A NEW METHOD FOR THE DETECTION OF BRAIN STEM IN

TRANSCRANIAL ULTRASOUND IMAGES

Josef Schreiber

∗

, Eduard Sojka

∗

, La

cezar Li

cev

∗

, Petra

nou

rilov

†

, Jan Gaura

∗

Department of Computer Science,

†

Department of Applied Mathematics

SB - Technical University of Ostrava, 17. listopadu 15/2172

Ostrava - Poruba, Czech Republic

David

Skoloud

ık

Neurological clinic, Faculty Hospital in Ostrava-Poruba, 17.listopadu 1790, Ostrava - Poruba, Czech Republic

Keywords:

Ultrasound images, brain stem, detection, noise, speckle, Parkinson’s disease, object recognition.

Abstract:

Transcranial sonography is to date the only method able to detect structural damage of the brain tissue in the

Parkinson’s disease patients. The problem is that the images provided by this method often suffer from a

very poor quality, which makes the ﬁnal diagnosis strongly dependent on experience of examinating medical

doctor. Our objective is to create a method that should help to minimize the physician’s subjectivity in the ﬁnal

diagnosis and should provide more exact information about the processed ultrasound images. The method

itself is divided into two phases. In the ﬁrst one, we try to locate the position of a minimal window containing

the brain stem in the analyzed image. In the second phase, we locate and measure the echogenic substantia

nigra area.

1 PARKINSON’S DISEASE

Parkinson’s disease (PD) belongs to the neurodegen-

erative diseases affecting mostly older people. It is a

chronic progressive disease that occurs if the nerve

cells in a part of the midbrain, called the substan-

tia nigra, die or are impaired. These nerve cells

produce dopamine, an important chemical messen-

ger that transmits signals from the substantia nigra to

other parts of the brain. These signals allow coordi-

nated movement. If the dopamine-secreting cells in

the substantia nigra die, the other movement control

centers in the brain become unregulated. Neuroimag-

ing methods are increasingly used as diagnostic tools

in patients presenting with parkinsonism. However,

brain computed tomography (CT) and magnetic res-

onance imaging (MRI) examinations are only able to

detect other aetiology than PD (Ressner, 2007). This

is why these traditional displaying methods like CT

and MRI are not considered to be conclusive (Bog-

dahn, 1998).

In 1995 Becker et al. published a study dealing

with the diagnostics of PD using transcranial sonog-

raphy (Becker, 1995). They showed that increased

echogenicity of substantia nigra is closely associated

with PD. Later, their research was followed by other

authors (Berg, 1999), (Berg, 2001). They proved that

hyperechogenic substantia nigra can be found in more

than 91% patients with PD. Nowadays the transcra-

nial sonography is considered to be the best possible

diagnostic tool for a detection of structural damage

of brain tissue in Parkinson’s disease patients. Ul-

trasonic imaging is based on detecting reﬂected and

scattered waves arising as a response to the emitted

wave with various frequencies. In general, the higher

the frequency is, the better and more detailed out-

put images can be obtained. Unfortunately, in the

case of transcranial sonography, we need to deal with

the scull that behaves as a barrier stopping all high-

frequency waves. This means that only the low fre-

quency probes (1-4 MHz) may be used. As a conse-

quence of this, transcranial sonography provides im-

ages of signiﬁcantly lower quality (see Figure 1).

The interpretation of ultrasound images is gener-

ally a difﬁcult task and the opinion of different med-

ical doctors is generally equivocal. The problem is

even more serious in transcranial sonography. Even if

the image is carefully evaluated by a physician, there

478

Schreiber J., Sojka E., Li

cev L., Šk

nourilová P., Gaura J. and Školoudík D. (2008).

A NEW METHOD FOR THE DETECTION OF BRAIN STEM IN TRANSCRANIAL ULTRASOUND IMAGES.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 478-483

DOI: 10.5220/0001057604780483

 SciTePress

Figure 1: An example of processed sono image.

is a signiﬁcant inﬂuence of subjectivity. By two dif-

ferent medical doctors, a different diagnosis may be

determined from one image. Another problem is a

long time treatment when the progress of the disease

must be determined from old images and repeating of

examination is no longer possible (

Skoloud

ık, 2007).

That is why there appeared the demand to create a tool

that would help physicians to objectivize the diagno-

sis process. The processing of ultrasound images is

widely discussed in literature. Various methods were

presented for an image segmentation (Noble, 2006),

(Boukerroui, 2003), (Bosch, 2002), noise and speckle

reduction (Magnin, 1982), (Rakotomamonjy, 2000),

(Kerr, 1986) or image enhancement (Lee, 1980), (Sat-

tar, 1997). However, according to our knowledge,

there are no studies dealing with processing the brain-

stem transcranial images or with computerized recog-

nition of objects in the substantia nigra area.

Since the image segmentation is strongly depen-

dent on the character of the processed image we need

to realize that the usage of classic ultrasound image

segmentation methods will be limited. Still, some in-

teresting studies dealing with ultrasound images were

published. Ballard et al. (Ballard, 1982) presented

region-based segmentation methods such as the re-

gion growing where they used the homogeneity of

inner regions in the images. Such approach is ob-

viously impropriate for transcranial images process-

ing since the images contain too much ultrasound

speckle. Mishra et al. (Mishra, 2006) presented

the active contour method in combine with the ge-

netic algorithms for the endocardial border detection.

A slightly different approach was used by Mignotte

(Mignotte, 2001). They used a statistical external en-

ergy in a discrete active contour for the segmentation

of parasternal images. Their work was followed by

many optimization efforts, e.g., Heitz (Heitz, 1994).

The method provided relatively good results and is

recommended for the noisy ultrasound images. Still,

Figure 2: An endocardial (left) and transcranial (right )ul-

trasound image. The different level of noise is evident.

if we compare the level of noise in the images the

method was tested on with transcranial ultrasound im-

ages, we see that our analysed images have signiﬁ-

cantly worse quality.

Another possible method for an ultrasound image

segmentation are the level sets, using an adaption of

the fast marching method. In 2003, Yan et al. (Yan,

2003) presented the purely edge-based version of this

method for the endocardial boundary detection. It was

later improved by Lin et al.(Lin, 2003). Their method

combined an edge and a region information in a level

set approach across spatial scales and it assumes that

a boundary is a closed curve. The method is supposed

to work well with the images of reasonably good qual-

ity. Klinger et al. (Klinger, 1988) presented a study

dealing with the echocardiographic images, based on

mathematical morphology. As well as the previous

method, this one also assumes to work with good

quality images. In 1999, Rekeczsky et al. (Rekeczky,

1999) and Binder et al. (Binder, 1999) came with the

artiﬁcial neural network method. Binder used a 2-

layer backpropagation network to identify a 7x7 pix-

els region with good results. Unfortunatly, we try to

locate a region 120x120 pixels large which means a

signiﬁcant growth of input information for the neural

network. Even if we decide to use only some impor-

tant parts of the image to reduce the input, we still

have to deal with the possiblity that the selected re-

gion is strongly covered by ultrasound speckle. Our

proposed method is designed to use as much infor-

mation from analysed image as possible to avoid be-

ing misled by the high level of ultrasound noise and

speckle.

2 BRAIN STEM LOCALIZATION

The method we have developed for processing the

brain-stem transcranial sono images is divided into

two phases. In the ﬁrst phase, we try to locate the

position of a minimal window in the processed im-

age containing the brain stem. To do so, we use a

modiﬁed template matching algorithm. Since every

human being is unique, the brain stems may slightly

A NEW METHOD FOR DETECTION OF BRAIN STEM IN TRANSCRANIAL ULTRASOUND IMAGES

479

Figure 3: Examples of the images used for the template con-

struction. For need of this paper, the images are displayed

with the same size.

differ in shape and size. Moreover, the objects be-

ing sought inside the stem may have different posi-

tion, size, and echogenicity, depending on the disease

progress. Therefore, for creating the template of the

stem, we choose several images that best represent,

according to our opinion, various possible shapes and

sizes of the stem. We also consider the seriousness

of the disease by choosing images depicting the situa-

tion in various stages of the disease progress (from

healthy persons to persons with an advanced stage

of the disease). The selection of images that are

used for the template construction is important in

our method. Therefore, the selection of images was

widely discussed with medical doctors to best fulﬁll

previously mentioned parameters. Overall, there were

20 selected images used for the template construction.

Four examples of these images can be seen in Figure

We construct the template that is used for match-

ing by simple averaging the particular selected images

of the brain stem. Firstly, the images are normalized

to the same size. In our experimental implementation,

we use the size of 120 × 120 pixels. After normaliz-

ing the size, we normalize the images of stem also

with respect to the mean value and the variance of

brightness. We do so by using the following formula

(x) = ab

(x) + c, (1)

where b

(x), b

(x) stand for the normalized and orig-

inal brightness, respectively, at a pixel whose position

is described by a two-dimensional vector x, and a and

c are constants that must be determined for each par-

ticular image. For determining them, the mean value

of brightness, denoted by µ

, and the variance of

brightness, denoted by σ

, in the original images are

needed. Let Ω stand for the set of all pixels in the

brain-stem image and let N be the size of this set. We

have

∑

x∈Ω

(x), (2)

∑

x∈Ω

(x) − µ

)

. (3)

In each normalized image, the normalization of

brightness aims at achieving a certain required mean

value, denoted by µ

and a required variance of

brightness, denoted by σ

. Simple mathematics

yields the following formulas for a and c

a =

, c = µ

− aµ

. (4)

The effect of normalization can be seen in Figures

4, 5. In Figure 6, the set of example images from Fig-

ure 3 can be seen in normalized form. An example of

the template that was obtained by averaging the brain-

stem images using Equation 5 is depicted in Figure 7.

Figure 4: The histogram of original image.

Figure 5: The histogram of normalized image.

In the pattern matching algorithm, we will also use

the variance of brightness in particular pixels that can

be expressed as follows

(x) =

∑

j=1

n j

(x), (5)

(x) =

∑

j=1

n j

(x) − µ

(x))

. (6)

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

480

Figure 6: The brain-stem images from Figure 3 after the

normalization process.

Figure 7: The constructed template.

In the above formulas, M is the number of partic-

ular normalized brain-stem images that are used for

creating the template; b

n j

stands for the j-th such im-

age.

In the ﬁrst step of our method, we try to locate the

position of the brain stem. We introduce a possibil-

ity, denoted by π(u

, x), of the event that the template

point with the coordinates x corresponds to the image

point with the coordinates x + u

(Sojka, 2006). This

possibility may be determined from the difference of

brightness

∆b = b(x + u

) − t(x), (7)

where b(x +u

) is the brightness of the pixel with co-

ordinates x + u

in processed image, and t(x) is the

brightness in the corresponding template pixel. Let it

be pointed out that u

characterizes the template po-

sition that is just being processed.

We suppose that the possibility distribution may

be described by a certain chosen function ϕ. Figure

8 shows an example of such a function. For the con-

Figure 8: The distribution of possibility ϕ (we use the Gaus-

sian function).

struction of ϕ, we use the deviation σ

(x) that was

determined in Equation 6.

To obtain the possibility of the event that the im-

age pixel just being processed corresponds to the pixel

from the template, we use the following equation

π(u

, x) = ϕ(b(x + u

) − t(x), σ

(x)). (8)

To characterize the quality of matching at the po-

sition u

, we introduce the quantity S(u

) character-

izing the number of pixels, i.e., the ”net area” that can

successfully be matched to the template. We have

S(u

) =

∑

x∈Ω

π(u

, x), (9)

where Ω stands for the set of template pixels.

The ﬁnal goal is to ﬁnd the value of u that maxi-

mizes the value of S(u

). The value of u then deter-

mines the position of the window that should contain

the brain stem (Figure 9).

Figure 9: An image with the recognized brain-stem object.

It is obvious that during the brain-stem detec-

tion, each processed window from the analyzed image

must be normalized in the same way as the images

used for the template construction.

3 ANALYSIS OF BRAIN STEM

To obtain the information about the disease progress,

we now need to locate and measure the objects in-

side the brain stem, which is the second step of our

A NEW METHOD FOR DETECTION OF BRAIN STEM IN TRANSCRANIAL ULTRASOUND IMAGES

481

Figure 10: An image with the recognized brain stem and the

highlighted objects.

method. In the image, these objects appear as the

areas with a higher level of echogenicity inside the

substantia nigra area. It is a difﬁcult task to correctly

identify these objects because the areas may have in-

sufﬁcient contrast.

We locate the objects in the brain stem as the ar-

eas with a higher level of brightness that are found

using the region growing method. Homogenity of

brightness is a criterion that is used for growing. Af-

ter growing, the regions of interest are usually parti-

tioned into several smaller areas (Figure 10). There-

fore, morphological closing is carried out after grow-

ing to connect the sub-areas together. If it is required

by a doctor, the convex hull of the found area may

be computed too. The regions that have been found

are then checked for the shape and size, which sepa-

rates the objects of interest in the stem. The numerical

characteristics are then computed. For all recognized

objects, we determine the number of pixels the ob-

jects are composed of, their average brightness, and

the location of their gravity centers. Besides comput-

ing the characteristics of the objects, they can also be

highlighted in the images (Figure 10).

Naturally, there is also a possibility to correct the

obtained results manually, if necessary, and remove

possible unwanted objects that are considered to be

only a noise, ultrasound speckle or possibly even a

part that does not belong to the brain stem area.

4 ACHIEVED RESULTS

To test the succesfulness of our method in brain stem

localization, we used a sample of 170 images in which

we tried to locate the correct brain-stem position. The

result (the quality of recognition) was classiﬁed with

the marks between 1 and 3. The mark 1 means that the

position was recognized correctly and accurately. The

mark 2 means that the position was determined inac-

curately but not completely incorrectly. In this case,

the position was usually determined with an error up

to 10-15 pixels. The mark 3 means that the method

determined an incorrect position. For our set of test

images, we obtained the results that are summarized

in Table 1.

Table 1: The results of the brain-stem localization achieved

by the presented method. The ﬁrst column determines the

quality of recognition. The second one shows the number of

images recognized with corresponding quality and the last

column displays the overall percentage.

Quality of

recognition

Number of images Results in %

1 129 75,9

2 4 2,3

3 37 21.8

The mark 1 was achieved in nearly 76% of

images. This can be considered as a good result since

we have to realize that the method must deal with

images of various quality. The difference between

the good and and bad image is shown in Figure 11.

In the left image, we can clearly see the shape of the

brain stem. For our method, the right image is very

difﬁcult to determine the correct brain-stem location.

Figure 11: These images illustrate the difference between

the good quality and the bad quality images. While in the

left image, the shape and position of brain stem is obvious,

in the right image, two places with similar shape to the brain

stem may be found.

5 CONCLUSIONS

The computer processing of transcranial ultrasound

images is a complicated task. Images often suffer

from a very poor quality and they often have a high

level of noise and speckle. The objects that were rec-

ognized are often discontinuous, in worse cases even

incomplete. The objects inside the brain stem often

have insufﬁcient contrast and they are usually frag-

mented by ultrasound speckle. Still, the objective

evaluation of these images can be very helpful in the

Parkinson disease diagnostics and treatment. It can

help the medical doctors to determine the correct di-

agnosis as well as the level of the disease progress.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

482

The exact and objective information about the ex-

amination from particular date can especially help in

longer time diagnostics when repeating of the exami-

nation is no longer possible.

Our method may be divided into two phases. At

ﬁrst, it attempts to correctly identify the position of

brain stem in processed image. This phase is crucial

in overall diagnostics and this paper focuses mostly

on this part. In the second phase, we detect the ob-

jects of interest in the brain stem. The detection of

existence, shape, size, and echogenicity of these ob-

jects is a valuable contribution to the diagnostics of

Parkinson’s disease.

Achieved results obtained during testing make us

believe that the method we have developed for the de-

tection and analysis of the brain stem in transcranial

ultrasound images is successful. From the tested im-

ages, we obtained good results. In 76% of cases, the

position of the brain stem was correctly determined.

ACKNOWLEDGEMENTS

Presented results had been obtained during solving

the grant project code T401940412 supported by the

Academy of Sciences of the Czech Republic.

REFERENCES

Ballard, D.H., B. C. (1982). Computer vision. Prentice-

Hall, Englewood Cliffs, NJ.

Becker, G., B. U. G. C. e. a. (1995). Transcranial color-

coded real-time sonography of intracranial veins. nor-

mal values of blood ﬂow velocities and ﬁndings in

superior sagittal sinus thrombosis. Journal of Neu-

roimaging, 5:87–94.

Berg, D., B. G. Z. B. T. O. H. E. e. a. (1999). Vulnerability

of the nigrostriatal system as detected by transcranial

ultrasound. Neurology, 53:1026–1031.

Berg, D., S. C. B. G. (2001). Echogenicity of the substantia

nigra in parkinson’s disease and its relation to clinical

ﬁndings. Journal of Neurology, 248:684–689.

Binder, T., S. M. M. D. S. H. B. T. M. G. P. G. (1999). Ar-

tiﬁcial neural networks and spatial temporal contour

linking for automated endocardial contour detection

on echocardiograms: A novel approach to determine

left ventricular contractile function. 25(7):1069–1076.

Bogdahn, U., B. G. S. F. (1998). Echoenhancers and tran-

scranial color duplex sonography. Blackwell Science,

Berlin.

Bosch, J.G., M. S. L. B. N. F. K. O. S. M. R. J. (2002). Au-

tomatic segmentation of echocardiographic sequences

by active appearance motion models. 21(11):1374–

1383.

Boukerroui, D., B. D. N. J. B. O. (2003). Segmentation

of ultrasound images - multiresolution 2d and 3d al-

gorithm based on global and local statistics. Pattern

Recognition Letters, 24(4-5):779–790.

Heitz, F., P. P. B. P. (1994). Multiscale minimization of

global energy functions in some visual recovery prob-

lems. 59:125–134.

Kerr, A.T., P. M. F. F. H. J. (1986). Speckle reduction in

pulse echo imaging using phase insensitive and phase

sensitive signal processing techniques. 8:11–28.

Klinger, J.W.J., V. C. F. T. A. L. T. (1988). Segmentation

of echocardiographic images using mathematical mor-

phology. 35(11):925–934.

Lee, J. (1980). Digital image enhancement and noise ltering

by use of local statistics. PAMI-2(2):165–168.

Lin, N., Y. W. D. J. (2003). Combinative multi-scale level

set framework for echocardiographic image segmen-

tation. 7(4):529–537.

Magnin, P.A., v. R. O. T. F. (1982). Frequency compound-

ing for speckle contrast reduction in phased array im-

ages. Ultrasonic Imaging, 4:267–281.

Mignotte, M., M. J. (2001). A multiscale optimization ap-

proach for the dynamic contour-based boundary de-

tection issue. 25(3):265–275.

Mishra, A., D. P. G. M. K. (2006). A ga based approach for

boundary detection of left ventricle with echocardio-

graphic image sequences. 21(11):967–976.

Noble, J.A., B. D. (2006). Ultrasound image segmentation:

A survey. IEEE Transactions on medical imagining,

25(8):987–1010.

Rakotomamonjy, A., D. P. M. P. (2000). Wavelet-based

speckle noise reduction in ultrasound b-scan images.

22:73–94.

Rekeczky, C., T. A. V. Z. R. T. (1999). Cnn-based spa-

tiotemporal nonlinear ﬁltering and endocardial bound-

ary detection in echocardiography. 27(1):171–207.

Ressner, P., v. D. H. P. K. P. (2007). Hyperechogenicity of

the substantia nigra in parkinson’s disease. Journal of

Neuroimaging, 17(Issue 2):164–167.

Sattar, F., F. L. S. G. L. B. (1997). Image enhancement

based on nonlinear multi-scale method. 6:888–895.

Sojka, E. (2006). A motion estimation method based on

possibility theory. In Proceedings of IEEE ICIP, pages

1241–1244.

Skoloud

ık, D., F. T. B. P. L. K. R. P. Z. O. H. P. H. R. K. P.

(2007). Reproducibility of sonographic measurement

of the substantia nigra. Ultrasound in Medicine & Bi-

ology, 33(9):1347–1352.

Yan, J.Y., Z. T. (2003). Applying improved fast marching

method to endocardial boundary detection in echocar-

diographic images. 24(15):2777–2784.

A NEW METHOD FOR DETECTION OF BRAIN STEM IN TRANSCRANIAL ULTRASOUND IMAGES

483