Page 1

Assessment of ventricular mechanical dyssynchrony by short-axis MRI

M. Mischi, Member, IEEE, H.C.M. v.d. Bosch, A.H.M. Jansen, R.M. Aarts, Fellow, IEEE, and H.H.M. Korsten

Abstract—Nowadays, patients with symptomatic heart fail-

ure and intraventricular conduction delay can be treated with

a cardiac resynchronization therapy. Electrical dyssynchrony is

typically adopted to represent myocardial dyssynchrony, to be

compensated by cardiac resynchronization therapy. One third

of the patients, however, does not respond to the therapy. There-

fore, imaging modalities aimed at the mechanical dyssynchrony

estimation have been recently proposed to improve patient

selection criteria. This paper presents a novel fully-automated

method for regional mechanical left-ventricular dyssynchrony

quantification in short-axis magnetic resonance imaging. The

endocardial movement is described by time-displacement curves

with respect to an automatically-determined reference point.

These curves are analyzed for the estimation of the regional

contraction timings. Four methods are proposed and tested for

the contraction timing estimation. They were evaluated in two

groups of subjects with and without left bundle branch block.

The standard deviation of the contraction timings showed a

significant increase for left bundle branch block patients with all

the methods. However, a novel method based on phase spectrum

analysis shows a better specificity and sensitivity. This method

may therefore provide a valuable prognostic indicator for heart

failure patients with dyssynchronous ventricular contraction,

adding new possibilities for regional timing analysis.

I. INTRODUCTION

Almost two million people in the United States suffers

from conduction system disease, i.e., the electrical depo-

larization signal that induces the muscle contraction is not

properly conducted over the myocardium [1]. This problem,

which is typically characterized by a prolongation of the QRS

complex in the electrocardiographic measurement, leads to

a dyssynchronous contraction of the left ventricle (LV). The

major effects are a reduction of the cardiac efficiency and

ejection fraction, which results in a progressive deterioration

of the LV function.

Cardiac resynchronization therapy (CRT) is an established

therapy for patients with symptomatic heart failure and

prolonged QRS duration [2]. Nevertheless, not all patients

respond to CRT and LV reverse remodeling is appreciated in

slightly more than 50 % of patients [3]. This lack of response

is typically attributed to inappropriate patient selection and,

in particular, to the poor predictive value of QRS prolun-

gation [4], [5]. Several studies on mechanical dyssynchrony

M. Mischi is with the Eindhoven University of Technology, the Nether-

lands.

H.C.M. v.d. Bosch and A.H.M. Jansen are with the Depts of Radiology

and Cardiology of the Catharina Hospital in Eindhoven, the Netherlands.

R.M. Aarts is with Philips Research in Eindhoven and with the Eindhoven

University of Technology, the Netherlands.

H.H.M. Korsten is with the Catharina Hospital in Eindhoven and with

the Eindhoven University of Technology, the Netherlands.

suggest that mechanical dyssynchrony is a better predictor

for CRT response than electrical dyssynchrony [6], [7].

Mechanical dyssynchrony can be assessed using imaging

modalities based on magnetic resonance imaging (MRI) or

ultrasound echo-Doppler. In general, myocardial velocity

timing can be evaluated by ultrasound tissue doppler imaging

(TDI) [6], [7] and myocardial strain timing can be evaluated

by MRI tagging [8]. However, TDI is limited by the acoustic

windows through the ribs and the anisotropic sensitivity

of the method (Doppler frequency shifts only occur for

longitudinal motion), while MRI tagging requires extensive

computation and shows a low sensitivity for reduced wall

thickening (hypokinetic ventricle).

This paper presents a novel method for regional mechani-

cal LV dyssynchrony quantification based on image analysis.

The method is fully automated. Short-axis MRI cines are

analyzed by an algorithm that is developed on the radial

framework of the image segmentation approach proposed

in [9], where the endocardium was detected by high-pass

filtering and thresholding along a beam of rays originating

in the ventricular center, referred to as seed point. The seed

point was manually determined. The threshold was defined

either manually or on the basis of the image histogram.

The proposed segmentation algorithm is based on dynamic

morphology and radial multiscale analysis of the LV basal

view [10], [11]. This results in a substantial advantage with

respect to the method in [9] and several other methods

proposed in the literature, as no user interaction or threshold

definition is required.

MRI is chosen for the high contrast along the endocar-

dium, which results in an increased segmentation robustness.

A short axis basal view permits to assess the synchronicity

of the basal segments, which is reported to be of prognostic

value in CRT [5], [6]. This allows reducing the complexity

of the diagnostic system from three to two dimensions.

In this study, segmentation is only a preprocessing step

for the dyssynchrony analysis. Based on the endocardial

tracking, the regional endocardial time-displacement curves

(TDCs) are determined. Several methods can then be im-

plemented to assess the contraction dyssynchrony along the

endocardium. To this end, we have adapted and implemented

the systolic time method [5], [12] and the inner product

method [13], [14]. In the latter, the entire cardiac cycle is

used to the determine the local contraction timing. We also

propose two alternative methods that are based on the cross

correlation of the entire TDCs.

The developed dyssynchrony analysis was validated on

two groups of patients with and without left bundle branch

Proceedings of the 29th Annual International

Conference of the IEEE EMBS

Cité Internationale, Lyon, France

August 23-26, 2007.

SaD03.5

1-4244-0788-5/07/$20.00 ©2007 IEEE 6011

Page 2

(a) Motion

(b) Tracked contour

Fig. 1.

in the motion image (a) used for detecting and tracking the endocardium (b)

is circumscribed by inner and outer borders on a beam of rays originating

in the ventricular center, i.e., the seed point. For illustrative purpose the

number of displayed rays is limited to 8.

A short-axis basal view of the left ventricle. The region of interest

block (LBBB). All the methods revealed a significant in-

crease of the standard deviation of the contraction timings

as a LBBB was present. However, one of the proposed

cross correlation methods showed a better sensitivity and

specificity for LBBB diagnosis.

II. METHODOLOGY

A. MRI cine acquisition and regularization

The MRI cines for the dyssynchrony analysis were ac-

quired at high temporal resolution (one hundredth of the

cardiac period) on a 1.5 T MRI scanner (Gyroscan In-

tera, Philips Medical Systems) equipped with a five-element

phased array coil. A balanced steady-state free precession

pulse sequence was adopted. Breath hold was used to avoid

motion artifacts. Going from base to apex, the first slice

that did not show the mitral valve structures was chosen for

the dyssynchrony analysis (see Fig. 1). In order to avoid

contour blurring, image regularization was performed by

combination of morphological proper-closing and proper-

opening operators with a nine element kernel [15].

B. Contour detection

The detection of the endocardial contour in the selected

MRI slice is fully automatic. The standard deviation of the

pixel intensities during the cine is analyzed and a binary

image generated by applying the threshold proposed by Otsu

[16]. The resulting binary image facilitates the automatic

detection of a region for the LV endocardium search, as this

is the structure that shows the most of the movement and,

therefore, the largest gray-level variations.

For the detection of the endocardial region, the Hough

Transform (HT) is employed [17]. The HT is a method that

tests the foreground pixels in a binary image against a defined

curve equation, which in our case is a circle. In fact, the

LV motion in a short axis view has a circular symmetry.

The 21-HT implementation is adopted due to its efficient

computation [18]. The results are the seed point coordinates

and the average radius of the endocardial motion circle.

Based on the 21-HT results, a region of interest (ROI) for

the endocardium search is defined as the endocardial range of

motion, i.e., the inner and outer borders of the endocardium

in the binary motion image (see Fig. 1). The border coordi-

nates are defined along the rays originating in the seed point.

Possible outliers, e.g., due to the papillary structures, are

corrected by means of a continuity condition on the border

derivative in polar coordinates, which is bounded within the

values assumed in case the seed point is located on the border

itself (worst condition). Once the borders are defined, the

seed point is relocated in the center of mass of the inner

border to increase the localization accuracy of the following

analysis. The defined ROI is expanded (20 % inwards and

outwards) before the contour tracking.

C. Contour tracking

Once a ROI for the contour detection in the original cine

is defined, the endocardium could be tracked along the rays

originating from the seed point by means of a high-pass filter

and a threshold [9]. However, as the definition of a threshold

is always critical, here we propose the replacement of the

radial filter in [9] by a more elaborated algorithm, which

does not require any threshold.

The multiscale analysis for edge detection proposed by

Canny et al. [11] is modified and adapted to the proposed

application [10]. A multiscale analysis with a Gaussian deriv-

ative mother wavelet is performed on each ray originating

from the seed point with an angular resolution of 5 degrees

and on each frame of the MRI cine. For each ray the

multiscale analysis produces a scalogram W(n,s) where n

is the distance (in pixels) from the seed point and s is the

scale. A linear function of n is used to weigh W(n,s) and

enhance the values at shorter distance from the seed point.

This is performed to reduce the sensitivity of the algorithm

to fatty structures at larger distance around the heart.

The edges (with negative slope) for each scale correspond

to the (local) maxima of W(n,s) along a ray. Without the

need for a threshold, the location ? n along a ray of the

? n = (n : µ(n) > µ(m),∀n ?= m),

s ∈ S of the normalized W(n,s), i.e.,

µ(n) = min

S

dominant edge, i.e., the endocardium, can be defined as

(1)

where µ(n) is defined as the minimal value for all the scales

?

W(n,s)

?

max

N

[W(n,s)]

?−1?

.

(2)

The solution of (1) results in an efficient compromise be-

tween localization accuracy and noise sensitivity.

In order to increase the robustness of the system, a

temporal continuity constraint is integrated in the system by

calculating µ?(n) over the output scales of three subsequent

frames together. The temporal continuity condition results

from the use of high temporal resolution MRI scans.

D. Dyssynchrony analysis

The result of the segmentation algorithm is a function

r(θ,k), which defines the edge distance from the seed

point along a θ degree ray at the frame k. Dyssynchrony

estimations result from the analysis of the TDCs r(θ,k) for

different rays.

6012

Page 3

To improve the accuracy of the timing analysis, the TDC

time resolution is doubled by linear upsampling. The TDCs,

which include both systole (contraction) and diastole (ex-

pansion), may resemble one period of a sinusoidal function.

When no conduction system dysfunction is present, the

ventricle contracts synchronously along its basal segments

and r(θ,k) shows a similar behavior in time for each angle.

Instead, when a conduction system block is present, the

TDCs for different rays are dyssynchronous.

Most segmentation-based methods for dyssynchrony eval-

uation are based on the estimation of the systolic time [12].

This approach might be more sensitive to noise as it focuses

on one moment of the cardiac cycle (end systole). The

analysis of the complete cardiac cycle was also proposed

to compare lateral to septal phase differences [13], [14].

Under the assumption of a sinusoidal TDC, the mutual phase

difference was defined as the arccosine of the inner product

between the lateral and septal TDCs [14]. Another approach

used the inner product with a reference sine and cosine

curve (one cycle) to derive an absolute phase delay [13].

These methods are however based on the assumption of

a sinusoidal TDC and they are therefore sensitive to TDC

shape variations.

For a complete validation of the proposed methods, both

the systolic time and the inner product approaches are

adapted and integrated in our MRI short-axis-view analysis.

The systolic time is defined as the time to the minimum

value of the TDC. The inner product method provides a

relative time delay (fraction of the cardiac period) between

two TDCs corresponding to the angles θi and θj that is

given as arccos(< r(θi,k),r(θj,k) >)/2π, where < .,. >

represents the inner product. This delay is estimated for all

the TDCs with respect to the TDC with the largest SNR.

Here we also propose the use cross correlation methods

for the assessment of the mutual delays between TDCs from

different angles. The same reference defined for the inner

product method is also adopted for the cross correlation

method and the regional contraction timing for each angle

is defined as the time at which the cross correlation with

the reference TDC shows its maximum. Typically 72 rays

are employed to span the entire circle. Figure 2 shows a

regional contraction timing plot for two subjects with and

without a LBBB. The contraction timing in the presence of

a LBBB shows increased variations.

The time resolution of this cross correlation method is

limited by the sampling period of the signals. This problem

can be overcome by phase spectrum analysis. If a Discrete

Fourier Transform of two TDCs r(θi,k) and r(θj,k) is

performed, the difference of the phase spectra results ap-

proximately in a line. This is due to the fact that r(θi,k)

and r(θj,k) are time translations of the same curve. If the

slope of the phase difference line is determined, then the

delay between r(θi,k) and r(θj,k) can be estimated with

no time resolution constraints.

The phase analysis is performed after the cross correlation

re-alignment. As a result, the phase difference is bounded

between −π and π with no phase wrapping issues. Ob-

Fig. 2.

without (dashed) LBBB.

The regional contraction timings for two subjects with (solid) and

viously, the assumption of a pure temporal translation is

not completely fulfilled, and the phase difference line may

show a poor SNR. A least square error method is there-

fore employed for the interpolation of the phase difference

line, whose slope can be estimated by a weighted linear

regression. The weights for the linear regression are derived

by the analysis of the amplitude of the frequency spectrum

[19]. They are determined as the minimum value between

the amplitudes of the spectra of the two TDCs. As a result,

the higher energy components, which typically represent the

myocardial movement, provide a larger contribution to the

regression line determination.

III. RESULTS

Ten subjects (five normals and five with LBBB) were

tested for quantification of the regional contraction timings.

LBBB was diagnosed by analysis of the QRS duration.

The performance of the proposed segmentation method was

validated on four evenly distributed image frames per cardiac

cycle. A total number of 40 images was evaluated by

measuring the percent area difference between manually

and automatically delineated endocardial boundaries. The

area was defined as the surface that was delimited by the

delineated boundary and the absolute area difference (error)

was defined as the area that was covered by only one of the

two delineations. The manual segmentation was taken as the

reference method since no gold standard for segmentation

of cardiac images exist. The results show a correlation

coefficient between the two methods R = 0.99. Bland-

Altman [20] analysis resulted in an average area error equal

to 9.8 % while the corresponding standard deviation was less

than 3.8 %. Therefore, the segmentation algorithm is suitable

for the following cardiac dyssynchrony analysis [9].

The validation of the dyssynchrony analysis was per-

formed by estimating the standard deviation of the regional

contraction timing assessed by the systolic time, inner prod-

uct, cross correlation, and phase difference method for each

subject. In Table I the population averages and standard

deviations for the four indicators are given. All the values

are reported as a fraction of the cardiac cycle. According

to the results, separable ranges can be defined by each

method for either population (with and without LBBB).

6013

Page 4

TABLE I

MEAN AND STANDARD DEVIATION OF THE FOUR INDICATORS.

Method

Systolic time

Inner product

Cross correlation

Phase difference

LBBB

0.1500 ± 0.0365

0.1178 ± 0.0436

0.1422 ± 0.0317

0.1468 ± 0.0191

Normals

0.0316 ± 0.0121

0.0436 ± 0.0275

0.0246 ± 0.0220

0.0268 ± 0.0171

TABLE II

SPECIFICITY AND SENSITIVITY OF THE FOUR INDICATORS.

Method

Systolic time

Inner product

Cross correlation

Phase difference

Specificity

99.71 %

90.68 %

98.89 %

99.96 %

Sensitivity

99.00 %

89.28 %

98.30 %

99.95 %

The evaluation of a diagnostic method is typically based on

sensitivity and specificity analysis. To this end, a threshold

must be defined that is used to decide whether a patient

has a LBBB. Assuming our statistical samples to be well

represented by Gaussian distributions defined by the means

and standard deviations in Table I, inference by Bayes rule

can be used to obtain the posterior class probability for each

class (normals and LBBB) by each of the four methods. As

a result, a threshold can be derived such that the posterior

class probability is equal for the two classes. The thresholds

for the systolic time, inner product, cross correlation, and

phase difference method are 0.0650, 0.0800, 0.0749, and

0.0838, respectively. Based on these thresholds, specificity

and sensitivity can be calculated for each method as given

in Table II, where the phase difference method shows the

best results.

IV. DISCUSSION AND CONCLUSION

A novel method is presented that permits the automatic

evaluation of the intraventricular mechanical dyssynchrony

based on high time-resolution MRI cines of a LV short-

axis basal view. The analysis is fully automatic and does

not require any user interaction.

It is known that the ventricular movement comprises a

twist about its long axis. This twist is however limited to

less than 5 degrees in the basal plane [21], and it is therefore

negligible for the presented analysis. Ventricular translations

due to respiration are also negligible since breath-hold MRI

is performed.

Four algorithms have been implemented for the timing

analysis. They are based respectively on systolic time, inner

product, cross correlation, and phase difference analysis.

Validation on ten patients showed that all methods are

suitable for LBBB detection based on the standard deviation

of the estimated regional contraction timings. The analysis of

the entire cardiac cycle by phase difference analysis seems

however to be the most promising in terms of sensitivity and

specificity, although a more extensive validation is necessary.

The improved results obtained by the phase difference

against the inner product and cross correlation methods were

expected. In fact, the first one assumes the TDC to be fully

represented by a sinusoidal function while the second one

has a limited time resolution. Compared to the systolic time

method, the analysis of the complete cardiac cycle, also

including diastole, represents a valid alternative, which might

show interesting predictive value for the selection of CRT

responders.

REFERENCES

[1] S. Iuliano, S. G. Fisher, P. E. Karasik, et al., “QRS duration and

mortality in patients with congestive heart failure,” American Heart

Journal, vol. 143, pp. 1085–1091, Jan. 2002.

[2] W. Abraham, W. Fisher, A. Smith, et al., “Cardiac resinchronization

in chronic heart failure,” New England Journal of Medicine, vol. 346,

no. 24, pp. 1845–1853, 2002.

[3] C. C. Leclercq and D. Kass, “Retiming the failing heart: Principles

and current clinical status of cardiac resynchronization,” J. Am. Coll.

Cardiol., vol. 39, no. 2, pp. 194–201, Jan. 2002.

[4] A. Kashani and S. Barold, “Significance of QRS complex duration in

patients with heart failure,” J. Am. Coll. Cardiol., vol. 46, no. 12, pp.

2183–2192, 2005.

[5] C.-M. Yu, W.-H. Fung, H. Lin, et al., “Predictors of left ventricular

reverse remodeling after cardiac resynchronization therapy for heart

failure secondary to idiopathic dilated or ischemic cardiomiopathy,”

Am. J. Cardiol., vol. 91, pp. 684–688, 2002.

[6] A. Jansen, F. Bracke, J. van Dantzig, et al., “Optimization of pulsed

wave tissue Doppler to predict left ventricular reverse remodeling after

cardiac resynchronization therapy,” J. Am. Soc. Echocardiogr., vol. 19,

no. 2, pp. 185–191, 2006.

[7] M. Penicka, J. Bartunek, B. D. Bruyne, et al., “Improvement of

left ventricular function after cardiac resynchronization therapy is

predicted by tissue Doppler imaging echocardiography.” Circulation,

vol. 109, no. 8, pp. 978–983, 2004.

[8] A. Lardo, T. Abraham, and D. Kass, “Magnetic resonance imaging

assessment of ventricular dyssynchrony: Current and emerging con-

cepts,” J. Am. Coll. Cardiol., vol. 46, no. 12, pp. 2223–2228, 2005.

[9] M. Mischi, A. A. C. M. Kalker, and H. H. M. Korsten, “Cardiac im-

age segmentation for contrast agent videodensitometry,” IEEE Trans.

Biomed. Eng., vol. 52, no. 2, pp. 277–286, Feb. 2005.

[10] S. K. Setarehdan and J. J. Soraghan, “Automatic cardiac LV boundary

detection and tracking using hybrid fuzzy temporal and fuzzy multi-

scale edge detection.”

[11] J. Canny, “A computational approach to edge detection,” IEEE Trans.

Pattern. Anal. Mach. Intell., vol. 8, no. 6, pp. 679–698, Nov. 1986.

[12] S. Kapetanakis, M. T. Kearney, A. Siva, et al., “Real-time three-

dimensional echocardiography: A novel technique to quantify global

left ventricular mechanical dyssynchrony,” Circulation, vol. 112, no. 7,

pp. 992–1000, 2005.

[13] O. A. Breithardt, C. Stellbrink, A. P. Kramer, et al., “Echocardio-

graphic quantification of left ventricular asynchrony predicts an acute

hemodynamic benefit of cardiac resynchronization therapy.”

[14] O. Grebe, A. M¨ uller, N. Merkle, et al., “Estimation of intra- and

interventricular dyssynchronization with cardiac magnetic resonance

imaging,” in Comp. in Cardiol. 2003, vol. 30, Thessaloniki, Sept. 2003,

pp. 741–743.

[15] R. M. Haralick, S. R. Sternberg, and X. Zhuang, “Image analysis using

mathematical morphology,” IEEE Trans. Pattern. Anal. Mach. Intell.,

vol. 9, no. 4, pp. 532–550, July 1987.

[16] N. Otsu, “Threshold selection method from gray-level histograms,”

IEEE Trans. Syst. Man. Cybern., vol. 9, no. 1, pp. 62–66, Jan. 1979.

[17] R. O. Duda and P. E. Hart, “Use of the Hough transformation to detect

lines and curves in pictures,” Comm. ACM, vol. 15, pp. 11–15, 1972.

[18] H. K. Yuen, J. Princen, J. Illingworth, and J. Kittler, “Comparative

study of Hough Transform methods for circle finding,” Image Vision

Comput., vol. 8, no. 1, pp. 71–77, Feb. 1990.

[19] C. Houtman, D. Stegeman, J. van Dijk, and M. Zwarts, “Changes in

muscle fiber conduction velocity indicate recruitment of distinct motor

unit populations,” J Appl Physiol, vol. 95, pp. 1045–1054, 2003.

[20] J. Bland and D. Altman, “Statistical methods for assessing agreement

between two methods of clinical measurement,” Lancet, vol. 237, no.

8476, pp. 307–310, Feb. 1986.

[21] A. S. Abbasi, L. M. Eber, R. N. Macalpin, and A. A. Kattus,

“New noninvasive method for assessment of left ventricular rotation,”

Circulation, vol. 112, pp. 3149–3156, 2005.

6014