Fully Automated Lung Volume Assessment from MRI in a

Population-based Child Cohort Study

Tatyana Ivanovska

, Pierluigi Ciet

, Adria Rerez-Rovira

, Anh Nguyen

, Harm Tiddens

Liesbeth Duijts

, Marleen de Bruijne

and Florentin W¨org¨otter

Department for Computational Neuroscience, Georg-August-University, G¨ottingen, Germany

Erasmus MC: University Medical Center Rotterdam, Rotterdam, Netherlands

Keywords:

MRI, Lung, Segmentation and Volumetry, Child Cohort.

Abstract:

In this work, a framework for fully automated lung extraction from magnetic resonance imaging (MRI) inspi-

ratory data that have been acquired within a on-going epidemiological child cohort study is presented. The

method’s main steps are intensity inhomogeneity correction, denoising, clustering, airway extraction and lung

region reﬁnement. The presented approach produces highly accurate results (Dice coefﬁcients ≥ 95%), when

compared to semi-automatically obtained masks, and has potential to be applied to the whole study data.

1 INTRODUCTION

Magnetic resonance imaging (MRI) is a non-invasive,

non-ionizing 3D imaging method that is increasingly

applied in research settings. Numerous MR data are

acquired from thousands of subjects (V¨olzke et al.,

2011; Hetterich et al., 2015).

Thereafter, the parameters of interest, such as lung

or liver volumes, need to be extracted from the im-

ages. Manual processing is rather unfeasible due to

time constraints and inter- and intra-observer variabil-

ity. Therefore, automated methods for segmentation

of different organs from MRI are developed (Balafar

et al., 2010; Setarehdan and Singh, 2012; Ivanovska

et al., 2014; Toennies et al., 2015).

However, in a child cohort MR technique is not

so easily implementable, since the participants might

be scared of the closed environment or the necessity

to lie still for some amount of time. Moreover, chil-

dren anatomy varies widely due to the different stages

of growth. Thus, anatomical assumptions used to de-

velop algorithms for adult data might not hold, and,

therefore, the methods implemented for adult subjects

may not be directly applicable.

Here, we discuss the data from a population-based

prospective study with a child cohort and, in particu-

lar, a lung segmentation problem from these data, and

propose a fully automated solution for lung segmen-

tation.

The paper is organized as follows. In Section 3,

the general study information as well as MR protocols

are presented. The ultimate goals of the pulmonary

study and the structured step-by-step tasks are formu-

lated in Section 4.1. In Section 2, related works are

discussed. The algorithmic solution is proposed in

Section 4.2. The results and ﬁndings are presented

and discussed in Section 5. The Section 6 concludes

the paper.

2 RELATED WORK

In recent years, multiple automated approaches for

lung segmentation were proposed (Ivanovska et al.,

2016; Ivanovska et al., 2012; Tustison et al., 2015;

Kohlmann et al., 2015; Heimann et al., 2012) have

been proposed. The approaches for detection of

the lung volumes from anatomical MR scans can be

roughly separated in two groups: classical intensity-

based and model-based methods. The model-based

methods consist from a prior model or atlas construc-

tion (Tustison et al., 2015; Tustison et al., 2011), and

require usually a signiﬁcant amount of training data.

The intensity-based methods (Ivanovska et al.,

2012; Kohlmann et al., 2015) rely on some low-level

features and general considerations on human lung

anatomy. Such methods are usually fast and require

no prior training.

Although there are some methods for airway sys-

tem segmentation available, there are only few meth-

ods that have been proposed for analysis of child

airway MR data. Heimann et al. (Heimann et al.,

Ivanovska T., Ciet P., Perez-Rovira A., Nguyen A., Tiddens H., Duijts L., de Bruijne M. and WÃ˝uergÃ˝uetter F.

Fully Automated Lung Volume Assessment from MRI in a Population-based Child Cohort Study.

DOI: 10.5220/0006075300530058

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 53-58

ISBN: 978-989-758-227-1

2012) presented a method for automated scoring of

regional lung perfusion in children with cystic ﬁbro-

sis using contrast enhanced MRI. They used a com-

bined intensity- and model-based approach. Arens et

al. (Arens et al., 2003) applied a fuzzy connectedness

approach to perform a detailed analysis of the upper

airway (Liu et al., 2002) in children with obstruc-

tive sleep apnea. Thayyil et al. (Thayyil et al., 2009)

used a semiautomatic method for non-invasive in-

ternal organ weight measurement using post-mortem

MR imaging in fetuses, newborns, and children.

3 MATERIALS

3.1 Information on the Generation R

Study

The Generation R Study is a population-based

prospective cohort study, which is initiated in Rotter-

dam, the Netherlands, from fetal life until adulthood.

The study is designed to identify the early environ-

mental and genetic causes and causal pathways lead-

ing to normal and abnormal growth, development and

health during fetal life, childhood and adulthood. The

study focuses on six areas of research:

• maternal health;

• growth and physical development;

• behavioral and cognitive development;

• respiratory health and allergies;

• diseases in childhood;

• health and healthcare for children and their par-

ents.

Main exposures of interest include environmental, en-

docrine, genetic and epigenetic, lifestyle related, nu-

tritional and socio-demographic determinants. In to-

tal, 9778 mothers with a delivery date from April

2002 until January 2006 were enrolled in the study.

Response at baseline was 61 %, and general follow-

up rates until the age of 6 years exceed 80 %. Data

collection in mothers, fathers and children include

questionnaires, detailed physical and ultrasound ex-

aminations, behavioral observations, and biological

samples. A genome and epigenome wide association

screen is available in the participating children. From

the age of 5 years, regular detailed hands-on assess-

ments are performed in a dedicated research center in-

cluding advanced imaging facilities such as Magnetic

Resonance Imaging (MRI). Eventually, results forth-

coming from the Generation R Study contributeto the

development of strategies for optimizing health (Jad-

doe et al., 2012).

3.2 MRI Protocol

The goal of the MRI-study, which comprises ap-

proximately 4000 children aged 9, is to assess car-

diac, pulmonary, fat, and liver parameters. The MR

scanner is a 3-T MR 750w (GE Healthcare, Mil-

waukee, WI, USA). The pulmonary sequence is the

3D Spoiled Gradient Echo (SPGR) at end-inspiration

and end-expiration repeated twice (2 insp and 2 exp).

The parameters are: TR=repetition time: 1.6 ms,

TE=echo time: 0.7 ms, Flip angle=2

◦

; Average: 0.75,

FOV=40cm; Matrix= 200× 200, in-plane resolution

2× 2× 2 mm

, 32 channel torso coil. Additionally, a

mock scanner is used, in which children can practice

to lie within the MR scanner in a friendly way and get

used to the scanner protocols.

3.3 Test Set and Expert Annotations

To evaluate the approach, we randomly selected ten

subjects and evaluated two inspiratory scans for each

of them, i.e., twenty datasets were used as a test set.

We asked an experienced observer to semi-

automatically measure lung volumes including tra-

chea with a simple global thresholding, which was a

reasonable trade-off between the time and measure-

ment accuracy.

It has to be noted that with such and approach tra-

chea was included in the lung volume, whereas the

automated method produces the results without the

tracheal volume. This does not appear to be a prob-

lem though, due to the fact that trachea is a relatively

small organ, when compared to lungs, and its volume

is a minor addition (about 200-300 ml) to the total

lung volume. Moreover, the spirometry parameters

do not exclude the tracheal volume as well. However,

in the automatic approach, the trachea is detected and

excluded, since we are interested in separate lung vol-

umes as well as furthertracheal analysis is planned for

future work.

The expert evaluated each dataset two times. In

Section 5, two rounds of measurements are denoted

as R

and R

4 METHODS

4.1 Formulated Task

The ultimate goal of the research is to fully analyze

the pulmonary system in a child cohort study using

the available MRI data. The sub-tasks include (but

are not limited to)

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

1. assessment of lung volumes from inspiratory MR

scans;

2. their correlation with spirometry data;

3. analysis of tracheal dimensions;

4. analysis of expiratory lung data.

In this work, we propose a solution to the ﬁrst sub-

task, namely, a fully automated framework for lung

volume extraction in inspiratory MR scans.

4.2 Proposed Approach

The proposed algorithm is the extended version of the

framework proposed by Ivanovska et al. (Ivanovska

et al., 2012). The framework scheme is shown in Fig-

ure 1.

Data Input

IIH Correction

Denoising

Clustering

Lung

Extraction

Trachea

Removal

Lung

Separation

Lung

Smoothing

LungVolumes

Figure 1: The framework scheme. The modules that differ

from the work of Ivanovska et al. (Ivanovska et al., 2012)

are marked green.

First, since the study data contain signiﬁcant in-

tensity inhomogeneity, it requires correction. The

recently proposed method N4ITK (Tustison et al.,

2010) is applied three times in a row, and the results

are of acceptable quality. The example result is shown

in Figure 2.

Second, the corrected data are denoised us-

ing a boundary preserving method, the classical

Anisotropic diffusion ﬁlter (Perona et al., 1994),

and the intensity clustering, namely, the Fuzzy C-

Means (Bezdek et al., 1984) method, is applied.

Third, the regions that are classiﬁed to the dark-

est intensity class are selected, and the background

is removed from further processing. Here, a special

region evaluation procedure is introduced to prevent

the extraction of non-lung regions. We apply a re-

gion growing-based procedure, which evaluates the

lung region overlap in consecutive slices. Since a

rather smooth lung area decay is expected, any re-

gions that have an overlap with regions in the previ-

ous slice smaller than a pre-selected threshold value

are disregarded from the further processing.

Fourth, the tracheal, namely, dark tubular regions

are identiﬁed using the Vesselness ﬁlter (Frangi et al.,

1998). It analyzes the eigenvalues of the Hessian ma-

trix. Thereafter, the ﬁlter’s response is overlaid with

the lung mask segmented in the previous step. Then,

the tracheal region is tracked and the cut is made,

where main bronchi enter the lung parenchyma. The

Vesselness ﬁlter allows one to detect tracheal regions

even in the cases, where there is no clearly visible

boundary between the trachea and the lung tissue.

Finally, the lung regions are separated with pre-

ﬁlled 3D Watershed (Roerdink and Meijster, 2000)

and smoothed with morphological operations (Sonka

et al., 2014) similar to the approach of Ivanovska et

al. (Ivanovska et al., 2011). The lung masks and vol-

umes are computed and saved to a ﬁle.

In Figure 3, we present 3D and 2D exam-

ple results, demonstrating the excluded trachea and

smoothed lungs.

5 RESULTS AND DISCUSSION

The automated approach was implemented in the

MeVisLab Framework (Heckel et al., 2009). The pa-

rameters were ﬁxed for the test set. Average compu-

tation time for each dataset is about 3-4 minutes on a

AMD Athlon II X4 860K, 4x 3700 MHz with 16 GB

DDR3-RAM.

To evaluate the expert readings and automatically

computed results, we use two metrics, namely, the

DICE coefﬁcient (Dice, 1945) and the Jaccard in-

dex (Real and Vargas, 1996). The deﬁnition of the

metrics is given below. For two sets, in our case, the

results obtained from the expert and the automatically

computed results, R

and R

, respectively, the overlap

coefﬁcients are computed as

DSC =

2|R

∩ R

| + |R

(1)

Jacc =

∩ R

∪ R

(2)

Since our observer analyzed the test set two times,

we assess the intra-observer variability and compare

the automatically obtained results R

to both expert

measurements, denoted R

and R

, respectively.

The comparison results (mean µ ± standard deviation

σ) for 20 test datasets are presented in Table 1. We

evaluated the intra-observervariability using the same

Fully Automated Lung Volume Assessment from MRI in a Population-based Child Cohort Study

Figure 2: Example slice in a sagittal view. Left: original data; Right: corrected data with three consecutive N4ITK cycles.

Figure 3: Example results from 2 datasets. Left: 3D results; Right: Segmentation results are overlaid with the original data in

axial projection.

metrics, and for R

vs R

the DICE and Jaccard co-

efﬁcients are 0.9674±0.00597and 0.9364± 0.02, re-

spectively.

As one can observe, the automatically computed

results lie close to both semi-automatically obtained

ground truth masks, and the Dice coefﬁcient is about

95%. The differences are due to the fact that the

expert neither excluded the tracheal region from the

evaluation nor smoothed the lung regions. In Figure 4,

an overlay example of the expert readings (white) and

the automated results (green) are documented. The

Table 1: Manual and automated result comparison.

Proposed approach Previous

method (Ivanovska

et al., 2011)

DICE Jaccard DICE Jaccard

vs. R

0.9521 ±

0.010

0.9083 ±

0.01

0.8901 ±

0.11

0.82 ±

0.0968

vs. R

0.95485±

0.018

0.9145 ±

0.01

0.8875 ±

0.0906

0.81 ±

0.1004

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

Figure 4: Comparison of manual readings (white) and auto-

mated results (green) in 2D.

expert readings differ from each other slightly (the

Dice coefﬁcient is about 97%), since the differences

are only due to the selected global threshold value.

Additionally, we applied the technique of

Ivanovska et al. (Ivanovska et al., 2011) to our test

set. That method was designed for and tested on a dif-

ferent sequence with a higher spatial resolution and

less artifacts. The results are also presented in Ta-

ble 1. The previous technique did not include any

intensity inhomogeneity correction and the trachea

removal procedure was based only on region grow-

ing in the segmentation mask. We assume that this

affected the results negatively (the DICE coefﬁcient

is less than 90%), since in some cases the parts of

lungs were either oversegmented or undersegmented

and misinterpreted as other structures and erroneously

removed. The proposed pipeline successfully over-

comes these problems and produces accurate results.

6 CONCLUSIONS AND FUTURE

WORK

In this paper, a fully automated approach for lung seg-

mentation in MRI data from the Generation R child

study. The results were applied to a sample of 20

datasets. Our expert established groundtruth in a

semi-automatic manner in two measurement sessions.

We assessed the segmentation accuracy by compar-

ing the automatically computed results to the expert

readings. Moreover, the comparison to a previously

established technique was also done. The proposed

framework produces highly accurate results and has a

potential to be applied to the whole pulmonary dataset

(above 4000 subjects).

Future extensions of the framework include analy-

sis of tracheal regions and segmentation of expiratory

scans.

REFERENCES

Arens, R., McDonough, J. M., Corbin, A. M., Rubin, N. K.,

Carroll, M. E., Pack, A. I., Liu, J., and Udupa, J. K.

(2003). Upper airway size analysis by magnetic res-

onance imaging of children with obstructive sleep ap-

nea syndrome. American journal of respiratory and

critical care medicine, 167(1):65–70.

Balafar, M. A., Ramli, A. R., Saripan, M. I., and Mashohor,

S. (2010). Review of brain mri image segmentation

methods. Artiﬁcial Intelligence Review, 33(3):261–

274.

Bezdek, J. C., Ehrlich, R., and Full, W. (1984). Fcm: The

fuzzy c-means clustering algorithm. Computers &

Geosciences, 10(2):191–203.

Dice, L. R. (1945). Measures of the amount of ecologic

association between species. Ecology, 26(3):297–302.

Frangi, A. F., Niessen, W. J., Vincken, K. L., and Viergever,

M. A. (1998). Multiscale vessel enhancement ﬁlter-

ing. In Medical Image Computing and Computer-

Assisted InterventationMICCAI98, pages 130–137.

Springer.

Heckel, F., Schwier, M., and Peitgen, H.-O. (2009). Object-

oriented application development with mevislab and

python. In GI Jahrestagung, pages 1338–1351. Cite-

seer.

Heimann, T., Eichinger, M., Bauman, G., Bischoff, A., Pud-

erbach, M., and Meinzer, H.-P. (2012). Automated

scoring of regional lung perfusion in children from

contrast enhanced 3d mri. In SPIE Medical Imaging,

pages 83150U–83150U. International Society for Op-

tics and Photonics.

Hetterich, H., Bayerl, C., Peters, A., Heier, M., Linkohr,

B., Meisinger, C., Auweter, S., Kannengießer, S. A.,

Kramer, H., Ertl-Wagner, B., et al. (2015). Feasibility

of a three-step magnetic resonance imaging approach

for the assessment of hepatic steatosis in an asymp-

tomatic study population. European radiology, pages

1–10.

Ivanovska, T., Buttke, E., Laqua, R., V¨olzke, H., and Beule,

A. (2011). Automatic trachea segmentation and evalu-

ation from mri data using intensity pre-clustering and

graph cuts. In Image and Signal Processing and Anal-

ysis (ISPA), 2011 7th International Symposium on,

pages 513–518. IEEE.

Ivanovska, T., Hegenscheid, K., Laqua, R., Gl¨aser, S., Ew-

ert, R., and V¨olzke, H. (2016). Lung segmentation of

mr images: A review. In Visualization in Medicine

and Life Sciences III, pages 3–24. Springer.

Ivanovska, T., Hegenscheid, K., Laqua, R., K¨uhn, J.-P.,

Gl¨aser, S., Ewert, R., Hosten, N., Puls, R., and

V¨olzke, H. (2012). A fast and accurate automatic

lung segmentation and volumetry method for mr data

used in epidemiological studies. Computerized Medi-

cal Imaging and Graphics, 36(4):281–293.

Fully Automated Lung Volume Assessment from MRI in a Population-based Child Cohort Study

Ivanovska, T., Laqua, R., Wang, L., Liebscher, V., V¨olzke,

H., and Hegenscheid, K. (2014). A level set based

framework for quantitative evaluation of breast tissue

density from mri data. PloS one, 9(11):e112709.

Jaddoe, V. W., van Duijn, C. M., Franco, O. H., van der Hei-

jden, A. J., van IIzendoorn, M. H., de Jongste, J. C.,

van der Lugt, A., Mackenbach, J. P., Moll, H. A., Raat,

H., et al. (2012). The generation r study: design and

cohort update 2012. European journal of epidemiol-

ogy, 27(9):739–756.

Kohlmann, P., Strehlow, J., Jobst, B., Krass, S., Kuhnigk, J.-

M., Anjorin, A., Sedlaczek, O., Ley, S., Kauczor, H.-

U., and Wielp¨utz, M. O. (2015). Automatic lung seg-

mentation method for mri-based lung perfusion stud-

ies of patients with chronic obstructive pulmonary dis-

ease. International journal of computer assisted radi-

ology and surgery, 10(4):403–417.

Liu, J., Udupa, J. K., Odhner, D., McDonough, J. M., and

Arens, R. (2002). Upper airway segmentation and

measurement in mri using fuzzy connectedness. In

Medical Imaging 2002, pages 238–247. International

Society for Optics and Photonics.

Perona, P., Shiota, T., and Malik, J. (1994). Anisotropic

diffusion. In Geometry-driven diffusion in computer

vision, pages 73–92. Springer.

Real, R. and Vargas, J. M. (1996). The probabilistic basis

of jaccard’s index of similarity. Systematic biology,

45(3):380–385.

Roerdink, J. B. and Meijster, A. (2000). The watershed

transform: Deﬁnitions, algorithms and parallelization

strategies. Fundamenta informaticae, 41(1, 2):187–

228.

Setarehdan, S. K. and Singh, S. (2012). Advanced algo-

rithmic approaches to medical image segmentation:

state-of-the-art applications in cardiology, neurology,

mammography and pathology. Springer Science &

Business Media.

Sonka, M., Hlavac, V., and Boyle, R. (2014). Image

processing, analysis, and machine vision. Cengage

Learning.

Thayyil, S., Schievano, S., Robertson, N. J., Jones, R.,

Chitty, L. S., Sebire, N. J., Taylor, A. M., group, M. M.

R. I. A. S. C., et al. (2009). A semi-automated method

for non-invasive internal organ weight estimation by

post-mortem magnetic resonance imaging in fetuses,

newborns and children. European journal of radiol-

ogy, 72(2):321–326.

Toennies, K. D., Gloger, O., Rak, M., Winkler, C., Klemm,

P., Preim, B., and V¨olzke, H. (2015). Image analysis

in epidemiological applications. it-Information Tech-

nology, 57(1):22–29.

Tustison, N. J., Avants, B. B., Cook, P. A., Zheng, Y., Egan,

A., Yushkevich, P. A., and Gee, J. C. (2010). N4itk:

improved n3 bias correction. Medical Imaging, IEEE

Transactions on, 29(6):1310–1320.

Tustison, N. J., Avants, B. B., Flors, L., Altes, T. A.,

de Lange, E. E., Mugler, J. P., and Gee, J. C. (2011).

Ventilation-based segmentation of the lungs using hy-

perpolarized 3he mri. Journal of Magnetic Resonance

Imaging, 34(4):831–841.

Tustison, N. J., Qing, K., Wang, C., Altes, T. A., and Mu-

gler, J. P. (2015). Atlas-based estimation of lung and

lobar anatomy in proton mri. Magnetic resonance in

medicine.

V¨olzke, H., Alte, D., Schmidt, C. O., Radke, D., Lorbeer,

R., Friedrich, N., Aumann, N., Lau, K., Piontek, M.,

Born, G., et al. (2011). Cohort proﬁle: the study of

health in pomerania. International journal of epidemi-

ology, 40(2):294–307.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications