Extracting Biomarkers from Dynamic Images

Approaches and Challenges

Jakub Nalepa

1,2

, Michael P. Hayball

1,3,4

, Stephen J. Brown

1,3,4

, Michal Kawulok

1,2

and Janusz Szymanek

Future Processing, Gliwice, Poland

Silesian University of Technology, Gliwice, Poland

Feedback PLC, Cambridge, U.K.

Cambridge Computed Imaging, Cambridge, U.K.

Keywords:

DCE Imaging, Biomarkers, Medical Image Analysis.

Abstract:

Imaging technologies have developed rapidly over the past decade proving to be valuable and effective tools

for diagnosis, evaluation and treatment of many conditions, especially cancer. Dynamic contrast enhanced

imaging using computed tomography or magnetic resonance has been shown particularly effective and has

been intensively studied to allow for assessing the vascular support of various tumours and other tissues. In

this paper, we discuss current approaches and most important challenges in extracting markers from such

dynamic images. These difﬁculties have to be resolved in order to ultimately improve patient care.

1 INTRODUCTION

Imaging technologies have developed extremely

rapidly over the past decade proving to be valuable

and effective tools for diagnosis, evaluation and treat-

ment of many conditions, especially cancer. Cross-

sectional imaging such as magnetic resonance (MR)

and computed tomography (CT) have been shown

particularly effective. The pixel data in these im-

ages holds signiﬁcant information which can be un-

covered by post-processing analytical software. This

is currently not widely done due to the complexity of

this process. It results in many conditions being un-

detected and miss-diagnosed leading to incorrect or

prolonged treatment, increased costs and unnecessary

emotional upset.

Therefore, there is need of early diagnostic and

predictive factors to reﬁne the management of patients

with various types of cancer, e.g., in order to differen-

tiate between benign and malignant, guide the use of

surgical adjuncts, response evaluation and help deter-

mine patients that are at risk of early reoccurrence,

requiring intense monitoring and follow-up.

Dynamic contrast enhanced (DCE) biomarkers

were proven to be correlated with physiological and

molecular processes which can be observed in tu-

mour angiogenesis (these processes are morpholog-

ically characterised by an increased number of micro-

vessels, which are extremely difﬁcult to image di-

rectly using well-known imaging techniques (Miles,

2002)). Therefore, DCE biomarkers can robustly as-

sess tumour characteristics and stage, and provide

an independent indicator of prognosis, enabling risk

stratiﬁcation for patients with various types of cancer.

The ﬁrst description of the application of multi-

slice DCE computed tomography (DCE-CT) for

measuring tumour perfusion dates back to early

90’s (Miles et al., 1993). The process of the DCE

analysis involves acquiring time series images and in-

vestigating temporal changes of injected contrast ma-

terial (very often referred to as tracer) attenuation

in vessels and tissues. Such biomarkers extracted

from dynamic images have been investigated in sev-

eral clinical trials (Coenegrachts et al., 2012). They

were also used for determining the drug effects in

various cancers (thus monitoring the therapy progress

and treatment response), which include glioma (Ford

et al., 1996), rectal, renal (Fournier et al., 2010),

lung (Ng et al., 2007) cancers, carcinoids, and numer-

ous others (Miles, 2002). Importantly, DCE imag-

ing produces reproducible measurements which have

been validated against many reference methods—this

reproducibility of results is extremely important in

clinical applications which are aimed at helping in di-

520

Nalepa, J., Hayball, M., Brown, S., Kawulok, M. and Szymanek, J.

Extracting Biomarkers from Dynamic Images - Approaches and Challenges.

DOI: 10.5220/0006652505200525

In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 520-525

ISBN: 978-989-758-276-9

agnosis. It is a non-invasive technique and can be eas-

ily incorporated into standard protocols, thus does not

require an additional imaging modality. Finally, the

linear relationship between the contrast medium con-

centration and CT attenuation value implies an easier

quantiﬁcation compared with MRI (Miles, 2002).

Quantitative imaging—according to the Quantita-

tive Imaging Biomarkers Alliance (which is the ini-

tiative of the Radiological Society of North Amer-

ica, looking into improving the practicality of quan-

titative biomarkers by reducing variability across de-

vices, patients, time, and sites

)—is the process of

extraction of quantiﬁable features from medical im-

ages (e.g., dynamic contrast enhanced images). These

image-derived features should allow for assessing the

status of a disease, injury, or chronic conditions rela-

tive to normal. The ﬁeld of quantitative imaging en-

compasses additional activities, including data anal-

ysis, along with the display and reporting methods.

Importantly, these procedures should directly lead to

extracting accurate, precise and reproducible metrics,

which can be coupled with additional, physiologically

relevant patient parameters to improve the treatment

outcome and prognostic efﬁciency of imaging.

A quantitative imaging biomarker is then an ob-

jective characteristic derived from an in vivo image.

This characteristic is measured on a ratio or interval

scale as indicators of biological/pathogenic processes

or a response to the intervention and/or treatment.

The biomarkers may be extracted from a plethora of

currently available image modalities (e.g., computed

tomography, magnetic resonance or positron emission

tomography), and coupling them may further boost

their prognostic efﬁciency. Optimising the perfor-

mance of the derived biomarkers can directly affect

the patient treatment pathway.

This paper is structured as follows. In Section 2,

we discuss the procedure of extracting biomarkers

from dynamic images. It is interleaved with the ref-

erences to the literature, which highlight the current

advances in this ﬁeld. Section 3 concludes the paper

and serves as an outlook to our future research.

2 EXTRACTING BIOMARKERS

FROM DYNAMIC IMAGES

The process of extracting biomarkers from dy-

namic images is visualised in Figure 1. The

ﬂowchart presents the standard steps in the processing

pipeline—the images (e.g., DCE-CT or DCE-MR) are

For more details on QIBA see:

http://www.rsna.org/qiba/

acquired using an appropriate acquisition protocol.

It is worth noting that novel volumetric acquisition

techniques allow for decreasing the radiation dose as

pointed out by Miles et al. (Miles et al., 2012). Then,

the acquired images are registered (in order to min-

imise the impact of e.g., the patient motion on the

DCE results) and segmented (at a minimum, an ar-

terial blood vessel must be identiﬁed to enable analy-

sis).

The DCE processing involves generating

time/density curves for volumes of interest (VOIs)

along with the parametric maps, which are later

analysed for extracting useful (quantiﬁable) metrics.

Finally, these results should be safely stored since

they will be re-investigated in the future. This

investigation will help understand how the current

treatment affects the patient condition, and will allow

for choosing the best care pathway.

Acquire

images

Segment

DCE

process

Present &

investigate

Store

Figure 1: Extracting biomarkers from dynamic images.

In the following subsections, we discuss in more

detail the pivotal steps of deriving biomarkers from

dynamic images. Although we focus on the DCE-CT

imaging, this processing strategy is quite generic and

can be easily tailored to other modalities too.

2.1 Image Registration

4D acquisition (3D image stack acquired in time,

which becomes the fourth dimension) techniques suf-

fer from motion artifacts which may be caused by ir-

regularities in patient breathing or heartbeat. These

differences between images should be corrected be-

fore any curve analysis since they are of the non-

pathological origin and can easily jeopardise the in-

vestigation (Hachama et al., 2010). This issue is tack-

led by the image registration process which maps im-

ages to a common space to correct geometric differ-

ences. On the one hand, rigid transformations al-

low for overcoming translational and rotational dis-

placements, whereas non-rigid transformations can

account for deformable changes in tissue shape.

The registration is crucial for an accurate DCE

analysis since it can inﬂuence the shape of resulting

enhancement curves (obtained either for segmented

automatically or manually annotated tissues of inter-

est). Importantly, some peaks in these curves can

be attenuated by the registration process, as shown

Extracting Biomarkers from Dynamic Images - Approaches and Challenges

521

in (Hachama et al., 2010). These peaks correspond to

the patient breathing periods which occur after peri-

ods of apnea. If these peaks were not removed during

the registration process, then they could be mistakenly

interpreted during the further curve analysis (either by

an automated algorithm or by the operator).

Incorporating advanced registration techniques to

minimise the impact of motion artifacts on time-

attenuation curve analysis is a pivotal task—it became

a vital research topic and some registration techniques

emerged during the years (Hou et al., 2014). Nonethe-

less, there is a need for fast and efﬁcient algorithms

(both rigid and non-rigid) which will help mitigate the

risk of inaccurate analysis (which, in turn, affects the

biomarker extraction process).

Speeding up the registration process may be ac-

complished by designing and implementing parallel

algorithms which may possibly run either on mod-

ern multi-core processor or graphics processing unit

architectures, which are widely available nowadays.

We already showed that co-operative parallel algo-

rithms are extremely efﬁcient in solving a variety of

complex optimisation problems (Nalepa and Błocho,

2015). Investigating the possibility of fusing the re-

sults of various registration algorithms run in parallel

(so that the execution time of this processing step is

not increased) to retrieve higher quality results is of

high research interest.

The recent advances and trends in medical image

registration are gathered in a number of interesting

surveys and reviews (Hill et al., 2001; Oliveira and

Tavares, 2014). These papers also highlight the po-

tential future research directions.

2.2 Image Segmentation

Automatic image segmentation (very often referred to

as the image labelling) algorithms identify particular

structures (or tissues) in the input medical image. It

has been recently shown that this step is crucial for

the 4-dimensional CT (3D stack of images along the

temporal dimension) and can signiﬁcantly affect the

image-guided therapy (Martin et al., 2015). This issue

is especially challenging for target volumes which are

subject to motion (e.g., the lung cancers are subject to

varying magnitudes of respiratory motion). Discrep-

ancies (in the annotated region’s shape and/or area)

between the automatic and manual segmentation (or

even the manual segmentation undertaken by two or

more experienced radiology experts) may lead to dif-

ferences in treatment planning. Therefore, there is

a need for accurate (and robust against low-quality

data) 4D image segmentation techniques.

In recent segmentation algorithms, images are

Figure 2: Example CT images containing lungs of various

shapes and characteristics.

very often denoised at ﬁrst, and then the tissue of in-

terest is manually outlined to provide a seed point for

further analysis (Hou et al., 2014). Then, the image

is segmented into the foreground (tissue) and back-

ground (usually air) parts, and undergoes the mor-

phological processing (although images can provide

a lot of information, only some—usually one or two

within the image—tissue areas are of interest). The

3D spatial and temporal domains are exploited with

the shape encoding, propagation and reﬁnement tech-

niques. Since the consecutive images most probably

share the similarities between tissues and their shapes,

these similarities can be used to increase the qual-

ity of the initial segmentation (the third dimension

is utilised). The contour information (which may be

represented e.g., as the control points of the B-spline

corresponding to a given image segment) of one im-

age is therefore used for reﬁning the control points in

the following image. Although it is a relatively easy

way of propagating the information between frames,

this approach helps boost the segmentation accuracy.

Such propagation is commonly used in segmentation

techniques which analyse the neighbouring frames in

stacks of images.

Segmenting medical images has been summarised

in numerous interesting surveys (Vitulano et al., 1997;

Wei and Li, 2014). A very recent paper by Smistad

et al. (Smistad et al., 2015) indicates that exploiting

parallel architectures (which are easily accessible at

the moment) is an important issue. Combining the in-

formation reﬂecting various image features, including

intensity, texture and spatial characteristics appears as

a very promising research direction. This combined

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

522

feature domain can be utilised for new 2D segmen-

tation (for individual images) techniques, and can be

coupled with the 3D spatial and temporal image re-

lations to provide accurate segmentation into organs

(i.e., to classify the organs in the input images prop-

erly). Also, weakly supervised segmentation tech-

niques are gaining attention too (Jia et al., 2017).

Four example images containing lungs of various

characteristics, positions and shapes are given in Fig-

ure 2, along with the example volumetric rendering of

segmented lungs presented in Figure 3. These exam-

ples show how difﬁcult the segmentation task is—a

well-performing algorithm should deal well with col-

lapsed or malformed lung structures. Also, it is pos-

sible that a tumour appears on the lung edge, hence

it inﬂuences its shape. These images come from

the Cancer Imaging Archive

repository (Clark et al.,

2013), and it was used during the 2015 SPIE Med-

ical Imaging Conference (the LUNGx Segmentation

Challenge (Armato et al., 2015)). Organs are very

often segmented from whole-body scans which im-

poses the additional difﬁculty of locating the frames

containing these organs in the ﬁrst step (other frames

can be pruned and removed from further analysis).

Figure 3: Lungs segmented from a whole-body CT scan us-

ing our lung segmentation algorithm (Walczak et al., 2017).

See https://wiki.cancerimagingarchive.net; Last ac-

cess: November 11, 2017.

2.3 DCE Processing

DCE biomarkers are derived from analysing the

movement of contrast dye into and out of soft tis-

sue in the human body following injection of a bolus

of dye. The semi-quantitative parameters which can

be derived from DCE imaging with the use of time-

attenuation curves reﬂect and quantify tumour vascu-

larity (O’Connor et al., 2011). These measures en-

compass (but are not limited to):

• Peak enhancement—the maximum contrast con-

centration.

• Perfusion normalised to cardiac output—the peak

enhancement corrected for dose of contrast

medium and contrast sensitivity.

• Standardised perfusion value—the peak enhance-

ment corrected for the patient weight and iodine

calibration factor.

• Area under the time-attenuation curve.

• Maximum upslope.

Derived values (quantifying physiological param-

eters) also require the time-attenuation curve from the

supplying artery, referred to as the arterial input func-

tion (AIF). These additional DCE parameters include:

• Perfusion (regional tumour blood ﬂow)—the

amount of blood (contrast) ﬂowing through the

unit volume of tissue.

• Permeability—the amount of contrast leaving the

blood stream.

• Blood volume—the fraction of the tissue which is

blood.

• Mean transit time—the time indicating how long

the contrast takes to pass through the given tissue

vasculature.

• Time to peak—the time taken to reach the maxi-

mum contrast concentration.

The extracted features are depicted as paramet-

ric maps (for better visualisation and easier interpre-

tation). Although comparing DCE results for var-

ious images obtained using software delivered by

one manufacturer (e.g., GE Healthcare or Siemens

Healthcare) is quite intuitive, such comparisons are

extremely challenging (or even impossible) if differ-

ent DCE analysis approaches are exploited.

As discussed by Miles et al. (Miles et al., 2012),

there are three tracer kinetics models currently in use:

the Fick’s principle (implemented by e.g., Philips and

Toshiba), the Patlak model (Siemens), and the John-

son and Wilson model (GE Healthcare). Because of

differences in the underlying physics of these mod-

els, the results cannot be compared directly. This is

Extracting Biomarkers from Dynamic Images - Approaches and Challenges

523

an important real-life issue—it is not possible to com-

pare DCE results retrieved for the same patient e.g., at

different sites.

3 CONCLUSIONS AND

OUTLOOK

Although DCE is widely used and accepted in clin-

ical practice, there are a number of technical is-

sues (including the interoperability of medical sys-

tems (Chmielewski and Stapor, 2016)) which remain

unresolved and should be tackled to provide robust

and efﬁcient personalised medical care. The guide-

lines and recommendations (concerning the system

requirements, quality assurance, radiation dosimetry,

patient preparation and many others which may in-

ﬂuence the examination) which help optimise the use

of the DCE imaging in a day-to-day oncology care,

along with the current status of DCE have been sum-

marised in a survey by Miles et al. (Miles et al., 2012).

DCE biomarkers have been available for some 25

years, but the improvements in acquisition technol-

ogy have accelerated the adoption recently. Modern

CT and MR scanners are able to acquire time series

images of 3D volumes routinely and most equipment

manufacturers offer a DCE analysis package. How-

ever, though a number of analysis algorithms have

been already published, each manufacturer applies

different analysis techniques leading to difﬁculties in

comparing results. Therefore, there is a need for a

more generally available analysis system for clinical

and research use. This issue was raised by Miles et

al. as the most important research direction, which

should be undertaken as fast as possible.

Our current research is focused on incorporat-

ing machine learning and evolutionary (e.g., memetic

(Nalepa and Blocho, 2016), being the hybrid ap-

proaches coupling evolutionary algorithms with var-

ious local-search procedures) approaches at various

steps of the DCE processing pipeline. We plan to in-

corporate support vector machines, which are a su-

pervised classiﬁer applied successfully in a range of

pattern recognition tasks (Nalepa and Kawulok, 2014;

Nalepa and Kawulok, 2016). Interestingly, segment-

ing medical images resembles the problem of de-

tecting and segmenting skin in colour images, which

was the topic of our earlier research (Kawulok et al.,

2014a; Kawulok et al., 2014b). The results obtained

in our previous works on skin detection can be ben-

eﬁcial to implement in emerging medical image seg-

mentation techniques. Finally, deep neural networks

(DNNs) are being intensively developed to segment

various image modalities (Liskowski and Krawiec,

2016). Such classiﬁcation engines can be provided

with annotated images and are aimed at extracting

features automatically. Since the medical data sets are

very often extremely imbalanced (the majority class

examples represent healthy tissue), the data augmen-

tation step is critical and was shown to dramatically

affect the performance od the deep convolutional lay-

ers. Also, determining the appropriate deep architec-

ture is an important problem which attracted research

attention (Lorenzo et al., 2017). Tailoring the DNN

architecture can allow for exploiting the additional

knowledge about the medical data (e.g., the 3D rela-

tionships) which has not been extensively investigated

in the literature so far.

ACKNOWLEDGEMENTS

This research was supported by the National

Centre for Research and Development under the

Innomed Research and Development Grant No.

POIR.01.02.00-00-0030/15.

REFERENCES

Armato, III, S. G., Hadjiiski, L., Tourassi, G. D., Drukker,

K., Giger, M. L., Li, F., Redmond, G., Farahani, K.,

Kirby, J. S., and Clarke, L. P. (2015). Guest editorial:

LUNGx challenge for computerized lung nodule clas-

siﬁcation: reﬂections and lessons learned. Journal of

Medical Imaging, 2(2):020103.

Chmielewski, M. and Stapor, P. (2016). Medical data uni-

ﬁcation using ontology-based semantic model struc-

tural analysis. In Swiatek, J., Borzemski, L., Grzech,

A., and Wilimowska, Z., editors, Information Systems

Architecture and Technology: Proceedings of 36th In-

ternational Conference on Information Systems Archi-

tecture and Technology – ISAT 2015 – Part III, pages

139–151, Cham. Springer International Publishing.

Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J.,

Koppel, P., Moore, S., Phillips, S., Mafﬁtt, D., Pringle,

M., Tarbox, L., and Prior, F. (2013). The cancer imag-

ing archive (tcia): Maintaining and operating a public

information repository. Journal of Digital Imaging,

26(6):1045–1057.

Coenegrachts, K., Bols, A., Haspeslagh, M., and Rigauts,

H. (2012). Prediction and monitoring of treatment

effect using T1-weighted dynamic contrast-enhanced

magnetic resonance imaging in colorectal liver metas-

tases: Potential of whole tumour ROI and selec-

tive ROI analysis. European Journal of Radiology,

81(12):3870 – 3876.

Ford, J., Miles, K. A., Hayball, M. P., Bearcroft, P., Blee-

han, N., and Osborn, C. (1996). A simpliﬁed method

for measurement of blood-brain barrier permeability

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

524

using CT: preliminary results and the effect of RMP-

7. Quantitative imaging in oncology, 1:1–5.

Fournier, L. S., Oudard, S., Thiam, R., Trinquart, L., Banu,

E., Medioni, J., Balvay, D., Chatellier, G., Frija, G.,

and Cuenod, C. A. (2010). Metastatic renal carci-

noma: Evaluation of antiangiogenic therapy with dy-

namic contrast-enhanced CT. Radiology, 256(2):511–

518.

Hachama, M., Desolneux, A., Cuenod, C. A., and Richard,

F. J. (2010). A classifying registration technique for

the estimation of enhancement curves of DCE-CT

scan sequences. Medical Image Analysis, 14(2):185

– 194.

Hill, D. L. G., Batchelor, P. G., Holden, M., and Hawkes,

D. J. (2001). Medical image registration. Physics in

Medicine and Biology, 46(3):R1.

Hou, Z., Wang, Y., Thng, C. H., Ng, Q.-S., Goh, V.,

and Koh, T. S. (2014). Automatic region-of-interest

segmentation and registration of dynamic contrast-

enhanced images of colorectal tumors. Physics in

Medicine and Biology, 59(23):7361.

Jia, Z., Huang, X., Chang, E. I. C., and Xu, Y. (2017).

Constrained deep weak supervision for histopathol-

ogy image segmentation. IEEE Transactions on Med-

ical Imaging, 36(11):2376–2388.

Kawulok, M., Kawulok, J., and Nalepa, J. (2014a).

Spatial-based skin detection using discriminative

skin-presence features. Pattern Recognition Letters,

41:3 – 13.

Kawulok, M., Kawulok, J., Nalepa, J., and Smołka, B.

(2014b). Self-adaptive algorithm for segmenting skin

regions. EURASIP Journal on Advances in Signal

Processing, 2014(1):170.

Liskowski, P. and Krawiec, K. (2016). Segmenting retinal

blood vessels with deep neural networks. IEEE Trans-

actions on Medical Imaging, 35(11):2369–2380.

Lorenzo, P. R., Nalepa, J., Kawulok, M., Ramos, L. S., and

Pastor, J. R. (2017). Particle swarm optimization for

hyper-parameter selection in deep neural networks. In

Proceedings of the Genetic and Evolutionary Compu-

tation Conference, GECCO ’17, pages 481–488, New

York, NY, USA. ACM.

Martin, S., Johnson, C., Brophy, M., Palma, D. A., Bar-

ron, J. L., Beauchemin, S. S., Louie, A. V., Yu,

E., Yaremko, B., Ahmad, B., Rodrigues, G. B., and

Gaede, S. (2015). Impact of target volume segmen-

tation accuracy and variability on treatment planning

for 4d-ct-based non-small cell lung cancer radiother-

apy. Acta Oncologica, 54(3):322–332.

Miles, K. A. (2002). Functional computed tomography in

oncology. European Journal of Cancer, 38(16):2079

– 2084.

Miles, K. A., Hayball, M. P., and Dixon, A. K. (1993).

Functional images of hepatic perfusion obtained with

dynamic CT. Radiology, 188(2):405–411.

Miles, K. A., Lee, T.-Y., Goh, V., Klotz, E., Cuenod, C.,

Bisdas, S., Groves, A. M., Hayball, M. P., Alonzi,

R., and Brunner, T. (2012). Current status and guide-

lines for the assessment of tumour vascular support

with dynamic contrast-enhanced computed tomogra-

phy. European Radiology, 22(7):1430–1441.

Nalepa, J. and Błocho, M. (2015). Co-operation in the par-

allel memetic algorithm. International Journal of Par-

allel Programming, 43(5):812–839.

Nalepa, J. and Blocho, M. (2016). Adaptive memetic al-

gorithm for minimizing distance in the vehicle rout-

ing problem with time windows. Soft Computing,

20(6):2309–2327.

Nalepa, J. and Kawulok, M. (2014). A memetic algorithm

to select training data for support vector machines. In

Proc. of the 2014 Annual Conference on Genetic and

Evolutionary Computation, GECCO ’14, pages 573–

580, New York, USA. ACM.

Nalepa, J. and Kawulok, M. (2016). Adaptive memetic

algorithm enhanced with data geometry analysis to

select training data for SVMs. Neurocomputing,

185:113 – 132.

Ng, Q.-S., Goh, V., Carnell, D., Meer, K., Padhani, A. R.,

Saunders, M. I., and Hoskin, P. J. (2007). Tumor

antivascular effects of radiotherapy combined with

combretastatin A4 phosphate in human non-small-cell

lung cancer. International Journal of Radiation On-

cology, 67(5):1375 – 1380.

O’Connor, J. P. B., Tofts, P. S., Miles, K. A., Parkes,

L. M., Thompson, G., and Jackson, A. (2011). Dy-

namic contrast-enhanced imaging techniques: Ct and

mri. The British Journal of Radiology, 84(spe-

cial issue 2):S112–S120. PMID: 22433822.

Oliveira, F. P. and Tavares, J. M. R. (2014). Medical image

registration: a review. Computer Methods in Biome-

chanics and Biomedical Engineering, 17(2):73–93.

PMID: 22435355.

Smistad, E., Falch, T. L., Bozorgi, M., Elster, A. C., and

Lindseth, F. (2015). Medical image segmentation on

GPUs a comprehensive review. Medical Image Anal-

ysis, 20(1):1 – 18.

Vitulano, S., Ruberto, C. D., and Nappi, M. (1997). Differ-

ent methods to segment biomedical images. Pattern

Recognition Letters, 18(11):1125 – 1131.

Walczak, M., Burda, I., Nalepa, J., and Kawulok, M.

(2017). Segmenting lungs from whole-body ct

scans. In Kozielski, S., Mrozek, D., Kasprowski,

P., Małysiak-Mrozek, B., and Kostrzewa, D., edi-

tors, Beyond Databases, Architectures and Structures.

Towards Efﬁcient Solutions for Data Analysis and

Knowledge Representation: 13th International Con-

ference, BDAS 2017, Ustro

n, Poland, May 30 - June 2,

2017, Proceedings, pages 403–414, Cham. Springer

International Publishing.

Wei, J. and Li, G. (2014). Automated lung segmenta-

tion and image quality assessment for clinical 3-d/4-d-

computed tomography. IEEE Journal of Translational

Engineering in Health and Medicine, 2:1–10.

Extracting Biomarkers from Dynamic Images - Approaches and Challenges

525