Intelligent algorithms for cell tracking and image segmentation

International Journal of Computer Science & Information Technology (IJCSIT) Vol 6, No 5, October 2014
INTELLIGENT ALGORITHMS FOR CELL TRACKING
AND IMAGE SEGMENTATION
Ashraf A. Aly1, Safaai Bin Deris2, and Nazar Zaki3
1Information Technology Department, Al khawarizmi College, UAE
2Faculty of computer Science and Information Systems, University Technology Malaysia
3College of Information Technology, UAE University, UAE
ABSTRACT
Sensitive and accurate cell tracking system is important to cell motility studies. Recently, researchers have
developed several methods for detecting and tracking the living cells. To improve the living cells tracking
systems performance and accuracy, we focused on developing a novel technique for image processing. The
algorithm we propose presents novel image segmentation and tracking system technique to incorporate the
advantages of both Topological Alignments and snakes for more accurate tracking approach. The results
demonstrate that the proposed algorithm achieves accurate tracking for detecting and analyzing the
mobility of the living cells. The RMSE between the manual and the computed displacement was less than
12% on average. Where the Active Contour method gave a velocity RMSE of less than 11%, improves to
less than 8% by using the novel Algorithm. We have achieved better tracking and detecting for the cells,
also the ability of the system to improve the low contrast, under and over segmentation which is the most
cell tracking challenge problems and responsible for lacking accuracy in cell tracking techniques.
KEYWORDS
Cell tracking, segmentation enhancement, Active Contour, Topological Alignments.
1. INTRODUCTION
Tracking cell mobility is an important part of many biological processes. Tissue cells of multi
cellular organisms mobilize during embryologic development, generation of new blood vessels,
cancer metastasis, and immune response. Understanding the mechanisms of cell motility is
essential part for curative and preventative treatments to many diseases. Cell tracking and
segmentation with high accuracy is important step in the cell motility research. For instance,
tracking the number and velocity of rolling leukocytes is essential to understand and successfully
treat inflammatory diseases [15]. Sensitive tracking for moving cells is important to do
mathematical modeling to cell locomotion. Moreover, Zimmer [17] modified the snake model to
track the movement of the cells and segment the first frame. Another research by Mukherjee [19]
he developed a technique to handle segmentation process and tracking problem simultaneously.
Li [20] used a technique with two stages; the first one is a tracker and a filter to detect the cell and
also the cells which move in and out of the image area. Coskun [12] used imaging data to solve
the inverse modeling problem to determine the mobility analysis of the cells. Recently, a number
of researchers have been created automated techniques to track and detect the cells mobility.
Segmentation is an essential part in many signal processing techniques and its applications.
Texture analysis is important in many areas such as image processing, determination of the object
shape, scene analysis. The process of segmentation depends on the determination of the best
positions of the points which represent the image. The purpose of image segmentation is to
partition an image into meaningful regions based on measurements taken from the image and
might be grey level, colour, texture, depth or motion. Usually the process to determine the image
DOI:10.5121/ijcsit.2014.6502 21

starts with image segmentation as initial step. The goal of image segmentation is to cluster pixels
into salient image regions, i.e., regions corresponding to individual objects, surfaces, or natural
parts of objects. In this paper we managed to introduce a novel technique for image segmentation
and cell tracking.
2. RELATED WORK
In recent years, there has been significant research efforts toward the development of automated
methods for segmentation and cell tracking for living cells as in [1][2][3][4][5]. Most of the time,
images from microscopic studies are corrupted during the recording process and due to the noise
from the electronics devices, which affect the quality of the image.
Cell tracking with good accuracy is important in microscopic imaging studies. For instance,
Image analysis of leukocytes cells is essential part for curative and preventative treatments to
many diseases and also important to understand and successfully treat inflammatory diseases as
in Ray et al. [23]. Sensitive tracking for moving cells is important to do mathematical modelling
to cell locomotion. Zimmer [17] modified the Active Contour model to detect the mobility of the
moving cells and also handle the cell division by providing an initial segmentation for the first
frame. Mukherjee et al. [19] developed an algorithm by using threshold decomposition computed
via image level sets to handle tracking problem and segmentation simultaneously. Li [20]
developed an algorithm with two levels, a motion filter and a level set tracker to handle the cell
detection and the cells that move in and out of the image. Coskun et al. [12] used imaging data to
solve the inverse modelling problem to determine the mobility analysis of the cells. Recently
there have been a number of researchers attempt to create automated algorithms to detect and
track the cells from microscopic images as in Mélange [6]; and Mignotte [10].
This paper discusses the tracking cell accuracy as important task in many biological studies to
understand the cell behaviour and the way in which cells interact with the world around them.
One of the major goals of tracking the mobility of living cells is to find the best way to increase
segmentation and tracking accuracy under weak image boundaries, over and under segmentation,
which the most cell tracking challenge problems and responsible for lacking accuracy in cell
tracking techniques.
3. ALGORITHM
Active Contour and Topological Alignments are used in image processing, particularly in
locating object boundaries. Each method has its own advantages and also limitations. Active
Contour (snakes), can locate the object boundaries dynamically and automatically from an initial
contour. The advantage of Snakes model is the ability of the model to give a linear determination
of the object shape at the convergence time, and no extra processing is needed. But Snakes model
require detecting strong image gradients to detect the contour. This actually limits the use of
Active Contour, because weak boundaries of the image frames and also frames with low contrast
will cause over and under segmentation which responsible for decreasing the accuracy of the
analysis. To mitigate the effect of this problem with the Active Contour model, and to improve
the performance of segmentation and cell tracking, we apply the Topological Alignments method
to increase the accuracy of cell tracking and detecting analysis. The Topological Alignments
method links segments between every frame and the next one; that will decrease the number of
false detections and also the false trajectories. In this paper, we introduce a novel technique based
on Active Contour in conjunction with Topological Alignments. We present tracking system and
image segmentation algorithm to incorporate the advantages of both Active Contour and
Topological Alignments to get a tracking system with high accuracy to detect and analyze the
22

mobility of the living cells. The novel technique proceeds in two steps: First we do initial
segmentation by applying the Topological Alignments and then transformed the output into the
input of the Active Contour model to begin the analysis to detect the cells boundaries and
determine the mobility of the cells.
Active Contour models (snakes) goal is to apply segmentation process to an image by doing
deformation to the initial contour towards the boundary of the object of interest. We do that by
deforming an initial contour to minimizing the energy function which defined on contours, as in
[13] [15]. We have here two components which represents the energy function; the first part is the
potential energy component, and the potential energy component is small when the contour is
aligned to the image edge, and the second part is the internal deformation energy component, and
the component is small when the contour is smooth. Both components are contour integrals with
respect to a parameter of the contour.
Active Contour can be represented by two models depends on the characteristics of the image;
edge-based models and region based models. The advantage of Active Contour model is the
ability of the model to give a linear determination of the object shape at the convergence time,
and no extra processing is needed. But Snakes model require detecting strong image gradients to
detect the contour. This actually limits the use of Active Contour, because weak boundaries of the
image frames and also frames with low contrast will cause over and under segmentation which
responsible for decreasing the accuracy of the analysis.
The Topological Alignments method based on linking the segmentation of two frames in the
video sequence as in [17] [25]. From the output of the segmentation procedure, the method finds
the maximum weighted solutions between two pairs of frames and then match the segments. The
method can deal with low contrast images and shape cells and improves the filtration efficiency.
Figure (1), represents the proposed framework of this research. Starts by loading the cell
sequence, and pre–processing stage starts to smooth the image and remove noise, segmentation
stage starts to detect the WBC outline boundaries, and segments the image boundary and then,
extracts the WBC mobility data. The following algorithm describes the general cell tracking main
steps used in this research for automated enhancement technique.
Algorithm
Step 1) Cell image pre-processing and enhancement to reduce the cell image noise.
Step 2) Cell detection; using level set Active Contour segmentation.
Step 3) Cell tracking; using Topological Alignments method, and transforms the raw images into
a file that encodes the boundaries of every cell at every time point, and then compute the motion,
and shape parameters. We implemented the algorithm in C++ under Windows XP operating
system.
The datasets in this research consists of three video sequences protocols of WBC as test data sets
for the tracking experiments. Measuring the quality of an automated cell tracking procedure
requires a ground truth annotation to compare the output of the computationally produced
tracking with the ground truth annotation.
The first test data set consists of 32 microscopic video sequences and the temporal resolution is
30 frames /s (3-sec duration) previously used in Jung et al. (1998). These sequences are captured
from experiments, and recorded via a CCD camera system (model VE-1000CD; Dage-MTI) on a
Panasonic S-VHS recorder. The video frames is recorded at a spatial resolution of 320 x 240
pixels, where the pixel-to-micron ratio is 2.47 pixels /micron horizontally and 2.34 pixels /micron
23

vertically, the leukocyte motion direction is known a priori and is from left to right hand
direction.
Figure 1. Proposed framework for cell tracking system: segmentation and tracking transforms the images of
a cell into a file that extracts the boundaries of every cell at every time point to compute the shape and
24
motion parameters.
The second data set consists of 16 leukocyte sequences previously used in Scott et al. (2001). The
video sequence is recorded at a spatial resolution of 320 x 240 pixels (where the pixel-to-micrometer
ratio is 2.47 pixels /mm in the horizontal direction and 2.34 pixels /mm in the vertical
direction) and a temporal resolution of 30 frames/s. All cells were tracked for at least 90 frames (a
period of 3 sec). The leukocyte motion direction is from right to left hand direction for the 16
microscopic video sequences.
The third data set (self-created detect) consists of a set of video sequences of 40 live cells. The
video recordings are made by attach a charge-coupled device (CCD) camera to a microscope. The
video frames were recorded at a spatial resolution of 640×480 pixels, where pixel-to-μm ratio
was 4.94 pixels/μm horizontally, and 4.68 pixels/μm vertically. The leukocyte motion direction is
from left to right hand direction. The temporal resolution was 30 frames /sec. Each video
sequence included 91 frames, more details about the (self-created detect) lab experiment to get
mobility analysis data for live cells. Standard operating procedure (S.O.P) is used for the ground
truth annotation.
3.1 Cell Image Pre-Processing and Enhancement
Image pre-processing is required to remove the image noise. Median filtering by Kasturi et al.
(1995) is used to reduce the existing high noise level in the input frames. The median filter is
an effective method that can suppress isolated noise without blurring sharp edges. Specifically,
the median filter replaces a pixel by the median of all pixels in the neighborhood.

25
3.2 Cell Segmentation and Tracking Using the Proposed Technique
The aim of cell segmentation algorithms is to partition the image into perceptually similar
regions. The goal of segmentation is to simplify and/or change the representation of an image into
something that is more meaningful easier to analyze, and to produce a binary image from the grey
scale image. In addition, segmentation process also helps to remove noise from the image. There
are a few methods to produce segmentation, such as edge and region based segmentation,
statistical classification, and thresholding.
One of the famous methods in cell segmentation is Active Contour based segmentation (Such as
level set Active Contour, Gradient Vector Flow (GVF) Active Contours), and also thresholding
methods is one of the simplest and famous method in image segmentation.
The enhanced technique is using the level set Active Contour by Lee et al. (2004), in conjunction
with Topological Alignments method by Palaniappan et al. (2009), to segment and track the cells,
and to track the cell motion and shape evolution. Quantifying cell morphology and motility
generally requires segmenting and tracking individual cells from large image sequences.
Segmentation means identifying the boundaries of meaningful objects (here: cells) in an image;
tracking means linking these objects across different time frames. In practice, a segmentation and
tracking algorithm typically produces a data file that contains a description of the boundary of
each cell at each time point, for example of the (x,y) coordinates of a finite list of points along
that boundary as shown in Figure 2.
These extracted data can be used to quantify parameters characterizing cell shape and motion,
generally without the need to go back to the original images. Most of the process is concerned
with the segmentation and tracking stage, which is very often the most difficult part of the
workflow that begins with image acquisition and ends with the analysis of quantitative results,
and get computation of shape and motion parameters (known width of the leukocyte(~ 4 m)).
Because the shape size and position information about cells is known a priori, it can incorporate
these as parameters in our energy function to provide a more robust algorithm. Active Contour
shape can be constrained to known elliptical and circular cell shapes. Active Contour size is
known because the approximate cell size from the scale of the video is known.
Figure 2. Cell image analysis processes: segmentation and tracking transforms the raw images into a file
that encodes the boundaries of every cell at every time point, and allowing computation of shape and
motion parameters (Zimmer et al., 2005).
It is importance to combine the shape and size constraints in capturing a leukocyte for tracking.
The position constraint incorporates future cell positions and provide updates for the (x, y)
position. In addition, the Active Contour re-parameterizes itself (keeping uniform sampling)
according to the sampling energy term. The assumed initial radius of the snake tracker is integral

to effective tracking. This parameter allows the Active Contour to formulate the size constraint,
which governs the relative size of the circle or ellipse shape constraint. An accurate value of the
initial radius as compared to the actual radius of the target cell realizes a more effective Active
Contour tracker.
Level set Active Contour segmentation: Level set theory, a formulation to implement Active
Contours by Lee et al. (2004), was proposed by Osher et al. (1988). Redefine the entire domain
of an image I(x, y) as a disjoint set of subsets. They represented a contour implicitly via a two
dimension continuous function  (x, y) :  defined on the image plane. The function
(x, y) is called level set function, and a particular level, usually the zero level, of (x, y) is
defined as the contour
Topological Alignments Tracking Method: level set Active Contour is used to do segmentation
to the cell frames, and Topological Alignments method is used to track the cells. The study
addresses the problem of linking segmentations of two consecutive frames in the video sequence.
Starting from the output of a conventional segmentation procedure, the study aligns pairs of
consecutive frames through assigning sets of segments in one frame to sets of segments in the
next frame. The Topological Alignments method represent the segmentation of two images from
the video sequence as m and n, index set P = {1,..., m}, and an index set Q = {1,..., n}. The
method assumes that cells move moderately between two consecutive frames.
Topological Alignments Method by Palaniappan et al. (2009) has two stages approach: The first
stage, the method apply a segmentation procedure on every frame (here, level set Active Contour
is used for cell segmentation). The second stage is the linking stage, where the topological
alignment links segments between each frame i and the next frame i + 1. Now, in order to track
one cell, the method is matching a set of segments in frame i with another set of segments in
frame i + 1. Being based on overlap of segment groups in the two frames, the method can be
considered as a topological alignment between two images.
The method links segments between each frame and the next frame, reducing the number of false
detections and false trajectories. The Topological Alignments technique achieve this through
finding maximum weighted solutions to a generalized "bipartite matching" between two
hierarchies of segments, where they derive weights from relative overlap scores of convex hulls
of sets of segments.
3.3 The Main Steps of the Proposed Enhancement Tracking Algorithm:
1. Level set Active Contour is used for cell segmentation to identify individual cells in every
single frame, then the segmentation data for each cell are stored in a file. The first step toward
extracting leukocyte shape is to find a closed contour that satisfies a leukocyte shape prior and
minimizes energy functional for image segmentation.
2. Topological Alignments method is applied to track the cells, where the data file contains a
description of the boundary of each cell at each time point, consisting for example of the (x, y)
coordinates of a finite list of points along that boundary.
a. The main point in this procedure is to assign a weight w(p, q) to matching segments sets p  P
and q  Q.
26

where segmentation of the first image into m segments with an index set P = {1,..., m},
and the segmentation of the second image into n segments with an index set Q = {1,...,
n}.
b. Measure weights based on the "relative overlap" of the convex hulls of p, q, and identify p`
 P with the convex hull of the area covered by all segments in P, i.e. A( p) :   a(x) xp ,
where a(x) denotes the area covered by segment x and X denotes the convex hull of a set X of
points in the plain.
c. Assign the relative overlap of p and q as their weight, formally defined as:
w( p, q) : A( p)  A(q) / A( p)  A(q) and assuming that cells move moderately between
two consecutive frames. Sets of segments that achieve a relative overlap close to 1 should be
considered as one cell.
d. For each p  P and q  Q, we introduce a binary variable Xp, q, where Xp, q = 1 if and only
if p = S T i and q = i .
e. Maximize over valid partitioning only, where max 
w( p ,q )X p ,q and,
27
p P ,qQ
correspondingly, overlapping subsets from Q by introducing constraints as Xp,q  Xp,,q,  1.
f. The constraint matrix C resulting from equation (3.12), C is the incidence matrix of the
bipartite graph B = (L  R, E), where L = {pp'|p, p'  P} and R = {qq' | q, q'  Q}, and E
introduces one edge for each constraint.
3. Repeat steps (a-f) until all frames of the image sequence are processed. The result as a function
of time enable estimation of the position and deformation of the corresponding cells for each
frame in the sequence and update shape and position.
4. Segmentation and tracking transforms the raw images into a file that encodes the boundaries of
every cell at every time point, directly allowing computation of shape and motion parameters.
By using Topological Alignments to track the cells, the enhanced technique addresses the
problem of linking segmentations of two consecutive frames in the video sequence. The output of
a conventional segmentation procedure is used to align pairs of consecutive frames through
assigning sets of segments in one frame to sets of segments in the next frame. The procedure
achieve this through finding maximum weighted solutions to a generalized bipartite matching
between two hierarchies of segments, and derive weights from relative overlap scores of convex
hulls of sets of segments.
These extracted data for shape and position can be used to quantify parameters characterizing cell
shape and motion. Segmentation and tracking transforms the raw images into a file that encodes
the boundaries of every cell at every time point, and allow the computation of shape and motion
of the cells.
3.4 Manual Method-Ground Truth (Self-Created Detect)
The manual method (ground truth) used to track the mobility of the cells and manually collecting
data from the recordings. Based on the Stripe Source Diffusion Technique, which developed by
G. Grimes and F. Barnes (Grimes et al., 1973), to track the cells using microscopic technique.

Manual tracking measurements were obtained by allowing an operator to observe the cells
movement; its movements were tracked on the computer monitor with the frames.
This manual method investigate the human leukocyte cells mobility changes under the effects of
radio frequency fields on the ability of human leukocyte to follow concentration gradients of
cyclic-AMP are reported. Blood from healthy adult donors is exposed in vitro to 990 MHz a
signal generator at levels of 100 V/m and magnetic flux densities of 0.26 μT, and to fields from a
NOKIA cellular phone for approximately 3 seconds intervals. Once the data are collected, and
usually calculate the chemotactic index, the initial concentration, the concentration gradient, and
the diffusion constant.
Squared grid sheet is placed on the screen so that the mobility of the cells could be tracked on the
screen as shown in Figure 3. The screen is calibrated so that the diameter of one white blood cell
(4 μm) is equal to one square on the screen. The collected data is used to calculate the mobility of
the cells, and allow the computation of shape and motion of the cells.
28
Figure 3. Squared grid sheet is placed on the monitor.
Shape Index: Shape index is an indicator for the shape change. (Koenderink and van Doorn,
1992) developed a single-value, angular measure to describe local surface topology in terms of
the principal curvatures. It is a quantitative measure of the surface shape. Where n is the surface
normal at position x, y. This shape index is defined as:
S k k 
2 arctan 2 1 k k

 1 2
(1)

k 
k
2 1

   



   

 

 

 

 

y x n
y y n
x x n x y n
H
( )
( )
( )
( )
(2)
0 4 8 12 16
20
20
16
12
8
4
μm
μm

where shape index (S) defines the shape, and the principal curvatures are the maximum curvature
k k k k H  kI  0 1 and minimal curvature 2 . 1 and 2 can be found by solving for k, where I is
the identity matrix. Where H matrix is the eigenvalues of the shape and  ( n )
,
x x

 denote the x and y components of the parenthesized vector respectively. The collected
29
( n )

( n )

 , x y
 , y x
( n )

and y y
data (video sequences) is used as a protocol to test the automated methods, and to compare the
results from the manual method (ground truth) and the automated methods.
3.5 Accuracy Performance Measure
The performance of the enhanced cell tracking technique is measured by how well the system can
track the WBC. Two methods are used to evaluate the enhanced technique.
• Percentage of frames tracked. If a computed cell center is within one cell radius of the
manually observed cell center, then we consider that frame as tracked. The percentage is
computed as the ratio of number of frames tracked to the total number of frames in the
sequence.
N
f  (3)
total
P N
where fP is the percentage of frames tracked, N is the number of frames tracked, and total N is
the total number of frames in the image sequence.
• The second method is used to measure the performance of our enhanced technique is
calculating the Root Mean Squared Error (RMSE) between the manual (ground truth) and
the computed displacement. In addition, compare the RMSE achieved with the RMSE for
the other methods (Jung et al., (1998), and also earlier observation of Ley et al., (1996)).
The RMSE (in microns) describes how accurately the tracker tracks the cell as compared
predicted (computed) to the actual (ground truth or manual) data. RMSE gives the standard
deviation of the model prediction error. A smaller value indicates better model performance. The
root mean square error (RMSE) is giving a sense of the predicted values error. Also how close the
predicted values are to the actual values. The RMSE mathematical formula is giving by:
X X
n
RMSE
n
i actual i predicted i   
 1
2
, , ( )
(4)
where Xactual is actual values and Xpredicted is predicted values, and i represent the current predictor,
and n represents the number of predictors. The combination of the percentage of frames tracked
and the RMSE yields the qualitative performance Ratings.
3.6 Validation and Benchmarks
To validation and evaluate the quality of an automated cell tracking procedure, a comparison
have been made between the manually marked data, also known as ground-truth, of three video
sequences protocols and the automated extracted data, to compare a computationally produced

tracking with the ground truth annotation. In addition, compare the experiment results with the
other research methods results such as Jung et al., (1998), Scott et al. (2001), and earlier
observation of Ley et al. (1996).
4. RESULTS AND DISCUSSION
In order to obtain validation of our approach, we have tested the algorithms in by using three
different protocols. Some results for both manual and automated tracking of the leukocytes are
given in help of a tracking system as function of time and position. The cell velocity is calculated
from the mass center across the frames. Some results for both manual and automated tracking of
the leukocytes are given in Table 1. Average movement velocity was υ = 4.2 ± 0.4 μm/min
(manual) and 4.8 ± 0.5 μm/min (automatic), consistent with earlier observations in [23]. The
RMSE between the manual and the computed displacement was less than 12% on average. The
Active Contour method gave a velocity and change shape RMSE of less than 11%, improves to
less than 7% by using the novel algorithm presented here. Our results indicate better
segmentation and more accurate tracking for detecting and analyzing the mobility of the living
cells. We have achieved better tracking and detecting the cells, also the ability of the system to
improve the low contrast, some results are shown in Figure 4, Figure 5 and Figure 10, Figure 11
show the novel velocity and the shape index for the cells. The novel image processing technique
used in the tracking system successfully address the major problems associated with tracking
cells, image with low contrast; more accurate tracking for detecting and analyzing the mobility of
the living cells.
The results show the advantages of using our novel techniques to enhance the image and the
effect on detecting the cell contours with better and accurate segmentation. The results indicate
improvement in segmentation performance by using the Topological Alignments, which leads to
improve cell detection results. Our results indicate better segmentation and more accurate
detecting and analyzing leukocytes cells, also the ability of the system to improve the low
contrast, under and over segmentation, some results shown in Figure 6, and Figure 7.
30
(a) (b)
Figure 4. Better accuracy segmentation by using the novel algorithm.(a) By using Active Contour
algorithm (b) By using the novel algorithm we get better contrast and detection.

31
(a) (b) (c)
Figure 5. Example image from grayscale images (a) Original image with noise, (b) Cell detection by using
Active Contour, (c) Better detection by using the Novel algorithm.
(a) (b) (c)
Figure 6. Example image from grayscale images (a) Over segmentation problem by using Active Contour
(b) Over segmentation problem solved, by using the Novel algorithm.
(a) (b)
Figure 7. Example image from grayscale images (a) Over segmentation problem by using Active Contour
(b) Under segmentation problem solved, by using the Novel algorithm.
Table 1. Some mobility results for both manual (ground truth) and enhancement
automated tracking of the leukocytes cells.
Cell
VelGT (μm/s) VelExp (μm/s) GT Shape.Index Exp Shape.index
1 7.0 7.4 0.8 0.83
2 3.5 3.5 0.9 0.83

32
3 2.5 2.8 0.5 0.53
4 2.1 2.4 0.73 0.8
5 3.1 4.1 0.83 0.83
6 5.6 6.4 0.5 0.56
7 6.8 6.8 0.43 0.53
8 7.5 6.5 0.83 0.93
9 1.3 1.6 0.96 1.0
10 2.4 2.9 0.56 0.56
GT Vel :GROUND TRUTH ExVel p :EXPRIMENTAL. (ENHANCED TECHNIQUE) VELOCITY
GT Shape.Index :GROUND TRUTH Exp Shape .index :EXPERIMENTAL (ENHANCED TECHNIQUE)
(CHANGE SHAPE INDICATOR)
The average movement velocity obtained is υ = 5.5 ± 0.4 μm/s (experimental) and 5.8 ± 0.5 μm/s
(ground -truth), consistent with earlier observations of Jung et al. (1998), and also earlier
observation of Ley et al. (1996). The RMSE between the manual (ground-truth) and the
computed displacement was less than 8% on average, where the RMSE observed by Jung et al.,
(1998) was 12%. The average velocity error was less than 12% improves to less than 8% by using
the enhanced algorithm (experimental) presented here, where a lower RMSE indicates a higher
accuracy. The RMSE calculated by using the values from Table 1 and the formula in Equation 4.
The experimental results indicate improvements in WBC segmentation Figure 6, Figure 7, and
more accurate tracking for detecting and analyzing the mobility of the WBC, and achieved better
tracking of the cells and solving the segmentation problems, also the ability of the system to
reduce the noise as shown in Figure 5, Figure 6, Table 1, and Figure 8(a), shows a comparison of
average velocities (per cell) between the manually (ground-truth) recorded measurements and the
automated tracker results for the tracked cells. Figure 8(b), shows the percentage of sequences
with 100% frames tracked for tracking 32 sequences for the cells.
(a) (b)
Figure 8. (a) Some results from the experiments, velocity for a single WBC. Manually (ground- truth)
and automatically computed measurements. (b) Percentage of sequences with 100% frames tracked for
tracking 32 sequences.
The enhanced image processing technique used in the tracking system successfully address the
major associated with tracking cells, image with low contrast; better tracking for detecting and
analyzing the mobility of the living cells.

Another validation test for the enhanced technique by using set of video sequences for 16 live
cells previously used in Scott et al. (2001). In order to obtain another validation of our approach,
the experiment carry out a test by using set of video sequences previously used in Scott et al.
(2001). The experimental results indicate better and accurate tracking analysis by using our
enhanced algorithm to track the mobility of WBC from video sequences as shown in Figure 9(a),
and Figure 9(b) shows the percentage of sequences with 100% frames tracked. The enhanced
algorithm increased the accuracy of the results of the velocity calculations. The average
movement velocity result were consistent with the results obtained by Scott et al. (2001) and the
RMSE according to the enhanced technique results is less than 10%, where the RMSE for Scott et
al. (2001) was less than 12%, as shown in Table 2.
33
Table 2. Some mobility results for both manual (ground truth) and automated (enhanced tracking
technique) of the living cells.
Cell No VelGT (μm/s) VelExp (μm/s)
1 7.0 7.6
2 6.5 6.5
3 5.2 5.9
4 3.5 3.4
5 5.2 5.1
6 7.5 7.8
7 6.7 6.8
8 5.6 5.5
9 4.0 4.1
10 5.8 5.9
VelGT : GROUND TRUTH VelExp :EXPRIMENTAL. (ENHANCED TECHNIQUE) VELOCITY
(a) (b)
Figure 9. (a) Some results from the experiments, Comparison of average velocities (per cell) between the
automatic (automatically computed measurements) enhanced technique and manually (ground- truth)
tracked cells. (b) Percentage of sequences with 100% frames tracked for tracking 16 sequences.
The third protocol is a set of video sequences (self-created detected) by using a microscopic
technique as video sequences for 40 living cells to be used to test and evaluate the experimental
tracking system, and used to make comparison between automatic tracking and human controlled

tracking. The 40 living cells as video sequences collected manually to be used to test and evaluate
the enhanced tracking system (i.e., experimentally determined in this research.
The automated tracker was used to compute the corresponding 40 cell positions and the cell shape
change. Manual tracking measurements were obtained by allowing an operator to observe the
cells movement, its movements were tracked on the computer monitor with the help of a tracking
system as function of time and position. The speed of a cell is calculated as the displacement of
its center of mass across frames. Results for both manual and automated tracking of the
leukocytes are given in Table 3.
34
Figure 10. Cells velocity, the WBC Figure 11. ShapeindexGT (ground
velocity-self created detect( GT Vel : truth) the WBC change shape
ground truth) indicator, self- created detect.
The root mean squared error (RMSE) between the automated computed displacements and the
manually measured displacements was less than 8% on average, where it was 12% of Jung et al.
(1998). The results indicates better mobility analysis for speed and changing shape due to the
improvement in segmentation performance by using the Topological Alignments, which leads to
improve cell tracking results as shown in Table 3.
Table 3. Some mobility results for both manual (ground truth) and automated
tracking of the leukocytes cells.
Cell GT Vel (μm/s Exp Vel (μm/s) GT Shape.Index Exp Shape.index
1 4.2 4.4 0.46 0.5
2 2.7 2.4 0.9 0.9
3 5.5 5.7 0.93 0.9
4 2.3 2.4 0.4 0.4
6 4.7 4.9 0.53 0.56
7 5.4 5.7 0.76 0.83
8 6.7 6.6 0.93 0.93
9 2.2 2.3 1.0 1.0
10 4.1 4.4 0.86 0.9
GT Vel :GROUND TRUTH Exp Vel :EXPRIMENTAL. (ENHANCED TECHNIQUE) VELOCITY
GT Shape.Index :GROUND TRUTH Exp Shape .index :EXPERIMENTAL(ENHANCED TECHNIQUE)
(CHANGE SHAPE INDICATOR)
The results from Table 3., Figure 12(a), shows the ability of the enhanced technique to segment
and track the WBC with good accuracy, based on the results in Jung et al. (1998), and also the

experiment results achieved from the manual tracking. As shown in Figure 12(a) and Figure
12(b). The comparison of average velocities (per cell) between the manual tracking and the
automated enhanced technique shows the similarity between the velocity average values for most
of the cells as shown in Figure 12(a), which indicate the high accuracy of using the enhanced
technique to track the WBC. Figure 12(b) shows the percentage of sequences with 100% frames
tracked.
35
(a) (b)
Figure 12. Some results from the experiments, Comparison of average velocities (per cell) between the
manually recorded measurements and the automated tracker results for the tracked cells. (b) Percentage of
sequences with 100% frames tracked for tracking 40 sequences.
5. CONCLUSION
Active Contour and Topological Alignments are used in image processing, particularly in
locating object boundaries. Each method has its own advantages and also limitations. Active
Contour (snakes), can locate the object boundaries dynamically and automatically from an initial
contour. The advantage of Snakes model is the ability of the model to give a linear determination
of the object shape at the convergence time, and no extra processing is needed. But Snakes model
require detecting strong image gradients to detect the contour. This actually limits the use of
Snakes, because weak boundaries of the image frames and also frames with low contrast will
cause over and under segmentation which responsible for decreasing the accuracy of the analysis.
To mitigate the effect of this problem with the Snakes model, and to improve the performance of
segmentation and cell tracking, we apply the Topological Alignments method to increase the
accuracy of cell tracking and detecting analysis. In our experiments, we compared our algorithm
with traditional snake. The results show that the algorithm can demonstrate the segmentation
accuracy under weak image boundaries, low contrast, under and over segmentation of living cells,
which the most cell tracking challenge problems and responsible for lacking accuracy in cell
tracking techniques. Our results indicate better segmentation and more accurate tracking for
detecting and analyzing the mobility of the living cells. We have achieved better tracking and
detecting for living cells, also the ability of the system to enhance the segmentation for low
contrast, under and over segmentation problem. In this paper we focused on solving the under and
over segmentation and low contrast problems, however in our future work we will continue our
research over the cell tracking by using shape descriptions and other features and consider other
problems with image segmentation such as images with high noise. Although the proposed
methods produce very robust and promising results, there are still a few aspects could be
improved. The required computation is relatively larger than traditional methods increasing the
convergence time. The fast advance of computer processor may solve this problem.

REFERENCES
[1] Jie,S.,Hao, F.,Guoping, Q.,Philip, K., and Mohammad, I.(2013) "Segmenting Overlapping Cell
Nuclei In Digital Histopathology Images", 35th Annual International Conference of the IEEE (EMBS
Osaka).
[2] Siegel, R.,Naishadham, D.,Ahmedin, J.(2013) "Cancer Statistics", 2013‖, CA Cancer J Clin, 63:11-
36
30.
[3] Anders, E.,Dufort, P.,Daniel, F., and Stephen, L.(2013) "Medical Image Processing on the GPU: Past,
Present and Future" Medical Image Analysis.
[4] Xin, Q.,Fuyong, X.,David, Foran J., and Lin Y. (2012) "Robust Segmentation of Overlapping Cells in
Histopathology Specimens Using Parallel Seed Detection and Repulsive Level Set" IEEE Trans.On
Biomed. Eng., vol.59, no.3.
[5] Lee, A.D.(2012) "Digital Pathology: Data - Intensive Frontier in Medical Imaging", Proc. of IEEE,
vol. 100, no. 4.
[6] Melange, T.,Nachtegael, M.,and Kerre, E.2011.Fuzzy Random Impulse Noise Removal From Color
Image Sequences. IEEE Trans. 20: 1023 – 1035.
[7] Danuser, G.,Meijering, E., and Smal, I.2011.Tracking in molecular bioimaging.Signal Processing
Magazine IEEE, 23(3):46-53.
[8] Melange, T., Nachtegael, M., and Kerre, E. 2011. Fuzzy Random Impulse Noise Removal From
Color Image Sequences. IEEE Trans. 20: 1023 – 1035.
[9] Danuser, G., Meijering, E., and Smal, I. 2011. Tracking in molecular bioimaging. Signal Processing
Magazine, IEEE, 23(3):46-53.
[10] Mignotte, M. 2010. A Label Field Fusion Bayesian Model and Its Penalized Maximum Rand
Estimator for Image Segmentation. IEEE Trans. 19: 1610 - 1624.
[11] Bradski, G. 000. The Open CV Library. Dr.Dobb’s Software Tools for the Professional
Programmer.
[12] Coskun, H., Li, Y., and Mackey, M.A.2007. Ameboid cell motility:A model and inverse problem,
with an application to live cell imaging data. Journal of heoretical Biology, 244(2):169–179.
[13] Kass, M., Witkin, A., and Terzopoulos, D. 1987. Snakes: Active Contour models.
InternationalJournal of Computer Vision, pages 321–331.
[14] Li, K., Miller, E., Weiss, L., Campbell, P., and Kanade, T. 2006. Online tracking of migrating and
proliferating cells imaged with phase - contrast microscopy”, Proc. of the 2006 Conf, on Computer
Vision and Pattern Recognition Workshop (CVPRW’06), pages 65–72.
[15] Mukherjee, D., Ray, N., and Acton, S., 2004. Level set analysis for leukocyte detection and
tracking. IEEE Trans Image Process, 13(4)562-72.
[16] Ray, N., Acton, S., and Ley, K. 2002. Tracking leukocytes in vivo with shape and
sizeconstrained Active Contours, IEEE Trans Med Imaging 21(10):1222–1235.
[17] Zimmer, C., Zhang, B., Dufour, A., Thebaud, S., Berlemont, V. and Marin, O. 2006. On the
Digital Trail of Mobile Cells. Signal Processing Magazine, 23(3):54-62.
[18] Miura, K., 2005. Tracking Movement in Cell Biology. Advances in Biochemical Engineering/
Biotechnology, 95:267-295.
[19] Mukherjee, D., Ray, N., and Acton, S. 2004. Level set analysis for leukocyte detection and
tracking. IEEE Trans Image Process, 13(4):562-572.
[20] Li,Y., Zheng,Y., Doermann, D., and Jaeger, S. 2008. Script-Independent Text Line Segmentation
Segmentation in Freestyle Handwritten Documents. IEEE Trans Pattern Anal Mach
Intell,30(8):1313-1329.
[21] Ersoy, I., Bunyak, F., Mackey, M., and Palaniappan, K. 2008. Cell Segmentation Using Hessian
Based Detection and Contour Evolution with Directional Derivatives. International Conference mon
Image Processing, 1804-1807.
[22] Jung, U., Norman, K., Ramos, C., Scharffetter-Kochanek, K., Beaudet, A., and Ley, K. 1998.
Transit time of leukocytes rolling through venules controls cytokine- induced inflammatory cell
recruitment in vivo. J. Clin. Invest. 102, 1526–1533.
[23] Ray, N., Acton, S., and Ley, K.2002. Tracking leukocytes in vivo with shape and size
constrained Active Contours”, IEEE Trans Med Imaging, 21(10):1222-1235.
[24] Debeir, O., Ham, P., Kiss, R., and Decaestecker C.2005. Tracking of migrating cells under phase-contrast
video microscopy with combined mean-shift processes. IEEE Trans Med Imaging,
24(6):697-711.

[25] Sacan, A., Ferhatosmanoglu, H., and Coskun, H. 2008. CellTrack: an opensource software for cell
37
tracking and motility analysis. Bioinformatics, 24(14):1647-1649.
[26] Rodrigo, M., and Leyza, B.2007.White blood cell segmentation using morphological operators and
scale – space analysis. Brazilian Symposiun on computer graphics and Image Processing, 1530-
1834/07.
[27] Scott, T., Klaus, W., and Klaus, L.2002. Automatic Tracking of Rolling Leukocytes in Vivo,
Microvascular Research, vol 63, pp. 139 – 148.

Intelligent algorithms for cell tracking and image segmentation

More Related Content

What's hot (16)

Viewers also liked (11)

Similar to Intelligent algorithms for cell tracking and image segmentation (20)

Recently uploaded (20)

Intelligent algorithms for cell tracking and image segmentation