Automatic tracking and deformation measurements of red blood cells flowing through a
microchannel with a microstenosis: the keyhole model
B. Taboada
a,b
, F.C. Monteiro
a
and R. Lima
a,b,c
*
a
ESTiG, IPB, C. Sta. Apolonia, 5301-857 Braganc a, Portugal;
b
CEFT, FEUP, R. Dr Roberto Frias, 4200-465 Porto, Portugal;
c
DEM, Universidade do Minho, Campus de Azure
´
m, 4800-058 Guimara
˜
es, Portugal
(Received 31 January 2014; accepted 20 August 2014)
This study aimed to assess the motion and its deformation index (DI) of red blood cells (RBCs) flowing through a
microchannel with a microstenosis using an image analysis-based method. For this purpose, a microchannel having a
smooth contraction was used and the images were captured by a standard high-speed microscopy system. An automatic
image-processing and analysing method was developed in a MATLAB environment to not only track the motion of RBCs
but also measure the DI along the microchannel. The keyhole model, tested in this study, proved to be a promising technique
to automatically track individual RBCs in microchannels.
Keywords: automatic tracking; keyhole model; red blood cells; deformation index; microfluidics
Introduction
Ever since the clinical significance of red blood cells
(RBCs) deformability became a possible way to diagnose
several pathologies, many methods of measuring this
phenomenon have been proposed. Some examples are the
RBC filtration (Gueguen et al. 1984), laser diffraction
ellipsometry (Shin et al. 2004), rheoscopy (Dobbe et al.
2002) and micropipette aspiration (Mokken et al. 1992).
Recently, by using a soft lithography technique, it was
possible to fabricate transparent micrometre-sized channels
to study motion (Abkarian et al. 2008; Lima et al. 2008,
2009, 2012; Fujiwara et al. 2009; Leble et al. 2011; Garcia
et al. 2012) and dynamical deformation (Yaginuma et al.
2011, 2013; Pinho, Yaginuma, et al. 2013; Faustino et al.
2014; Rodrigues et al. 2014) of cells flowing through
microchannels. However, the majority of both cell tracking
and deformation measurements are often made by using
manual methods. These methods are extremely time-
consuming and may introduce operator errors into the data.
During the last years, several researchers have been
developing different kinds of automatic particle tracking
(APT) tools either as plugin for Image J (Sbalzarini and
Koumoutsakos 2005; Smith et al. 2011; Meijering et al.
2012) or as a MATLAB module (Meijering et al. 2012;
Pinho, Gayubo, et al. 2013). Promising APT plugins for
Image J are ‘Particletracker’ (Sbalzarini and Koumoutsa-
kos 2005) and ‘SpeckleTrackerJ’ (Smith et al. 2011).
Although both plugins can perform automatic tracking in an
extremely fast way, these methods are still under
development as the tracking trajectories tend to fail,
especially at high haematocrits (Hcts). Recently, Pinho,
Gayubo, et al. (2013) developed a MATLAB module to
track automatically labelled RBCs flowing through a
microchannel. Using fluorescent labelling, several individ-
ual RBCs were able to track automatically as bright obje ct
against a darker background. However, this method still
requires some consuming time by the user to perform
tracking and is not able to measure RBCs deformability
flowing along the microchannel. Hence, it is crucial to
develop a fast automatic method that is able to not only
track individual RBCs but also measure their deformability.
In this study, we propose an automatic image analysis
technique based on the keyhole model to characterise the
motion and deformation of RBCs flowing through a
microchannel having a smooth contraction shape. In this
geometry, the mechanical properties of RBCs are under
the effect of both simple shear and extensional flows.
In this study, we propose an automatic image analysis
technique (MKHtrackCells), based on the keyhole-
tracking algorithm that describes the probable movement
of RBC (Reyes-Aldaroro et al. 2007). First, a sequence of
binary images containing segmented foreground obje cts
are obtained by pre-processing videos, and then tracks are
formed by linking the objects with common optical flow in
contiguous frames. Finally, we measure the deformation of
individual RBCs flowing through a microchannel having a
hyperbolic contraction. In this geometry, the mechanical
properties of RBCs are under the effect of a strong
extensional flow.
Optical flow segmentation is usually defined as a group
of pixels of similar intensity that are associated with
smooth and uniform motion information. Lo w-level
q 2014 Taylor & Francis
*Corresponding author. Email: ruimec@ipb.pt
Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2014
http://dx.doi.org/10.1080/21681163.2014.957868
Downloaded by [rui lima] at 11:45 28 September 2014
motion segmentation tries to group pixels with homo-
geneous motion vectors without taking any other
information apart from intensity or image gradient.
High-level motion segmentation divides the image into
regions that exhibit coherent motion and also uses other
image cues to produce image segments that correspond to
projections of real RBCs.
Our method takes the spatial atomic regions produced
by the watershed algorithm and a variational motion
estimation method and combines them into a complete
algorithm producing a re liable motion segmentatio n
framework that is used in the tracki ng step.
Materials and methods
Working fluids and microchannel geometry
The working fluid examined was composed of Dextran 40
containing , 2% of human RBCs (i.e. Hct, , 2%). The
blood used was collected from a healthy adult volunteer,
and ethylenediaminetetraacetic acid was added to the
samples to prevent coagulation. The blood samples were
then washed by centrifugation and then stored hermeti-
cally at 48C until the experiments were performed at a
temperature of , 378C.
The microchannels containing the smooth contraction
were produced in polydimethylsiloxane (PDMS) using a
standard soft-lithography technique from a SU-8 photo-
resist mould. The moulds were prepared in a clean room
facility by photo-lithography using a high-resolution
chrome mask. The geometry of the fabricated micro-
channel is shown in Figure 1. The channel has a constant
depth of 51 mm throughout the PDMS chip and the width
of the upstream and downstream channels was the same,
cf. W
1
¼ 50 mm. The minimum width in the contraction
region (W
2
)is10mm.
For the microfluidic experiments, the device contain-
ing the microchannel was placed on the stage of an
inverted microscope (IX71, Olympus, Tokyo, Japan) and
the temperature of the stage was adjusted to 378Cby
means of a thermoplate controller (Tokai Hit, Shizuoka,
Japan). The flow rate of 1 ml/min was controlled using a
syringe pump (KD Scientific Inc., Hollist on, MA, USA).
The images of the flowing RBCs were captured using a
high-speed camera (Phantom v7.1, Vision Research,
W
2
W
1
(a)
(b)
Figure 1. (a) RBCs flowing through a microchannel having a smooth contraction shape. W
1
¼ 50 mm and W
2
¼ 10 mm. (b) A schematic
of the microchannel in 3D.
B. Taboada et al.2
Downloaded by [rui lima] at 11:45 28 September 2014
Wayne, NJ, USA) with the frame interval of 8000 fps and
transferred to the computer for analysis. An illustration of
the experimental set-up is shown in Figure 2.
Image analysis
The proposed methodology for image analysis has five
major stages. First, we remove the backgr ound, noise
and some artefacts of the original movie, as a pre-
processing stage, obtaining an image only with the
RBCs. Next, we create an over-segmented image, based
on the initial magnitude gradient image, using the
watershed transform. The optical flow information of
these regions is obtained by using the variational method
proposed by Brox et al. (2004). After that, the cell
tracking links the atomic regions in contiguous frames,
according to their motion, to form the tracks by means
of a keyhole model proposed by Reyes-Aldaroro et al.
(2007). Finally, we measure the deformation index (DI)
of each RBC.
Pre-processing
At this stage, the image background is removed by
subtracting the average of all movie images from each
image. To improve the identification of RBCs, the image
contrast is adjusted by histogram expansion.
Images taken with digital cameras will pick up noise
from a variety of sources. As the watershed algorithm is
very sensitive to noise, it is desirable to apply a noise
reduction filter in the pre-processing step. Several filters
have been proposed in the literature to reduce the spurious
boundaries created due to noise. However, most of these
filters tend to blur image edges while they suppress noise.
To prevent this effect, we use the median filter. This filter
preserves image stru ctures while su ppres sing noise.
Figure 3 shows the result obtained after pre-processing
the image. For a better visualisation, we use the image
negative.
Segmentation
Image segmentation is a fundamental step in image
analysis systems, and aims to identify regions, the so-
called segm ents that have a specific meaning within
images. The regions have to be uniform with respect to
some parameter, such as image intensity, texture or
motion.
An ideal over-segmentation should be easy and fast to
obtain and should not contain too many segmented
regions, and it should have its region bound aries as a
superset of the true image region boundaries. In this
section, we present an algorithm step that groups pixels
into ‘atomic regions’. The motivations of this preliminary
grouping stage resemble the perceptual grouping task: (1)
abandoning pixels as the basic image elements, we instead
use small image regions of coherent structure to define the
optical flow patches. In fact, because the real world does
not consist of pixels, it can be argued that this is even a
more natural image representation than pixels as those are
merely a consequence of the digital image discretisation.
Watershed transform is a classical and effective
method for image segmentation in grey-scale mathemat-
ical morphology. For images, the idea of the watershed
construction is quite simple. An image is considered as a
topographic relief where for every pixel in position (x, y),
its brightness level plays the role of the z-coordinate in the
landscape. Local maxima of the activity image can be
thought of as mountain tops, and minima can be
considered as valleys.
In the flooding or immersion approach (Vincent and
Soille 1991), single-pixel holes are pierced at each
regional minimum of the activity image which is regarded
as topographic landscape. When sinking the whole surface
slowly into a lake, water leaks through the holes, rising
uniformly and globally across the image, and proceeds to
fill each catchment basin. Then, in order to avoid water
Figure 2. Experimental set-up: inverted microscope, high-
speed camera, syringe pump and thermoplate controller.
Figure 3. Up: original image. Down: pre-processed image.
Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization 3
Downloaded by [rui lima] at 11:45 28 September 2014
coming from different holes merging, virtual dams are
built at places where the water coming from two different
minima would merge.
We start by computing the image gradient magnitude
and remove the weakest edges by gradient minima
suppression (pre-flooding). Figure 4 shows the gradient
image, the image after applying a 5% of the gradient
maximum pre-flooding, and the image after applying the
watershed method.
Figure 5 shows the original RBC segmented by
watershed transform.
Optical flow
There is a strong interdependence between the definition
of the spatial support of a region and of its mot ion
estimation. On the one hand, the estimation of the motion
information of the region depends on the region of support.
Therefore, a careful segmentation of the regions is needed
in order to estimate the motion accurately. On the other
hand, a moving region is characterised by coherent motion
characteristics over its entire surface (assuming that only
rigid motion is permitted). Therefore, a precise estimation
of the motion is required in order to obtain an accurate
segmentation of the region.
All the motion estimation approaches assume that
there is a point correspondence between two consecutive
frames which induces dense motion vector field of an
image. No matter what method is used, at some sta ge we
need a mec hanism to assign each point to one of the
recovered motions. This mechanism must take into
account the smoothness of the world, i.e. the intuitive
notion that the points belonging to the same motion are
also spatially clustered in the image.
Motion information will be initially represented
through a dense motion vector field, i.e. it estimates
which one best relates the position of each pixel in
successive image frames. For the task at hand, we adopt a
high accuracy optical flow estimation based on a coarse-
to-fine warping strategy proposed by Brox et al. (2004 )
which can provide dense optical flow inform ation. This
method accelerates convergence by allowing global
motion features to be d etected immediately, but it also
improves the accuracy of flow estimation because it
provides better approximation of image gradients via
warping. This technique is implemented within a multi-
resolution framework, allowing the estimation of a wide
range of displacements.
Estimating optical flow involves the solution of a
correspondence problem. That is, which pixel in one frame
corresponds to which pixel in the other frame. In order to
find these corr espondences, one needs to define some
properties that are not affected by displacement. The
combined variational approach (Brox et al. 2004) differs
from usual variational approaches by the use of a gradient
constancy assumption. This assumption provides the
method with the capability to yield good estimation
results even in the presence of small local or global
variations of illumination.
Given two successive images of a sequence Iðx; y; tÞ
and Iðx þ u; y þ v; t þ 1Þ, we seek at each pixel x :¼
ðx; y; tÞ
T
the optical flow vector vð x Þ :¼ðu; v; 1Þ
T
that
describes the motion of the pixel at x to its new location
ðx þ u; y þ v; t þ 1Þ in the next frame.
.
Grey-value constancy assumption. It is assumed that
the grey value of a pixel is not changed by
displacement:
Iðx; y; tÞ¼Iðx þ u; y þ v; t þ 1Þ: ð1Þ
.
Gradient constancy assumption. A global change in
illumination both shifts and scales the grey values of
an image sequence. Shifting the grey values will not
affect the gradient. Although scaling the grey values
Figure 4. (a) Gradient image. (b) Result after the pre-flooding.
(c) Result after the watershed method.
Figure 5. Image with a black background and the segmented
regions with the original colour.
B. Taboada et al.4
Downloaded by [rui lima] at 11:45 28 September 2014
