FAST AND ROBUST LOCALIZATION OF

THE HEART IN CARDIAC MRI SERIES

A Cascade of Operations for Automatically Detecting the Heart in Cine MRI Series

Sebastian Zambal, Andreas Sch

ollhuber, Katja B

uhler and Ji

ı Hlad

uvka

VRVis Research Center, Donau-City-Strasse 1, Vienna, Austria

Keywords:

Localization of the Heart, Magnetic Resonance Imaging, Cardiac cine MRI.

Abstract:

This work presents a robust approach for fast initialization of an Active Appearance Model for subsequent

segmentation of cardiac MRI data. The method automatically determines AAM initialization parameters:

position, orientation, and scaling of the model. Four steps are carried out: (1) variance images over time

are calculated to ﬁnd a bounding box that roughly deﬁnes the heart region; (2) circle Hough-transformation

adapted to gray values is performed to detect the left ventricle; (3) thresholding is carried out to determine the

orientation of the heart; (4) the optimal initialization is selected using a mean texture model.

The method was evaluated on 42 MRI short axis studies coming from two MRI scanners of two different

vendors. Automatic initializations are compared to manual ones. It is shown that the proposed automatic

method is much faster than and achieves results qualitatively equal to manual initialization.

1 INTRODUCTION

Segmentation of cardiac structures from magnetic

resonance (MRI) images has been of great interest

in the medical imaging community (S

orgel and Vaer-

man, 1997; Stegmann and Pedersen, 2005; Lelieveldt

et al., 2001; van Assen et al., 2006). The great advan-

tage of model-based segmentation is that it incorpo-

rates prior knowledge about the segmented structures.

Active Appearance Models (AAMs) (Cootes

et al., 1998) are deformable models which describe

possible conﬁgurations of shape and gray values by

statistical analysis of a training data set. Several au-

thors have proposed the use of AAMs and their nu-

merous extensions to the problem of segmentation of

cardiac structures. Methods proposed so far com-

prise 3D AAMs (Mitchell et al., 2002), temporal

AAMs (Lelieveldt et al., 2001), and 3D+time AAMs

(Stegmann and Pedersen, 2005).

A concrete instance of an AAM is deﬁned by pa-

rameters comprising: position, scaling, orientation,

shape and texture parameters. Matching the model

to unseen data is equivalent to ﬁnding a conﬁgura-

tion of parameters that optimally ﬁt the model to the

the unseen data. The common proceeding is to place

the model onto unknown image data. Then deforma-

tions are iteratively applied until a difference mea-

sure such as root mean square (RMS) texture differ-

ence reaches a minimum. A problem often ignored

in literature is robust and fast automatic initialization

of the model, i.e. ﬁnding reasonable initial position,

orientation and scaling. A brute-force method itera-

tively tries out each and every conﬁguration. How-

ever this is very time-consuming since the number of

possible initializations is huge. In previous work it

has been suggested to perform AAM Search in par-

allel with multiple different initialization parameters

(Stegmann, 2000). However, this approach is quite

time consuming, especially when dealing with 3D

AAMs. To utilize AAM-based segmentation for car-

diac cine MR in daily clinical practice, a more efﬁ-

cient method for initialization is required.

Recently a method based on sparse Markov Ran-

dom Fields (MRFs) (Donner et al., 2007) has been

proposed for fast initialization of model-based seg-

mentation. However this method relies on feature ex-

traction which is not proven to deliver adequate re-

sults on cardiac cine MR data. Further more the run

times reported for solving the considered MRF are

in the order of a few seconds while the method we

341

Zambal S., Schöllhuber A., Bühler K. and Hladuvka J. (2008).

FAST AND ROBUST LOCALIZATION OF THE HEART IN CARDIAC MRI SERIES - A Cascade of Operations for Automatically Detecting the Heart in

Cine MRI Series.

In Proceedings of the Third International Conference on Computer Vision Theory and Applications, pages 341-346

DOI: 10.5220/0001074503410346

 SciTePress

present in this paper delivers the result after about one

second.

This work presents a method that automatically

determines the initial position parameters for an AAM

for segmentation of the human heart in MRI short axis

data. In earlier work S

orgel and Vaerman (S

orgel

and Vaerman, 1997) have introduced a method for

automatic heart localization for initialization of ac-

tive contours. The presented work extends this ap-

proach. Instead of initializing active contours the goal

of this paper is to initialize an Active Appearance

Model. A set of well-established image processing

algorithms is used including morphological operators

(Soille, 2002) and Hough-transformation for circles

(Davies, 1988). In contrast to previous work where

Fuzzy Hough-transformation (Philip et al., 1994b)

was applied to detect the left ventricle (Philip et al.,

1994a) we propose to perform Hough-transformation

directly on gray values as will be outlined later in this

paper.

This paper is organized as follows: In section 2 an

overview of the investigated MRI data is given. The

fully automatic method for localization of the heart is

described in detail in section 3. Validation and results

are presented in section 4 and the paper concludes

with section 5.

2 DATA

The 4D data considered in this work consists of cine

MRI short axis studies of 42 different patients. The

data was captured using two MR scanners from dif-

ferent vendors each one operating at a magnetic ﬁeld

strength of 1.5 Tesla. Each short-axis study con-

sists of 7 to 13 slices with pixel resolutions ranging

from 1.17mm to 1.68mm. The spacing between slices

ranges from 7.2mm to 12.0mm. Time-resolution lies

in the range of 11 to 27 time steps per patient study.

3 METHOD

Our method takes into account the complete four-

dimensional (3D + time) input data and computes the

initial parameters for the position of the model: po-

sition, orientation and scale-factor. The pipeline is

made up of four steps (see ﬁgure 1):

• extraction of the region of interest (ROI)

• localization of the LV

• calculation of LV-RV orientation

• model-based candidate selection

In each step elementary digital image processing al-

gorithms are used. This makes the method transpar-

ent, comprehensible, and easy to implement. In the

following the individual steps of the algorithm are ex-

plained in detail.

3.1 Extraction of the Region of Interest

In order to limit more complex calculations to a re-

stricted ROI, the ﬁrst step is a detection of the im-

age area which contains the heart. Over the period of

the cardiac cycle position and size of the heart vary

due to contraction of the myocardium. As a result the

strongest variations of gray values appear in the re-

gion inside the heart. The localization is derived from

the variations of gray values over time similarly as

proposed by S

orgel and Vaerman (S

orgel and Vaer-

man, 1997).

For every slice a variance image is calculated. The

gray values of a variance image correspond to the

variance of the according pixel over the complete car-

diac cycle. High gray values indicate strong variance

and thus the according pixels belong to the heart re-

gion with high probability. Figure 2 shows examples

of such variance images for different slices of a single

data set.

It has to be considered, especially for MRI data,

that high variance of a pixel’s gray value might pos-

sibly come from noise or imaging artifacts. To re-

duce the disturbing inﬂuence of noise the following

image processing steps are carried out for the indi-

vidual variance images. A threshold is selected such

that the according number of pixels above the thresh-

old approximately cover the area of the heart (roughly

10000mm

To eliminate single pixels and small pixel areas

a morphological cleaning is applied to each variance

image. The morphological structuring element that

is used is a 5 × 5 mask centered over the considered

pixel. If less than 11 pixels in this mask are set the

pixel is unset. If more than 15 pixels are set the center

pixel is set. Otherwise the old pixel value is kept.

This improves the results signiﬁcantly but in some

cases there still remain misleading pixels set. To in-

crease robustness all masks from all slices are consid-

ered jointly. A new mask is generated by summation

of the individual variance images. This gives a result

as depicted in ﬁgure 3(a). Outliers where misleading

variances appear in individual slices only are removed

with the following operation: All pixels which are set

in less than 25% of all slices are deleted. Figure 3(b)

shows an example of the outcome of this step. To fur-

ther reduce artifacts only the largest connected region

in the mask is considered (ﬁgure 3(c)). A bounding

VISAPP 2008 - International Conference on Computer Vision Theory and Applications

342

Figure 1: The pipeline: (a) ROI extraction, (b) LV localization, (c) LV-RV orientation, (d) model-based candidate selection.

Figure 2: Variance images for individual slices of a single data set.

box around it deﬁnes the ROI as shown in ﬁgure 3(d).

Since information from different time steps and dif-

ferent slices is combined it is argued that the resulting

mask robustly identiﬁes the region of the heart.

3.2 Localization of the Left Ventricle

While the ﬁrst step in the algorithm takes the full

4D data into account the rest of the algorithm is per-

formed on the central slice of end-diastole only. The

end-diastole is typically known since the individual

time steps are delivered as volumes sorted by time,

starting with the end-diastolic volume. Since the cap-

tured volume typically covers the left ventricle from

apex to base the center slice of a volume is taken for

further reﬁned localization of the heart.

The myocardium of LV has approximately the

shape of a circle. This fact motivates the use of

a Hough-transformation for circles (Davies, 1988).

To reduce the computational burden the Hough-

transformation is restricted to the ROI calculated in

the previous step.

Typically the ﬁrst step in Hough-transformation is

edge detection. In the experiments carried out on car-

diac MRI data it turned out that standard edge detec-

tion algorithms like Canny Edge Detection (Canny,

1983) give very poor results for many data sets. On

the one hand this is due to properties of MRI data.

On the other hand fuzzy anatomical structures such as

papillary muscles and trabeculae make it very difﬁcult

FAST AND ROBUST LOCALIZATION OF THE HEART IN CARDIAC MRI SERIES - A Cascade of Operations for

Automatically Detecting the Heart in Cine MRI Series

343

(a) (b) (c) (d)

Figure 3: Towards the ROI: (a) Sum of variations, (b) thresholded, (c) largest region extracted, (d) and resulting bounding

box.

to calculate meaningful image gradients that clearly

represent transitions between objects. Furthermore in

experiments it was observed that gray value distribu-

tions for MRI images ﬂuctuate signiﬁcantly inter and

even intra patient studies. This makes it very hard to

select generic parameters for an elaborate edge detec-

tion algorithm.

More elaborate approaches like Fuzzy Hough-

transformation (Philip et al., 1994b) try to circum-

vent the problem of structures deviating from perfect

circles. Anyway the problem of strongly misleading

edges in the region of papillary muscles remains.

In order to overcome the problems of edge de-

tection, the Hough-transformation is adopted to take

original gray-values rather than edge information as

input. The assumption is made that gray values of the

myocardium are signiﬁcantly darker than those of the

blood inside the ventricles. Thus, for transforming

the image into Hough-space low gray values in the

image are assumed to belong to the myocardium with

high probability. Using this approach the detected cir-

cle does not lie on the boundaries of the myocardium

but somewhere in between. As a result the circle is

detected robustly even if the shape of the left ventri-

cle deviates from the perfect circle. Compared to the

Fuzzy Hough-transformation the computational com-

plexity is even reduced (no gradient calculation is re-

quired). Note that the even darker gray values in the

lung region were excluded since they lie outside the

previously calculated ROI.

The Hough-space considered is a three-

dimensional space of parameters. Its axes are x,

y (position of the circle’s center) and r (radius of

the circle). x and y are constrained by the bounding

box deﬁning the ROI. r is restricted to an interval of

25mm to 40mm – a typical range of radii for LVs.

Although Hough-transformation as we have de-

scribed it works quite robustly there is still a small

chance that the highest evidence for a circle is not

correctly describing the LV. Our experiments showed

that the correct contour of the LV always corresponds

to one of the ﬁrst two largest peaks in Hough-space.

Thus the ﬁrst two most prominent candidates for po-

sition and scaling of the LV are considered at the last

stage of the algorithm.

3.3 Heart Orientation

From previous steps two possible candidates for the

left ventricle are extracted. Each of these candidates

described by position and size of the two Hough cir-

cles. What remains is to determine the orientation of

the heart, i.e. where the right ventricle is located rela-

tive to the left ventricle.

The ROI computed in section 3.1 is thresholded

such that the 20% brightest pixels remain. As exper-

iments showed, the two largest connected regions ro-

bustly correspond to the blood inside LV and RV. The

centroids of these regions already indicate the spa-

tial relation between the ventricles. It is however not

known which region corresponds to the left ventricle

and which to the right one. To resolve this ambiguity

the region with its centroid closer to the center of a

Hough candidate is identiﬁed as the left ventricle.

This way a unique orientation is assigned to both

Hough-candidates.

3.4 Model-based Candidate Selection

The preceding steps reduced the initialization search

space from millions (possibly every pixel with mul-

tiple different orientations) to two candidates: two

Hough circles for the myocardium of the LV together

with estimates of LV-RV orientation. Each of the two

candidates deﬁnes position, scaling, and orientation

for a possible initialization of the model.

To select the optimal candidate the root mean

squared (RMS) texture errors between model and

both initialization candidates are calculated. The can-

didate which produces the smaller error is identiﬁed

as the ﬁnal result.

VISAPP 2008 - International Conference on Computer Vision Theory and Applications

344

4 VALIDATION AND RESULTS

The method has been evaluated for a total of 42 MRI

studies. Automatic initializations have been com-

pared to manual ones: ﬁve users interactively initial-

ized the model. To assess the quality of the automatic

method three ﬁgures of merit have been evaluated:

average point-to-surface distance, texture error, and

time performance.

The average point-to-surface (PTS) distance is

calculated for the manually and automatically placed

mean model shape relative to the accurately done

manual ground truth segmentation. This was only

possible for a subset of 31 data sets where ground

truth segmentation was available. Figure 4 (top) sum-

marizes PTS measures achieved by users and the au-

tomatic method. It is observed that automatic initial-

izations come close to the manual ones. Please note

that the average discrepancy of 6mm only refers to

rigid initialization of the mean model. In this work

only initialization is investigated – no subsequent de-

formations of the model are applied in attempt to

achieve ﬁnal segmentations.

In order to evaluate initializations for which no

ground truth was given, the texture difference be-

tween mean model and image data was determined.

Figure 4 (middle) shows the quality of matches for

all (unsegmented) 42 data sets after manual/automatic

initializations with respect to texture difference. It is

again concluded that the automatic method generates

initializations qualitatively comparable to those of the

users. The ﬁnal results – the initializations – for all

validation data sets are visualized in ﬁgure 5 for the

central slices of end-diastole.

While similar in quality ﬁgure 4 (bottom) proves

another advantage of the automatization over user in-

teraction – the speed-up. The average initialization of

1 second has been achieved by a Java implementation.

5 CONCLUSIONS

This work has introduced an automatic and robust

method for localization of LV and RV in 4D cardiac

MRI data. The method has been designed with help

of few elementary image processing operators. The

Hough-transformation for circles was adapted to op-

erate on original image gray values instead of gradi-

ent magnitudes which makes the detection of the LV

highly robust. The overall quality of initialization has

been assessed by a user study. Time performance of

the method indicates a high potential for daily clinical

use.

REFERENCES

Canny, J. F. (1983). A variational approach to edge detec-

tion. In Proceedings of the Third National Conference

on Artiﬁcial Intelligence (AAAI), pages 54–58.

Cootes, T. F., Edwards, G. J., and Taylor, C. J. (1998). Ac-

tive appearance models. In ECCV, volume 2, pages

484–498.

Davies, E. (1988). A modiﬁed Hough scheme for general

circle location. Pattern Recogn., 7(1):37–43.

Donner, R., Micu

ık, B., Langs, G., and Bischof, H. (2007).

Sparse MRF appearance models for fast anatomical

structure localisation. In British Machine Vision Con-

ference.

Lelieveldt, B. P. F., Mitchell, S. C., Bosch, J. G., van der

Geest, R. J., Sonka, M., and Reiber, J. H. C. (2001).

Time-continuous segmentation of cardiac image se-

quences using active appearance motion models. In

Information Processing in Medical Imaging (IPMI),

pages 446–452.

Mitchell, S. C., Bosch, J. G., Lelieveldt, P. F., van der Geest,

R. J., Reiber, J. H. C., and Sonka, M. (2002). 3D ac-

tive appearance models: Segmentation of cardiac MR

and ultrasound images. IEEE Transactions on Medi-

cal Imaging, 21(9):1167–1178.

Philip, K., Dove, E., McPherson, D., Gotteiner, N., Stan-

ford, W., and Chandran, K. (1994a). Automatic detec-

tion of myocardial contours in cine-computed tomo-

graphic images. IEEE Transactions on Medical Imag-

ing, 13:241–253.

Philip, K., Dove, E., McPherson, D., Gotteiner, N., Stan-

ford, W., and Chandran, K. (1994b). The fuzzy hough

transform-feature extraction in medical images. IEEE

Transactions on Medical Imaging, 13:235–240.

Soille, P. (2002). Morphological Image Analysis. Springer

Verlag Berlin.

orgel, W. and Vaerman, V. (1997). Automatic heart local-

ization from a 4D MRI dataset. In Proceedings SPIE

Medical Imaging.

Stegmann, M. B. (2000). Active appearance models: The-

ory, extensions and cases. Master’s thesis, Informatics

and Mathematical Modelling, Technical University of

Denmark, DTU.

Stegmann, M. B. and Pedersen, D. (2005). Bi-temporal 3D

active appearance models with applications to unsu-

pervised ejection fraction estimation. In The Interna-

tional Symposium on Medical Imaging, volume 5747.

van Assen, H., Danilouchkine, M., Frangi, A., Ords, S.,

Westenberg, J., Reiber, J., and Lelieveldt, B. (2006).

SPASM: a 3D-ASM for segmentation of sparse and

arbitrarily oriented cardiac MRI data. Medical Image

Analysis, 10(2):286–303.

FAST AND ROBUST LOCALIZATION OF THE HEART IN CARDIAC MRI SERIES - A Cascade of Operations for

Automatically Detecting the Heart in Cine MRI Series

345

Figure 4: User study at a glance: Averaged point-to-surface distances for 41 data sets (top left), RMS texture errors for all

validation data sets (top right), and time performance for all validation data sets (bottom).

Figure 5: The result: Initialization of the AAM’s mean at the central slices.

VISAPP 2008 - International Conference on Computer Vision Theory and Applications

346