of the physicians’ anatomy
knowledge. The enabled hiding of vessels to reduce occlu-
sions was appreciated. The independence of speciﬁc MRI
scanners was emphasized positively. A suggestion was to let
the program perform pending automatic operations such as
pathline integrations for all new datasets at once. This way,
the concentrated waiting time could be used for other things.
7. Conclusion and Future Work
We presented the cardiac 4D PC-MRI data evaluation soft-
ware Bloodline that allows to process datasets within ten
minutes. It integrates a full preprocessing pipeline as well
as quantitative and qualitative data analysis. The use of
(semi-)automatic methods enables a ﬂuent workﬂow. Rea-
sonable defaults strongly reduce the necessity to adjust pa-
rameters. State-of-the-art visualizations can easily be created
and saved in order to share results.
Special functionality for the ventricles shall be provided
in the future. Another goal is to automatically generate clin-
ical reports. Hence, larger studies can be evaluated better and
gender- and age-speciﬁc norm values may be determined.
[AEIR03] ANTIGA L., EN E-I ORDACH E B., REMUZZ I A.: Cen-
terline Computation and Geometric Analysis of Branching Tubu-
lar Surfaces with Application to Blood Vessel Modeling. In Proc
Winter School Comput Graph (WSCG) (2003). 3
[BPM∗13] BORN S., PF EIFL E M., MAR KL M., GU TBE RLET
M., SCHEUERMANN G.: Visual Analysis of Cardiac 4D MRI
Blood Flow Using Line Predicates. IEEE Trans Vis Comput
Graph 19 (2013), 900 – 912. 2
[CRvdG∗14] CALKOE N E. E., ROEST A. A., VAN DER GEE ST
R. J., DE ROOS A., WESTENBERG J. J.: Cardiovascular Func-
tion and Flow by 4-Dimensional Magnetic Resonance Imaging
Techniques: New Applications. J Thorac Imag 29, 3 (2014), 185
– 196. 2
[DR04] DÌAZ C., ROBLES L. A.: Fast Noncontinuous Path
Phase-Unwrapping Algorithm Based on Gradients and Mask. In
Proc Iberoameric Cong Patt Recog (CIARP) (2004), pp. 116 –
[GNKP10] GASTE IGE R R., NEU GEB AUER M ., KUBISCH C.,
PREIM B.: Adapted Surface Visualization of Cerebral
Aneurysms with Embedded Blood Flow Information. In Proc
Eurographics Vis Comput Biol Med (2010), pp. 25 – 32. 3
[HFS∗11] HENNE MUT H A., FRIMAN O., SCHUMANN C.,
BOCK J., DRE XL J. , HUELLEBRAND M., MA RKL M., PEIT-
GEN H.-O.: Fast Interactive Exploration of 4D MRI Flow Data.
Proc of SPIE Med Imag 7964 (2011). 2
[HHM∗10] HOPE M. D., HO PE T. A., MEADOWS A. K., OR-
DOVAS K. G., URBA NIA T. H., AL LEY M . T., HIGGINS C . B.:
Bicuspid Aortic Valve: Four-Dimensional MR Evaluation of As-
cending Aortic Systolic Flow Patterns. Radiology 255, 1 (2010),
53 – 61. 2
[Hop99] HOPPE H .: New Quadric Metric for Simplifying Meshes
with Appearance Attributes. In Proc IEEE Vis (1999), pp. 59 –
[HSD13] HOPE M. D., SEDLIC T., DYVERFELDT P.: Cardiotho-
racic Magnetic Resonance Flow Imaging. J Thorac Imaging 28,
4 (2013), 217 – 230. 2
[HSU∗10] HEIBE RG E., SJÖGREN J., UG ANDER M., CA RLS-
SON M., ENGB LOM H., ARHEDE N H.: Design and Validation
of Segment - Freely Available Software for Cardiovascular Image
Analysis. BMC Med Imag 10 (2010), 1. 2
[KGP∗13] KÖH LER B ., GAST EIGER R., PR EIM U., THEI SE L
H., GUT BER LET M., PREIM B.: Semi-Automatic Vortex Ex-
traction in 4D PC-MRI Cardiac Blood Flow Data Using Line
Predicates. IEEE Trans Vis Comput Graph 19, 12 (2013), 2773 –
[KMP∗15] KÖHLER B., MEUSCHKE M. , PREIM U., FIS-
CHBACH K., GU TBERL ET M., PREIM B.: Two-Dimensional
Plot Visualization of Aortic Vortex Flow in Cardiac 4D PC-MRI
Data. In Proc Bildverarb Med (2015), p. to appear. 3
[KPG∗14] KÖH LER B ., PREI M U., GUT BER LET M., FIS-
CHBACH K., PR EIM B.: Robust Cardiac Function Assessment
in 4D PC-MRI Data. In Proc Eurographics Vis Comput Biol Med
(2014), pp. 1 – 10. 2,3
[KPG∗15a] KÖH LER B ., PREI M U., GROTHOFF M., GU TB ER-
LET M., FISC HBACH K., PRE IM B.: Motion-Aware Stroke Vol-
ume Quantiﬁcation and Vessel Visualization in 4D PC-MRI Data
of the Human Aorta. Int J Comp Assist Radiol Surgery (2015).
[KPG∗15b] KÖH LER B ., PREI M U., GROTHOFF M., GU TBER-
LET M., FISC HBACH K., PRE IM B.: Robust Cardiac Function
Assessment in 4D PC-MRI Data of the Aorta and Pulmonary
Artery. Comp Graph Forum (2015). invited submission. 3
[MFK∗12] MARKL M., FRY DRYCHOW IC Z A., KOZE RKE S.,
HOPE M. D., WIEBEN O.: 4D Flow MRI. J Magn Reson Imag
36, 5 (2012), 1015 – 1036. 2
[SCG∗14] SEMAA N E. M., CARR M., GU LSUN M., VAN OOIJ
P., STAL DER A. F., CA RR J. , COLLI NS J., MA RKL M., SPOT-
TISWOODE B.: Evaluation of an Optimized Post-processing
Tool for 4D Flow MRI Data Analysis in Healthy Volunteers and
Patients with Aortic Stenosis, Aortic Insufﬁciency, and Aortic
Aneurysm. In Proc Intl Soc Mag Reson Med (2014), p. 3948.
[TZG96] TAUBI N G., Z HANG T., GOLU B G.: Optimal Surface
Smoothing as Filter Design. In Proc Europ Conf Comput Vis.
1996, pp. 283 – 292. 3
[vPBB∗10] VAN PELT R., BES COS J. O., BREEU WER M.,
RACH EL E. C., GRÖLL ER M. E., T ER HAA R ROMENIJ B., VI-
LANOVA A.: Exploration of 4D MRI Blood Flow Using Stylistic
Visualization. IEEE Trans Vis Comput Graph 16, 6 (2010), 1339
– 1347. 2
[WCS∗93] WALKE R P. G., CRAN NEY G. B., SCHEIDEGGER
M. B., WASE LES KI G., POHOST G. M., YOGA NATHA N A. P.:
Semiautomated Method for Noise Reduction and Background
Phase Error Correction in MR Phase Velocity Data. J Magn Re-
son Imag 3, 3 (1993), 521 – 530. 2
The Eurographics Association 2015.