Content uploaded by Esther Puyol Anton
Author content
All content in this area was uploaded by Esther Puyol Anton on Jun 21, 2019
Content may be subject to copyright.
FULLY AUTOMATED MYOCARDIAL STRAIN ESTIMATION FROM CINE MRI USING
CONVOLUTIONAL NEURAL NETWORKS
Esther Puyol-Ant´
on1, Bram Ruijsink1,3, Wenjia Bai4,H
´
el`
ene Langet2, Mathieu De Craene2,
Julia A. Schnabel1Paolo Piro2, Andrew P. King1∗, Matthew Sinclair 4∗
1School of Biomedical Engineering & Imaging Sciences , King’s College London, UK
2Philips Research, Medisys, Paris, France
3Guy’s and St Thomas’ Hospital NHS Foundation Trust, London, UK
4Biomedical Image Analysis Group, Department of Computing, Imperial College London, UK
ABSTRACT
Cardiovascular magnetic resonance myocardial feature track-
ing (CMR-FT) is a promising method for quantification of
cardiac function from standard steady-state free precession
(SSFP) images. However, currently available techniques re-
quire operator dependent and time-consuming manual inter-
vention, limiting reproducibility and clinical use. In this pa-
per, we propose a fully automated pipeline to compute left
ventricular (LV) longitudinal and radial strain from 2- and
4-chamber cine acquisitions, and LV circumferential and ra-
dial strain from the short-axis imaging. The method employs
a convolutional neural network to automatically segment the
myocardium, followed by feature tracking and strain estima-
tion. Experiments are performed using 40 healthy volunteers
and 40 ischemic patients from the UK Biobank dataset. Re-
sults show that our method obtained strain values that were
in excellent agreement with the commercially available clini-
cal CMR-FT software CVI42 (Circle Cardiovascular Imaging,
Calgary, Canada).
Index Terms—Myocardial Strain, Automatic pipeline,
Machine learning, MRI
1. INTRODUCTION
Myocardial wall motion analysis (MWMA) allows for pre-
cise and comprehensive assessment of left ventricular (LV)
and right ventricular (RV) contractile function. Myocardial
strain and strain rate provide a relatively load-independent
quantitative evaluation of myocardial wall motion, and have
been shown to enable earlier and more sensitive detection of
myocardial diseases compared to global measures of cardiac
This work is funded by the Kings College London & Imperial Col-
lege London EPSRC Centre for Doctoral Training in Medical Imaging
(EP/L015226/1) and supported by the Wellcome EPSRC Centre for Medical
Engineering at Kings College London (WT 203148/Z/16/Z). This research
has been conducted using the UK Biobank Resource under Application Num-
ber 17806.
∗Joint last authors.
function, such as ventricular volumes and ejection fraction
[1]. In echocardiography, strain can be measured by track-
ing naturally occurring acoustic markers (‘speckles’) in the
myocardium throughout the cardiac cycle. However, the
limited acquisition windows severely restrict the ability to
interrogate total myocardial wall motion. Cardiac magnetic
resonance (CMR) is the current gold standard for assess-
ment of global and regional myocardial function, and does
not suffer from limited acquisition windows. Several CMR
imaging techniques have been proposed for strain analysis,
such as myocardial tagging, phase contrast velocity imaging,
displacement encoding, and strain encoding. Although all of
these CMR techniques provide useful information on myocar-
dial function, they are not typically used in routine clinical
CMR as they require additional imaging and complex, time-
consuming post-processing. Instead, CMR feature tracking
(CMR-FT) has been proposed as a more accessible MWMA
technique. By tracking features between consecutive frames
from steady state free precession (SSFP) cine acquisitions,
in a way analogous to speckle tracking echocardiography,
CMR-FT is able to derive strain and strain rate from rou-
tinely acquired CMR images. However, current CMR-FT
techniques typically require manual delineation of cardiac
volumes and frequent reassessment of annotations based on
tracking results, which is skill and experience dependent.
This results in increased processing times and a significant
degree of inter and intra-observer variability [2].
Related Work: The two most common commercially
available software packages offering CMR-FT are TomTec
(TomTec Imaging Systems, Unterschleissheim, Germany)
and CVI42 (Circle Cardiovascular Imaging, Calgary, Canada).
Both require manual segmentation of the end-diastolic (ED)
frame and identification of RV-LV and mitral valve insertion
points. Furthermore, manual readjustment of the automat-
ically propagated ED segmentations is frequently needed,
resulting in significant processing time and interobserver
variability, limiting its current use for clinical assessment in
978-1-5386-3636-7/18/$31.00 ©2018 IEEE 1139
2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)
April 4-7, 2018, Washington, D.C., USA
large groups of patients. Some semi-automatic methods have
been proposed, for example Mansi et al., [3] presented an im-
provement of the diffeomorphic Demons algorithm for cine
MR sequences. They compared the estimated strain using the
proposed motion tracking algorithm to tagged-MR estimated
strains for a healthy volunteer, and with ultrasound 2-D strain
for a patient with congenital pulmonary valve regurgitations.
However, their method required manual input for segmen-
tation and motion correction. Few automatic pipelines have
been proposed before and most either focus on one type of
strain, or on a single slice. For example, Jolly et al., [4]
proposed an automatic pipeline to measure LV mean mid-
wall Eulerian circumferential strain from cine SSFP. More
recently, Vigneault et al., [5] proposed an automatic pipeline
for estimation of circumferential cardiac strain using deep
learning, although did not make a direct comparison to any
clinical software.
Contributions: In this paper we propose a fully auto-
matic pipeline to quantify myocardial longitudinal, radial and
circumferential strain from cine MR sequences. The pipeline
enables fast and accurate assessment of LV strains and elimi-
nates manual intervention and inter and intra-observer vari-
ation. To validate the proposed method we compared the
obtained strain values with those computed using CVI42,a
widely used clinical tissue-tracking CMR software package.
2. MATERIAL
The study population consisted of 40 healthy volunteers and
40 ischemic patients from the UK Biobank Imaging Study
[6], with demographics displayed in Table 1. CMR imaging
was carried out on a 1.5 Tesla scanner (Siemens Healthcare,
Erlangen, Germany). Short-axis (SA) stacks covering the full
heart, and two orthogonal long-axis (LA) planes (2-chamber
(2Ch) and 4-chamber (4Ch) views) were available for each
subject (TR/TE = 2.6/1.10 ms, flip angle = 80◦). In-plane
resolution of the SA stack and LA images was 1.8mm, with
slice thickness of 8mm and 6mm for SA and LA respectively.
50 frames were acquired per cardiac cycle.
Table 1: Study demographics: end-diastolic volume (EDV);
end-systolic volume (ESV); ejection fraction (EF), all ex-
pressed as mean (standard deviation); and age expressed as
mean (min-max).
Demographics Healthy volunteers Ischemic patients
Study population, n40 40
Age (years) 60.20 (43-73) 66.75 (51-73)
LV-EDV (mL/m2) 141.43 (33.74) 175.47 (48.82)
LV-ESV (mL/m2) 55.61 (15.67) 84.28 (35.72)
LV-EF (%) 60.85 (4.78) 55.08 (7.93)
3. METHODS
The proposed framework for automatically quantifying my-
ocardial strain from cine MR sequences is summarized in Fig.
1, and each step is described below.
Fig. 1: Overview of the proposed framework for automatic
quantification of myocardial strain from cine MR sequence.
Automatic Segmentation Network: A fully-convolutional
network (FCN) with a 17 convolutional layer VGG-like ar-
chitecture was used for the automatic segmentation of the LV
myocardium and blood-pool at ED for SA and LA slices[7, 8].
Each convolutional layer of the network is followed by batch
normalisation and ReLU, except the last one, which is fol-
lowed by the softmax function. In the case of the SA stack,
each slice is segmented independently, i.e. in 2D. From the
segmentations, a bounding box was generated and used to
crop the image to only include the desired FoV, improving
pipeline speed and reducing errors in motion tracking.
Motion correction: Automatic SA and LA segmenta-
tions were used to correct breath-hold induced motion arte-
facts using the iterative registration algorithm proposed in
[7]. The motion-corrected LA/SA slices are used to correctly
identify mid-cavity SA planes for computing strain, deter-
mined by correspondence to the valves and apex identified in
the LA view.
Motion tracking: Motion tracking was performed on
each 2D plane in both SA and LA views using MIRTK; more
specifically, a 2D B-spline free-form deformation (FFD)
registration was used [9] to estimate LV motion between con-
secutive frames of the cine MR sequences.
Generation of layers and segments of the myocardium.
On the ED frame, the LV myocardium was divided into 5 lay-
ers and 6 segments as illustrated in Fig. 2, and described
below. From the SA and LA segmentations, the contours
defining the boundaries of the LV endocardium and epi-
cardium were extracted using standard morphological opera-
tions. Both contours were smoothed by fitting a spline with
the same number of equally-spaced points for both. Skele-
tonization was used to generate a centreline of the myocardial
segmentation in both the SA and LA views. Centrelines were
smoothed by fitting a further spline. Two additional contours
were generated at the midline of the centreline-epicardium
and centreline-endocardium, resulting in a total of 5 con-
centric trans-mural contours. In addition, six myocardial
1140
sectors were identified in each slice. In the LA views, the
myocardium was divided into 6 equally sized sectors (see
Fig. 2). In the SA view, RV-LV intersections were auto-
matically detected (RV1 and RV2 in Fig. 2) and used to
divide the septum and LV free-wall into respectively 2 and 4
equally sized sectors along the arc-length of the myocardial
centreline. Finally, all spline points were transformed with
the motion tracking deformation fields.
Strain computation: Myocardial strain defines the total
deformation of a region of tissue during the cardiac cycle rela-
tive to its initial configuration at the onset of the cardiac cycle,
and it is normally expressed in percentages. Three compo-
nents of myocardial strain (radial, rr, circumferential, cc, and
longitudinal, ll) are typically measured, and each component
used to quantify different aspects of cardiac function. More
specifically, the mean Lagrangian strain over the whole my-
ocardium for each strain component j(i.e. rr,cc or ll) at each
time point twas computed as follows:
Et
j,v =
S
s=1
K
k=1
1
SK
dt
j,v,k,s −dED
j,v,k,s
dED
j,v,k,s
(1)
where dED
j,v,k,s is the length at ED for the segment s, layer k
and view v(i.e. SA or LA) for strain component j. Radial
strain was computed with transmural distance (drin Fig. 2)
from SA/LA slices; circumferential strain using circumfer-
ential segment arclength (dcin Fig. 2) from SA slices; and
longitudinal strain using longitudinal segment arclength from
LA slices. Because there are five layers, we can calculate
the endocardial, epicardial, midwall, endo-midwall and epi-
midwall strain separately. Global strain was computed as the
average of the estimated strains from each segment and layer
to reduce noise.
Fig. 2:Left: Schema of the SA segments with radial distance
drand circumferential distance dc.Middle: SA segmentation
with 5 concentric contours within the LV wall and six sectors.
Right: LA segmentation split into 5 layers and 6 segments.
Colours represent segments.
4. RESULTS
The strain values obtained using the proposed automatic
method were compared with strain analysis obtained by an
expert CMR trained cardiologist using CVI42. Importantly,
the manual segmentations created in CVI42 were made inde-
pendently of the proposed automated framework and strain
results were computed as the mean of three analysis repeti-
tions, according to clinical consensus [2]. Furthermore, the
deep learning segmentation network was trained and opti-
mised using a separate cohort of the UKBB data set. The 2Ch
and 4Ch LA slices were used to determine LV longitudinal
strain and LV radial strain (Et
ll,LA and Et
rr,LA) alongside
the time to peak (TPK) strain duration. LV SA circumfer-
ential (Et
cc,SA ) and radial (Et
rr,SA) strains and the corre-
sponding TPK strain durations were calculated from three
mid-ventricular SA slices determined automatically in the
proposed approach. The same slices were also used in CVI42 .
Peak LV strain values obtained using the proposed method
and CVI42 were compared using a Welch’s t−test (significant
differences reported for p-value<0.05 with Bonferroni’s cor-
rection). Moreover, peak LV strain values between healthy
volunteers and ischemic patients were compared using a
t−test (significant differences reported for p-value<0.05
with Bonferroni’s correction) for the proposed method and
CVI42. Fig. 3 shows an example of the three strains estimated
for a healthy volunteer and an ischemic patient. Experiments
were carried out on a PC with a Intel Xeon CPU E5-1660 v3
with 31GiB RAM, and the run-time of the proposed pipeline
was 200s per subject.
Quantification of LV strain. Table 2 shows the average peak
strain values of 40 healthy volunteers and 40 ischemic pa-
tients using both methods. Paired sample t−tests, showed no
statistically significant difference between the two methods
for peak Err,LA and Err,SA in both patients and volunteers,
but showed a slight underestimation in Ell,LA and Ecc,SA
in healthy volunteers with the proposed method compared to
CVI42. Furthermore, an unpaired sample t−test showed that
the proposed method was successful in detecting a statisti-
cally significant decreased peak strain in ischemic patients
compared to healthy volunteers, similarly to CVI42.
Fig. 3: Examples of estimated strains for a healthy volunteer
and an ischemic patient. Figure shows global strain curves,
peak strain and TPK.
1141
Table 2: Comparison of strain results derived from CVI42 and
the proposed pipeline for the cohort, reported as mean (SD).
Asterisks indicate significant differences between proposed
methid and CVI42. Daggers indicate statistically significant
differences between healthy volunteers and ischemic patients.
Healthy volunteers
CVI42 Proposed
Peak Ell,LA(%) -20.26 (2.44) -18.17 (2.49)*
TPK Ell,LA(ms) 343.75 (44.41) 383.68 (53.83)
Peak Err,LA(%) 37.11 (7.52) 36.70 (7.97)
TPK Err,LA(ms) 343.75 (44.41) 387.33 (52.24)
Peak Err,SA(%) 43.07 (7.10) 42.31 (8.48)
TPK Err,SA(ms) 336.25 (36.04) 337.93 (33.97)
Peak Ecc,SA (%) -22.02 (2.11) -19.94 (2.56)*
TPK Ecc,SA (ms) 330.60 (40.63) 331.58 (46.92)
Ischemic patients
CVI42 Proposed
Peak Ell,LA(%) -17.19 (4.26)†-16.09 (4.29)†
TPK Ell,LA(ms) 367.50 (66.02) 366.90 (62.53)
Peak Err,LA(%) 29.67 (10.36)†31.54 (10.33)†
TPK Err,LA(ms) 339.00 (58.74) 369.64 (64.72)
Peak Err,SA(%) 32.32 (8.99)†31.70 (9.81)†
TPK Err,SA(ms) 355.27 (34.28) 349.73 (35.26)
Peak Ecc,SA (%) -18.22 (3.40)†-18.02 (3.78)†
TPK Ecc,SA (ms) 362.03 (45.56) 362.03 (61.84)
Variability between CVI42 and proposed method. The
variation in peak strain estimation between the two meth-
ods was assessed using Bland-Altman analysis and Intraclass
correlation coefficients (ICC) (see Fig. 4). The level of
agreement was defined as in [2]: excellent for ICC>0.74,
good for ICC=0.6-0.74, fair for ICC=0.40-0.59, and poor for
ICC<0.4. Results show excellent agreement between our
method and CVI42, with the lowest variation in peak Ell,LA
and Ecc,SA , whereas a larger spread was observed for peak
radial strain. The latter is in line with the larger intra and
inter observer variability seen in radial strain assessment in
previously published CMR-FT literature [1].
5. DISCUSSION AND CONCLUSIONS
Automatic quantification of cardiac function from cine CMR
sequences has the potential to increase accessibility of MWMA
for assessment of cardiac diseases by eliminating time con-
suming manual post-processing steps and reducing inter- and
intra-observer variation. In this paper, we have presented
a fully automated pipeline for LV strain estimation that in-
cludes segmentation, motion tracking and longitudinal, radial
and circumferential strain estimation from routinely acquired
CMR imaging. This is the first time that such a pipeline has
been described. We compared the performance of the pro-
posed method with CVI42, one of the two commercially avail-
Fig. 4: Bland Altman plots between CVI42 and the proposed
pipeline for peak global strain (%). Healthy volunteers are
shown as black circles and ischemic patients as blue trian-
gles. Dotted lines correspond to the mean difference between
methods, and dashed lines correspond to the limits of agree-
ment (95% confidence).
able software packages for CMR-FT. Our method showed ex-
cellent agreement with strain analysis manually obtained by
an expert in CVI42. The method slightly underestimated peak
longitudinal and circumferential strain compared to CVI42,
most likely reflecting minor differences in motion tracking
algorithms. However, this underestimation was consistent
throughout the range of strains observed in our study. We
also note that, although widely used clinically, tools such
as CVI42 cannot be considered as gold standards for strain
quantification due to their intra and inter observer variability.
Future work will focus on extending the strain computation
to a larger cohort and to multiple pathologies. Furthermore,
our method allows for assessment of regional 2D strain and
can be easily extended to 3D strain, which can be used in the
future for more comprehensive regional MWMA.
6. REFERENCES
[1] P Claus et al., “Tissue tracking technology for assessing
cardiac mechanics: principles, normal values, and clini-
cal applications,” JACC: Cardiovascular Imaging, vol. 8,
no. 12, pp. 1444–1460, 2015.
[2] A Schuster et al., “Cardiovascular magnetic resonance
feature-tracking assessment of myocardial mechanics:
Intervendor agreement and considerations regarding re-
producibility,” Clinical radiology, vol. 70, no. 9, pp. 989–
998, 2015.
[3] T Mansi et al., “Physically-constrained diffeomorphic
demons for the estimation of 3d myocardium strain from
1142
cine-mri,” Functional Imag. Modeling Heart., pp. 201–
210, 2009.
[4] M-P Jolly et al., “Automated assessments of circum-
ferential strain from cine CMR correlate with LVEF de-
clines in cancer patients early after receipt of cardio-
toxic chemotherapy,” Journal of Cardiovascular Mag-
netic Resonance, vol. 19, no. 1, pp. 59, 2017.
[5] DM Vigneault et al., “Feature tracking cardiac mag-
netic resonance via deep learning and spline optimiza-
tion,” Functional Imag. Modeling Heart., p. 183, 2017.
[6] SE Petersen et al., “UK Biobanks cardiovascular mag-
netic resonance protocol,” Journal of cardiovascular
magnetic resonance, vol. 18, no. 1, pp. 8, 2016.
[7] M Sinclair et al., “Fully automated segmentation-based
respiratory motion correction of multiplanar cardiac mag-
netic resonance images for large-scale datasets,” in
Med. Image Comput. Comput. Assist. Interv (MICCAI).
Springer, 2017, pp. 332–340.
[8] W Bai et al., “Semi-supervised learning for network-
based cardiac MR image segmentation,” in Med. Im-
age Comput. Comput. Assist. Interv (MICCAI), 2017, pp.
253–260.
[9] D Rueckert et al., “Nonrigid registration using free-form
deformations: application to breast mr images,” IEEE
Trans. Med. Imag., vol. 18, no. 8, pp. 712–721, 1999.
1143