Automated Segmentation of Upper Airways from MRI

Vocal Tract Geometry Extraction

Antti Ojalammi and Jarmo Malinen

Department of Mathematics and Systems Analysis, Aalto University, Otakaari 1, Espoo, Finland

Keywords:

MRI, 3D Image Processing, Automatic Surface Extraction, Vocal Tract.

Abstract:

An algorithm for automatically extracting a triangulated surface mesh of the human vocal tract from 3D MRI

data is proposed. The algorithm is based on a combination of anatomic landmarking, seeded region growing,

and smoothing. Using these methods, a mask is automatically created for removing unwanted details not

associated with the vocal tract from the MRI voxel data. The mask is then applied to the original MRI data,

after which marching cubes algorithm is used for extracting a triangulated surface. The proposed method can

be used for processing large datasets, e.g., for validation of numerical methods in speech sciences as well as

for anatomical studies.

1 INTRODUCTION

We propose an algorithm for extracting vocal tract

(VT) geometries from greyscale voxel images, produ-

ced by static 3D Magnetic Resonance Imaging (MRI).

The motivation for such an algorithm, requiring at

most minimal user intervention, lies in the need to

process large MRI datasets of the upper airways and

mouth area. As such, the voxel data produced by an

MRI scanner is suitable for patient examinations by

specialists such as radiologists, using the software and

the user interface provided by the scanner manufactu-

rer. Applications for the extracted surface models for

the VT include (i) producing anatomic models by ra-

pid prototyping for patient examinations, planning of

treatment, etc., by medical professionals, and (ii) ge-

nerating computational meshes for numerical simula-

tions of a wide range of biophysical phenomena rela-

ted to speech production, breathing, and swallowing

function.

MRI examinations are useful for improving the

current understanding of the relation between ana-

tomy, phonation, and articulation. Depending on the

particular purpose, MRI data is required both from he-

althy test subjects as well as from patients requiring,

e.g., surgical treatment or rehabilitation. Because of

the lack of ionising radiation, the MRI examinations

are an attractive alternative to X-ray Computed Tomo-

graphy that would, however, produce a better spatial

and temporal resolution. This aspect is particularly

important when imaging healthy test subjects without

an underlying medical condition that would warrant

the use of radiation.

A lot of earlier work has been done in segmen-

tation as well as image processing, both overall and

in general medical context; see, e.g., (Gonzalez and

Woods, 2001; Sharma et al., 2010). The main issues

in segmenting the MRI data of the VT are caused by

the quality of raw voxel data and the characteristics

of the MRI technology: (i) the inability of MRI to

isolate osseous structures from the air volume due to

the common low hydrogen content of both of these

media, and (ii) motion artefacts due to the scanning

time that may exceed 10 seconds for a single statio-

nary 3D image in high resolution. As such, both of

these problems can be alleviated (however, at a de-

ﬁnite cost) by choices in MRI sequences, their para-

meterisation, and even by using a scanner with ﬁeld

strength at 3 T or even higher. Further challenges are

associated to non-typical or even pathological anato-

mies in test subjects: in some VT conﬁgurations, the

air passage may be so narrow that a naively processed,

low-resolution MRI might lack any opening at all.

The proposed algorithm is able to extract a VT

surface taking into consideration each of these chal-

lenges. The proposed algorithm is completely au-

tomatic, and it derives its parameters from the MR

image data to be processed. The algorithm aims to be

as robust as possible, with the goal to extract a rea-

sonable candidate surface even from data having he-

avy motion artefacts or postural abnormalities. This

is preferred due to the large amount of image data,

Ojalammi A. and Malinen J.

Automated Segmentation of Upper Airways from MRI - Vocal Tract Geometry Extraction.

DOI: 10.5220/0006138300770084

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 77-84

ISBN: 978-989-758-215-8

Figure 1: An extracted isosurface from raw MRI without

any preliminary modiﬁcations. Observe that the osseous tis-

sue is merged with the air volume.

which then allows for the application of statistical

analysis on numerical results.

The robustness of VT surface extraction has been

validated by applying the algorithm to a dataset of 3D

MRI from one male and one female test subject ut-

tering vowels, comprising of 109 images.

This da-

taset is used as the reference of the characteristics of

the MRI data in this article. We show two examples

of challenging VT geometries for the extraction. In

Figure 8, the resolution of the MRI data is insufﬁ-

cient for resolving the piriform sinuses. In Figure 7

(which is not part of the validation data set), the ear-

lier version of the extraction algorithm (Aalto et al.,

2013) was unable to resolve the vicinity of the uvula

whereas the proposed algorithm is able reproduce the

opening as required.

Measurements are performed on a Siemens Mag-

netom Avanto 1.5T scanner using 3D VIBE MRI se-

quence (Rofsky et al., 1999). Further details about

the acquisition of the MRI data have been explained

in (Aalto et al., 2014, Section 3).

2 BACKGROUND

Vocal tract segmentation from MRI is a long-standing

technical challenge in speech research. Semi-

automated algorithms have been developed since the

MRI resolution and the scanning time ﬁrst became

practicable for capturing articulation (Niikawa et al.,

1996; Baer et al., 1987; Baer et al., 1991; Engwall and

The original dataset of the two test subjects contained

114 MR images, of which ﬁve were deemed as failed scans,

and hence, excluded from the validation set of this article.

Badin, 1999; Story et al., 1996; Story et al., 1998; Ta-

kemoto et al., 2004; Aalto et al., 2014; Aalto et al.,

2011). When using 3D MRI, experiments involving

prolonged vowel utterances have been most common.

There exist more generic softwares that can be

used for extracting VT geometries, but the particu-

lar challenges related to the head and neck area ana-

tomy require a highly tailored approach. For exam-

ple, the segmentation software Vascular Modelling

Toolkit (Vascular Modeling Toolkit, 2016) for medi-

cal data on blood vessels can be used due to the tubu-

lar shape of the VT. However, generic software – or

software mainly intended for other purposes – require

user input to deﬁne initial conﬁgurations, etc., as well

as various parameters. These parameter values can-

not always be directly inferred from the data, and the

user must usually proceed based on trial and error to

produce high quality segmented data. Moreover, the

sensitivity with respect to parameter values is an is-

sue, and different parts of the anatomy may beneﬁt

from different parameterisations. All this adds to the

amount of manual work which easily becomes prohi-

bitively high for large scale studies or in commercial

applications.

Segmentation approaches based on an estimated

VT centreline have been proposed (Poznyakovskiy

et al., 2015). This approach reduces the 3D segmenta-

tion task into two dimensions where, e.g., active con-

tour methods (i.e., snakes) can be used. With such

methods, a multitude of parameters is needed. More-

over, special care must be taken when generating the

VT centreline (lacking a unique deﬁnition due to the

complicated geometry) which is a non-trivial task in

itself. It should be pointed out that VT centrelines and

intersection surface areas are also required in some of

the speech acoustic models. It is, however, a different

matter to derive such centrelines from a triangulated

VT surface model, compared to deriving them from

raw voxel data. In two-dimensional sections, some

parts of the VT (such as piriform sinuses and the val-

leculae) may appear not connected even though they

are connected in three dimensions. This adds to the

complications when using an active contour method.

An almost automatic segmentation technique was

presented in (Aalto et al., 2013), and the current work

is based on the lessons learned since then. The earlier

approach requires artefact model geometries for the

maxilla and mandible that have to be created for each

test subject separately. The artefact models are then

automatically aligned with the surface extracted from

the target data in three dimensions in order to mask

away the osseous structures that would interfere with

the air volume in the VT as shown in Fig. 1. Both the

surface models had to be represented as point clouds

BIOIMAGING 2017 - 4th International Conference on Bioimaging

since the alignment process is based on point cloud

registration (Rusu and Cousins, 2011). Detecting and

correcting misalignments proved to be labourious.

3 OVERVIEW OF THE

ALGORITHM

The extraction of the air/tissue interface from MRI

data requires four major steps as follows:

1. Preprocessing: Firstly, the original voxel number

is increased 8-fold to accommodate a more pre-

cise estimate of air/tissue interface making opti-

mal use of the information contained in the grey

values of the MR image. The edge deﬁnition and

the contrast of the MR image is improved by using

standard image processing algorithms. An ini-

tial threshold for the grey value due to air volume

(showing as low intensity in MRI) is deﬁned for

Step 3.

2. Landmarking: Anatomic features near the mouth

and the oropharynx are detected as required in

Step 3.

3. Mask creation: A binary mask containing only the

exterior and the VT air column volume is genera-

ted by an iterative intensity based region growing

algorithm and smoothing. With the aid of anato-

mic landmarks, the mask excludes osseous struc-

tures that are not discernible from air in MRI data.

The threshold value from Step 1 is increased du-

ring the iteration, leading to a larger volume inter-

preted as air by the region growing algorithm.

4. Surface extraction: A triangulated surface mesh is

extracted from the original MRI data using a do-

wnsampled version of the mask created in Step 3.

The surface mesh produced by the marching cu-

bes algorithm is consistent with a level set of grey

values, with the aid of an empirically obtained

threshold value. Volume preserving smoothing is

applied to ﬁnalise the extracted mesh.

After further processing, the resulting triangulated

surface mesh can be used in numerical simulations,

for extraction of VT centrelines and cross-section

areas, and rapid prototyping of physical models.

4 PREPROCESSING

In order to improve contrast and enhance edges, MR

images must be preprocessed. The initial resolution

of the MR image data is quite low: 1.8 mm for the

female and 1.9 mm for the male test subject.

As a

ﬁrst step, we increase the number of voxels 8-fold

by linear interpolation in order to halve the distance

between adjacent slices. This gives us more room

for play when it comes to narrow parts of the vo-

cal tract, such as at the glottis or near the uvula.

Secondly, we run the symmetric nearest neighbour

(SNN) ﬁlter adapted from (Hong et al., 2004) with

full 3-dimensional connectivity in order to enhance

edge deﬁnition. Thirdly, the way how the MRI scan-

ner reconstructs the image may lead to wrap-around

where the back of the skull appears in front of the pa-

tient’s face. These imaging artefacts are removed by

discarding all but the largest connected component of

imaged tissue that is in contact with the air space in

front of the mouth.

After the aforementioned preliminary steps, an

initial threshold value is evaluated corresponding to

the grey level of air deduced from the histogram of

the MRI data. The initial threshold value is used for

separating the air and tissue components in the image

data as explained in Section 6.1. Histogram norma-

lisation is applied to voxels with grey values higher

than the threshold in order to distribute the tissue in-

tensities more evenly. This makes the gradient of the

grey values less steep near the air/tissue interface, al-

lowing more control and precision in mask creation

(Step 3 in Section 3) as a function of the improved,

higher threshold value.

Carrying out the steps described in Section 5, po-

sition information is obtained near the mouth and the

oropharynx. With the aid of it, the grey values at the

throat area are increased in order to remove low inten-

sity spots near the mandible due to motion artefacts.

An example of mid-sagittal head and neck MRI

slice before and after pre-processing is shown in Fi-

gure 2. It is worth noting that narrow passages around

the glottal area are not always detected as air volume

at this stage. This is acceptable since the algorithm

will later block out unwanted regions in order to be

able to iteratively raise the air/tissue threshold value

in Step 3.

5 ANATOMIC LANDMARKS

The extraction algorithm requires estimates for the lo-

cation of features near the mouth and the oropharynx.

We obtain these by using a facial proﬁle constructed

from the preprocessed image data.

The number of voxels is the same in all MR images of

the validation set. The voxels are isotropic, but the voxel

size varies depending the size of the test subject.

Automated Segmentation of Upper Airways from MRI - Vocal Tract Geometry Extraction

Figure 2: A mid-sagittal section of MRI data before and after the pre-processing steps described in Section 4.

We use the extreme anterior coronal section as a

seed for the region growing algorithm, allowing the

expansion along the rays pointing in the coronal di-

rection, only. Voxels with grey values below the ini-

tial threshold is marked with 1 (with the default value

being 0) until grey values corresponding to tissue are

met. This produces a binary image M of the same di-

mensions, say N

× N

, as the preprocessed data.

Summing over the coronal direction gives a matrix P

whose integer elements are the distances (in voxels)

from the most anterior of the coronal sections of the

preprocessed MR image to their nearest tissue voxels.

More precisely, the coordinates of M, indexed by

(i, j, k), correspond to sagittal, transverse, and coro-

nal directions, respectively. The two-dimensional dis-

tance proﬁle P

is given by

(i, j) =

∑

k=1

M(i, j, k). (1)

The coordinates of the nose can be easily found from

using peak detection, and we denote the lowest va-

lue of P

corresponding to this peak by the integer

≥ 0, (i.e., the distance of the nose to the most an-

terior coronal plane in the image). Unfortunately, the

peak value k

due to the nose may appear in many

elements of P

To obtain the proﬁle P

of the face shown in Fi-

gure 3 (right panel), we average P

over 2n +1 sagit-

tal sections, centred at the mid-sagittal section with

coordinate m. More precisely,

( j) =

m+n

∑

i=m−n

(i, j) where

m = round



mean



i| min

(i, j) = k



For the data used in this article, the value n = 20 was

used corresponding to the distance of 19 mm in both

directions.

The transverse position of the chin is quite easy to

ﬁnd from P

. The mouth position can now be found

from between the known positions of the chin and the

nose. Using the anatomic landmarks obtained above,

the anterior air/tissue interface of the oropharynx can

be located. Knowing the positions of these anatomic

details are sufﬁcient for constructing the mask (Step 3

in Section 3).

50 100 150 200

-40

-20

100

120

Figure 3: Left: Example of a distance function P

with

n = 10. Right: A facial proﬁle P

as a function of the

transverse coordinate.

6 MASK CREATION

We proceed to discuss Step 3 in Section 3. Because of

the restrictions imposed by the MRI scanner, a mask

is required for blocking out the osseous tissues that

would otherwise mix with the air column interior. The

mask is a binary array of the same 3D dimensions as

the postprocessed MR image.

We create the mask using seeded region growing

based on intensity values (Adams and Bischof, 1994).

The vocal tract has the unique advantage (over the os-

seous tissue volumes) of being connected to the air

volume outside the patient. Thus, the most anterior

coronal section of the voxel data is used as a seed.

BIOIMAGING 2017 - 4th International Conference on Bioimaging

Given the high contrast between air and soft tissue

voxel intensities, isolating the maxilla and the mandi-

ble (including parts of the dental structures) from the

air/tissue interface is the only remaining challenge.

6.1 Masking the Osseous Structures

In order to avoid regarding osseous structure as air

in the MR image, the passages caused by the lack of

teeth visibility must be closed. We carry out the clo-

sure by smoothing the preprocessed image data with a

Gaussian kernel (std= 0.65), thus spreading the high

intensity areas around the positions of teeth.

The closure is performed by iterating the follo-

wing steps, using the preprocessed MR image as

image data in the ﬁrst iteration:

1. The image data is smoothed (again), and a binary

mask is extracted from it as described above.

2. The surface area of the edge (obtained by morpho-

logical operations) of this mask is computed.

3. If a drop large enough is observed in the computed

surface area (see Figure 4), then the iteration is

terminated. Otherwise, return to the ﬁrst step of

this iteration.

We call the outcome of this process the initial ex-

traction which is a binary mask. The large drop in

the surface area is an indication that the passages due

to teeth have been closed. An example of such a drop

at the 5th iteration round is observed in Figure 4. We

point out that is important to use the surface area of

the mask edge rather than the volume of the mask

since the mandible and maxilla are thin structures

compared to the VT.

The iterative smoothing procedure described

above may also close off the narrow parts of the vo-

cal tract such as the passage near the uvula. Since

the initial extraction is only used for masking out the

teeth, this is acceptable. We ﬁrst expand the initial

extraction in all directions by a few voxels in order to

account for the loss of edges which are actually part

of the air column in the VT; see Figure 6 (left panel).

The expanded initial extraction is used to mask away

from the preprocessed image data the unwanted de-

tails in the anterior part of the mouth cavity; Figure 6

(middle panel). From this masked image, a reﬁned

mask is extracted without using any smoothing. The

reﬁned mask follows the outline of the VT, but due to

a possibly too low threshold for the grey values, it is

likely not to contain every detail of it – especially the

piriform sinuses and valleculae.

The reﬁned mask is applied to the preprocessed

image data. By increasing the threshold of the grey

values, a further reﬁned mask is produced from it.

1 2 3 4 5 6 7 8 9 10

2.5

3.5

×10

Figure 4: The computed surface area of the mask edge plot-

ted as a function of the number of smoothing steps.

Figure 5: Final extractions of the air/tissue interface from

a female (left) and male (right) test subject with the faces

removed for visual clarity. The vocal tract conﬁgurations

correspond to the vowels [œ:] and [y:], respectively.

This process of reﬁning masks and increasing the

threshold is repeated until either (i) the most inferior

transverse section of the image data intersects the air

column described by the mask, or (ii) a pre-deﬁned

maximum threshold value is reached. In the latter

case, the full length of the VT air column fails to be

extracted. The ﬁnal mask is obtained by expanding

the outcome a few voxels in all directions.

6.2 Mouth Opening

Smoothing used in Section 6.1, Step 1, has the unwan-

ted side effect of closing the mouth opening in parti-

cular for vowels where the opening is already small.

To cancel this side effect, we restore the connectivity

by creating a hole (using an auxiliary mask) to the

smoothed image at the mouth position using the pre-

processed MR image. The auxiliary mask is produced

by seeded region growing in the coronal direction, as

explained in Section 5.

Automated Segmentation of Upper Airways from MRI - Vocal Tract Geometry Extraction

Figure 6: From left to right: A sagittal section from (i) the expanded initial extraction, (ii) the masked image used for the

reﬁned extraction, and (iii) the ﬁnal mask.

7 SURFACE EXTRACTION

The ﬁnal surface is extracted using the marching cu-

bes algorithm (Lorensen and Cline, 1987). First, the

mask described in Section 6 is applied to the origi-

nal MR image to obtain an artefact-free, intermediate

version of the same voxel data. Second, the marching

cubes algorithm is applied to the intermediate ver-

sion, producing a triangulated surface of the air/tissue

interface. Finally, surface smoothing is applied to

the triangulated surface using the lowpass ﬁlter from

iso2mesh (Fang and Boas, 2009). The ﬁnal outcome

is a triangulated surface model of the air/tissue inter-

face of the VT.

8 DISCUSSION

For validation, the segmentation algorithm was app-

lied to the dataset of 3D MR images (N = 109) descri-

bed in Section 1. The algorithm was able to automa-

tically produce an anatomically and phonetically rea-

listic geometry for the air/tissue interface in all cases.

The geometries were visually inspected for anatomic

correctness, and they were compared with the original

MRI voxel data when necessary. One of the compa-

rison methods is by superimposing the air/tissue con-

tour from the masked MR image (right before surface

mesh extraction) on the original bitmap as shown in

Figure 9.

The quality of the 55 VT conﬁgurations from

the male test subject was found to be excellent, and

no segmentation errors were observed. Additionally,

physical models of the surfaces corresponding to

quantal vowels [A, i, u] were created using rapid pro-

totyping. The acoustic properties of the models were

measured using frequency sweeps, and the resulting

estimates of power spectra correspond very well with

those of the recorded speech signals measured during

the MR imaging (Kuortti et al., 2016).

So as to the 54 VT conﬁgurations from the female

test subject, the outcome is less satisfactory: anato-

mic details near piriform sinuses were lost in some

of the surface models. Inspecting the original MRI

voxel data leads to the conclusion that there is insuf-

ﬁcient resolution in the grey values produced by the

MRI scanner as can be seen in Figure 8. The high po-

sition of the larynx and the small dimension of ana-

tomic details make the MR examination of the upper

airways challenging in this test subject.

Figure 7: Extracted VT geometry of a test subject where

the uvula is blocking a non-typically large section of the

oropharynx.

It is possible to improve the visibility of piriform

sinuses and valleculae by manually adjusting the grey

values near larynx in the preprocessing stage descri-

bed in Section 4. The advantage of manual interven-

BIOIMAGING 2017 - 4th International Conference on Bioimaging

Figure 8: A sagittal section from a female test subject. The

circled area shows that the connection between the piriform

sinus and the VT cannot be resolved.

tion is that the specialist can incorporate her anatomic

expertise to the work ﬂow, and restrict her work to

a very small part of the voxel data where automatic

segmentation proves insufﬁcient. Here, manual cor-

rections were not carried out since the purpose is to

evaluate the quality of automatic surface extraction.

9 CONCLUSIONS

A method for automatically extracting VT surface

meshes from MR images has been proposed. Vali-

dation of the method has been carried out by subjecti-

vely evaluating results produced from two test sub-

jects. Additionally, the outlines of the extractions

were visually compared against the MR image data

in order to verify that no obvious regions have been

omitted. It is also noted that the proposed method

performs better than our previous approach, with the

added beneﬁt of not having to manually create arte-

fact models for each test subject.

ACKNOWLEDGEMENTS

The authors have received funding from Vilho, Yrjö

and Kalle Väisälä Foundation of the Finnish Aca-

demy of Science and Letters and Magnus Ehrnrooth

Foundation for funding this research. The test subject

data collection procedure was accepted by the Ethics

Committee of the Hospital District of Southwest Fin-

land.

The authors wish to express their gratitude to

R. P. Happonen, R. Parkkola, and J. Saunavaara for

cooperation in MRI acquisition.

Figure 9: A sagittal section of the original MRI data supe-

rimposed with an outline describing the extracted air/tissue

interface.

REFERENCES

Aalto, D., Aaltonen, O., Happonen, R.-P., Jääsaari, P., Ki-

velä, A., Kuortti, J., Luukinen, J.-M., Malinen, J.,

Murtola, T., Parkkola, R., Saunavaara, J., Soukka, T.,

and Vainio, M. (2014). Large scale data acquisition of

simultaneous MRI and speech. Appl. Acoust., 83:64–

75.

Aalto, D., Aaltonen, O., Happonen, R.-P., Malinen, J.,

Palo, P., Parkkola, R., Saunavaara, J., and Vainio, M.

(2011). Recording speech sound and articulation in

MRI. In Proceedings of BIODEVICES 2011, pages

168–173.

Aalto, D., Helle, J., Huhtala, A., Kivelä, A., Malinen, J.,

Saunavaara, J., and Ronkka, T. (2013). Algorithmic

surface extraction from MRI data: modelling the hu-

man vocal tract. In Proceedings of BIODEVICES

2013, pages 257–260.

Adams, R. and Bischof, L. (1994). Seeded region growing.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 16(6):641–647.

Baer, T., Gore, J. C., Boyce, S., and Nye, P. W. (1987). Ap-

plication of MRI to the analysis of speech production.

Magn. Reson. Imaging, 5:1–7.

Baer, T., Gore, J. C., Gracco, L. W., and Nye, P. W. (1991).

Analysis of vocal tract shape and dimensions using

magnetic resonance imaging. J. Acoust. Soc. Am.,

90(2):799–828.

Engwall, O. and Badin, P. (1999). Collecting and analy-

sing two- and three-dimensional MRI data for Swe-

dish. TMH-QPSR, (3-4/1999):11–38.

Fang, Q. and Boas, D. A. (2009). Tetrahedral mesh genera-

tion from volumetric binary and grayscale images. In

2009 IEEE International Symposium on Biomedical

Imaging: From Nano to Macro, pages 1142–1145.

Gonzalez, R. C. and Woods, R. E. (2001). Digital Image

Processing. Addison-Wesley Longman Publishing

Co., Inc., Boston, MA, USA, 2nd edition.

Automated Segmentation of Upper Airways from MRI - Vocal Tract Geometry Extraction

Hong, V., Palus, H., and Paulus, D. (2004). Edge Preser-

ving Filters on Color Images, pages 34–40. Springer

Verlag, Berlin, Heidelberg.

Kuortti, J., Malinen, J., and Ojalammi, A. (in press). Post-

processing speech recordings during MRI. Biomed

Signal Process Control.

Lorensen, W. E. and Cline, H. E. (1987). Marching Cubes:

A High Resolution 3D Surface Construction Algo-

rithm. SIGGRAPH Comput. Graph., 21(4):163–169.

Niikawa, T., Matsumura, M., Tachimura, T., Wada, T., Shi-

mizu, K., and Hashimoto, Y. (1996). Accurate mea-

surement of three-dimensional shapes of a vocal tract

and a dental crown using magnetic resonance ima-

ging: Japanese fricative consonants. J. Acoust. Soc.

Am., 100(4):2658–2658.

Poznyakovskiy, A. A., Mainka, A., Platzek, I., and Mü-

rbe, D. (2015). A Fast Semiautomatic Algorithm for

Centerline-Based Vocal Tract Segmentation. BioMed

Research International, 2015:e906356.

Rofsky, N., Lee, V., Laub, G., Pollack, M., Krinsky, G.,

Thomasson, D., Ambrosino, M., and Weinreb, J.

(1999). Abdominal MR imaging with a volume-

tric interpolated breath-hold examination. Radiology,

212(3):876–884.

Rusu, R. B. and Cousins, S. (2011). 3D is here: Point Cloud

Library (PCL). In Robotics and Automation (ICRA),

2011 IEEE International Conference on, pages 1–4.

Sharma, N., Ray, A., Shukla, K., Sharma, S., Pradhan, S.,

Srivastva, A., and Aggarwal, L. (2010). Automa-

ted medical image segmentation techniques. J. Med.

Phys., 35(1):3.

Story, B., Titze, I., and Hoffman, E. (1996). Vocal tract

area functions from magnetic resonance imaging. J.

Acoust. Soc. Am., 100(1):537–554.

Story, B. H., Titze, I. R., and Hoffman, E. A. (1998). Vo-

cal tract area functions for an adult female speaker

based on volumetric imaging. J. Acoust. Soc. Am.,

104(1):471–487.

Takemoto, H., Kitamura, T., Nishimoto, H., and Honda, K.

(2004). A method of tooth superimposition on MRI

data for accurate measurement of vocal tract shape

and dimensions. Acoust. Sci. Technol., 25(6):468–

474.

Vascular Modeling Toolkit (2016). http://www.vmtk.org.

Accessed Sept. 28th, 2016.

BIOIMAGING 2017 - 4th International Conference on Bioimaging