2010 International Conference on Pattern Recognition

Cell Tracking in Video Microscopy using Bipartite Graph Matching Ananda S. Chowdhury1, Rohit Chatterjee1, Mayukh Ghosh1, Nilanjan Ray2 {[email protected], falcon[email protected], [email protected], [email protected]} 1


Department of Electronics & Telecom. Engg. Jadavpur University Calcutta, India 700032.

Department of Computing Science University of Alberta Edmonton, Alberta, Canada T6G2E8.


tracking becomes very challenging due to the existence of poor contrast, high noise levels and a large possibility of cells touching each-other. Various computer vision-based approaches have been undertaken in the past decade for automated cell tracking. For example, see a sequential Bayesian framework by Ray et al. [3], a fast geometric active contour-based tracker by Li et al. [4], and an adaptive kernel-based tracking by Xie et al. [5]. A recent survey article on computerized tracking in cell biology can be found in the paper by Meijering et al. [6]. For cell tracking using graph matching, see the works of Sbalzarini and Koumoutsakos [7] and Liu et. al. [8]. Mosig et.al. has shown in [9] how bipartite matching can be applied for proper segmentation of cells in two consecutive frames. In an approach, different from [79], we combine bipartite graph matching with some properties of Gaussian distributions to propose two different methods for automated cell tracking. Geometrical centers of the cells, i.e., the centroids form the vertices of the bipartite graph. Distances between the cells in two frames and the cell velocities are used to form the weight matrix. In Method-I, the fast moving cells are identified using Gaussian distribution of Chebyshev distance between the cells in two frames and are excluded from the matching process. In Method-II, the false matches are removed, based on extreme values in the Gaussian distribution of the matched scores after the matching is performed. Two principal advantages of the proposed methods are computational efficiency and scalability. Both the methods are found to outperform i) k-NN [12] and ii) simple (with only distance information and no filtering) bipartite matching [13]. Method-II is found to yield higher accuracy than method-I.

Automated visual tracking of cells from video microscopy has many important biomedical applications. In this paper, we model the problem of cell tracking over pairs of video microscopy image frames as a minimum weight matching problem in bipartite graphs. The bipartite matching essentially establishes one-to-one correspondences between the cells in different frames. A key advantage of using bipartite matching is the inherent scalability, which arises from its polynomial time-complexity. We propose two different tracking methods based on bipartite graph matching and properties of Gaussian distributions. In both the methods, i) the centers of the cells appearing in two frames are treated as vertices of a bipartite graph and ii) the weight matrix contains information about distance between the cells (in two frames) and cell velocity. In the first method, we identify fast-moving cells based on distance and filter them out using Gaussian distributions before the matching is applied. In the second method, we remove false matches using Gaussian distributions after the bipartite graph matching is employed. Experimental results indicate that both the methods are promising while the second method has higher accuracy. Keywords: Video Microscopy, Cell Tracking, Bipartite Graph, Minimum Weight Matching.

1. Introduction Live-cell imaging experiments provide the scope of computer vision-based tracking, identification and analysis of cells in video microscopy [1]. Robust cell tracking in video microscopy has many important biomedical applications. Ianzini and Mackey, for example, developed a large scale digital cell analysis system to study the behavior of cells exposed to various stresses [2]. This problem of automated cell 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.601

2. Methods for Cell Tracking Different steps of method-I and method-II are 2448 2460 2456

shown below in figure 1:

between a cell i in the first frame and a cell j in the second frame is computed. This is given by [10]:

Image pre-proc. (2.1)

Pre-filtering (2.2)

Image pre-proc. (2.1)

d C (i, j ) = max( xi − x j , yi − y j )

For each cell i, we find out the minimum Chebyshev distance d Cmin . We observe from the plot of the minimum Chebyshev distances that they are Gaussian distributed i.e., d Cmin ~ N(μ, σ). Based on the properties of Gaussian distribution, we remove the ith cell in the first frame and the minimum Chebyshev distance yielding cell in the second frame if the following condition holds [12]: d Cmin (i ) > ( μ + 2σ ) (4)

Bipartite Matching (2.3)

Bipartite Matching (2.3)

Post-filtering (2.4)

Fig. 1. Flowchart for method-I (left column) and method-II (right column)

2.1. Image pre-processing Image pre-processing is required to properly identify the centroids of the cells, to be used as vertices of the bipartite graph. The pre-processing steps are: (a) Noise Removal: Median filtering [10] is used to reduce the existing high noise level in the input frames. (b) Entropy Thresholding: Entropy thresholding [11] is employed to segment the cells from the background. For an image, the entropy Sc for each grayscale class c (consisting of several grayscale values) can be computed using the grayscale histogram as follows: S c = ∑ p(k ) log 2 p(k ) (1)

2.3. Bipartite Matching A bipartite graph G = G (V1 ∪ V2 , E ) is constructed where cell centers in two frames form the disjoint vertex sets V1 and V2. Note that |V1| and |V2| are less in method-I as compared to that in method-II due to the pre-filtering of fast moving cells. Euclidean distance between a cell center i ∈ V1 (with coordinates (xi, yi)) and a cell center j ∈ V2 (with coordinates (xj, yj)) is given by: d E (i, j ) = ( xi − x j ) 2 + ( yi − y j ) 2

where p(k) is the probability of a pixel having a grayscale value k and Gc is the set of grayscale values for class c. In the context of binarization, c = 2 and the grayscale threshold T is chosen such that the total entropy S = S1 + S2 is maximized. (c) Connected Component Labeling: Connected component labeling [10] is employed to identify the individual cells. An 8-neighbour is used in this case. (d) Centroid Extraction: For each cell, the centroid (xc, yc) is extracted using the following equation: xc =

∑ xi i =1


For simple bipartite matching (used in the comparative study), we choose the value of weight wij of an edge eij ∈ E to be equal to dE(i, j). Since the cells under

k ∈G c



consideration have a certain velocity, which is equivalent to displacement over time (of separation between the frames), we modify the weight matrix elements by adding a mean displacement component. Let dx and dy be the average x-displacement and the average y-displacement of the cells. They can be estimated using the following equations: V2


∑ yi

dx =

i =1

, yc = (2) n n where xi and yi are pixel coordinates of the ith cell and n is the no. of pixels in that particular cell.

j =1




∑ xj −

∑ xi i =1


, dy =


∑ y j ∑ yi j =1 V2

i =1



While estimating dx and dy in method II, contributions from cells with very high match scores in simple bipartite matching are ignored. However, unlike method-I, no fast moving cells are pre-filtered out from the bipartite matching process. From equation (6), dx and dy can be considered as x-displacement and y-displacement of the center-of-mass of the first frame. The modified weight zij between a cell center i ∈ V1 and a cell center j ∈ V2 is given by:

2.2. Pre-filtering of Fast Moving Cells Some fast moving cells are identified in the following manner and are excluded before the bipartite matching is applied. A cell i in the first frame is deemed to be fast moving if the minimum Chebyshev distance between i and a cell in the second frame is very high. Use of Chebyshev distance gives better results because motion of the cells in one direction is found to be more dominant. Chebyshev distance

zij = ( x j − ( xi + d x )) 2 + ( y j − ( yi + d y )) 2

2457 2461 2449


The cell correspondences are obtained by applying minimum weight bipartite matching algorithm, whose time-complexity is O(n3), where n = max(|V1|,|V2|) [13].

ms for tracking 50 cells and about 7s for tracking 500 cells on a 2.2 GHz Pentium machine. The aforementioned values clearly demonstrate the computational efficiency and the scalability of the proposed solution.

2.4. Post-filtering of False Matches

Table 1. Percentage Cell Tracking Accuracy of four different methods for ten datasets

To remove incorrect correspondences, a postfiltering scheme is employed on the basis of edge weights of the matched cells. It is observed from the plot of the edge weights that they follow a Gaussian distribution i.e., zij ~ N(μ1,σ1). We remove the cell pair (i, j) from the match set, if the following condition is satisfied: zij > ( μ1 + 2σ 1 ) (8)

3. Experimental Results We have experimented with 10 pairs of frames of human monocytes observed from an in vitro assay, where the cells are rolling on human P-selectin. For each dataset, we show the results of tracking using four different algorithms, namely, a) k-NN, b) bipartite matching with simple weight matrix (using only distance information) and no filtering, c) Method-I and d) Method-II. Table 1 demonstrates the accuracy of the four methods for all ten datasets. Figure 2 graphically portrays the same. By accuracy, we mean the percentage of total number of cells that has been tracked correctly with respect to the reference, i.e., the manual tracking. We have used k=1 in the k-NN algorithm. Bipartite matching with simple weight matrix yields an improvement of about 26% from that of k-NN algorithm. Method-I yields a slightly better accuracy compared to the simple bipartite matching. Method-II gives best results among all the four methods used with about 14% improvement over method-I. Figure 3 shows the images of cell tracking using the four methods. The problem of cell tracking can be considered as a combinatorial pattern matching problem. Note that the k-NN algorithm behaves in a greedy manner by finding matches for each individual cell separately. In sharp contrast, bipartite matching takes into consideration all cells, while establishing one-to-one correspondences. Therefore, bipartite matching gives a much higher accuracy over k-NN. The use of velocity information in the weight matrix and pre-filtering of fast moving cells improve the performance of method-I compared to that of simple bipartite matching. However, it is imperative that method-I removes some (mostly fast moving) cells and lacks a post-filtering step. Method-II addresses both these issues and yields maximum accuracy. We use Matlab for implementing the bipartite matching algorithm. The execution time is approximately 100



1 2 3 4 5 6 7 8 9 10 mean s.d.

65.22 57.45 60.00 53.57 60.34 65.52 60.38 54.10 68.52 66.67 61.18 4.94

Simple Weight Matrix 71.74 68.09 80.00 71.43 72.41 82.76 71.70 77.05 88.89 91.11 77.52 7.53

Method I

Method II

84.78 82.98 86.00 75.00 81.03 82.76 81.13 73.77 72.22 86.67 80.63 4.92

95.65 93.62 98.00 85.71 87.93 93.10 86.79 93.44 92.59 91.11 91.80 3.73

Fig.2. Cell tracking accuracy using four different methods for ten datasets.

4. Conclusion In this paper, we propose two methods for tracking of human monocytes in a pair of frames, separated in time. Both the methods employ minimum weight bipartite graph matching and some properties of Gaussian distributions. Two principal advantages of the proposed methods are computational efficiency and scalability, which mainly arises from the polynomial time-complexity of the bipartite matching algorithm. As the centroids of the cells are used as vertices in the bipartite graph, the proposed methods are expected to work well even for cells with more complicated shapes. 2458 2462 2450

In future, we will work towards a synergistic combination of the two methods to further improve the tracking accuracy. In addition, we plan to make our tracking framework more robust by handling nonuniformities in the velocities of the various cells.

[6] E. Meijering, O. Dzyubachyk, I. Smal and W.A. van Cappellen. Tracking in cell and developmental biology. Seminars in Cell & Developmental Biology, 20(8): 894-902, 2009. [7] I.F. Sbalzarini and P. Koumoutsakos. Feature point tracking and trajectory analysis for video imaging in cell biology. Journal of Structural Biology, 151: 182-195, 2005. [8] M. Liu, A.K. Roy-Chowdhury and G. V. Reddy. Robust estimation of stem cell lineages using local graph matching. Proc. IEEE Conf. Computer Vision Pattern Recognition (CVPR) Workshops, 194-201, Miami, FL, USA, 2009. [9] A. Mosig et. al. Tracking cells in Life Cell Imaging Videos usng topological alignments. Algorithms for Molecular Biology, 4: 10, 2009. [10] R. Kasturi, R. Jain and B. Schunk. Machine Vision. McGraw Hill, NY, USA, 1995. [11] P.K. Sahoo, S. Soltani, K.C. Wong and Y.C. Chen. A Survey of Thresholding Techniques, Computer Vision, Graphics, and Image Processing 41:233-260, 1988. [12] E. Gose, R. Johnsonbaugh and S. Jost. Pattern Recognition and Image Analysis. Prentice Hall of India, 1997. [13] C.H. Papadimitrou and K. Steiglitz. Combinatorial Optimization: Algorithms and Complexity. Prentice Hall of India, 1982.

References [1] Q. Wu, F.A. Merchant and K.R. Castleman. Microscope image processing. Burlington, MA: Elsevier Academic Press; 2008. [2] F. Ianzini and M.A. Mackey. Development of the large scale digital cell analysis system. Radiation Protection and Dosimetry, 99:289-293, 2002. [3] N. Ray, G. Dong and S.T. Acton. Tracking multiple cells by correspondence resolution in a sequential Bayesian framework. Proc. IEEE Conference on Image Processing (ICIP), 705-708, Genoa, Italy, 2005. [4] K. Li, E.D. Miller, M. Chen, T. Kanade, L.E. Weiss and P.G. Campbell. Cell population tracking and lineage construction with spatiotemporal context. Medical Image Analysis, 12(5): 546-556, 2008. [5] J. Xie, S. Khan and M. Shah. Automatic tracking of Escherichia coli in phase-contrast microscopy video. IEEE Transactions on Biomedical Engineering, 56: 390-399, 2009.







Fig.3 Cell Tracking using Different Methods. (a) Cells in frame-1, (b) Cells tracked in frame-2 manually, (c) Cells tracked in frame-2 using k-NN, (d) Cells tracked in frame-2 using Simple Bipartite Matching, (e) Cells tracked in frame-2 using method-I, (f) Cells tracked in frame-2 using method-II.

2459 2463 2451

Cell Tracking in Video Microscopy Using Bipartite Graph ... - IEEE Xplore

Automated visual tracking of cells from video microscopy has many important biomedical applications. In this paper, we model the problem of cell tracking over pairs of video microscopy image frames as a minimum weight matching problem in bipartite graphs. The bipartite matching essentially establishes one-to-one ...

1MB Sizes 0 Downloads 315 Views

Recommend Documents

Cell Tracking in Video Microscopy Using Bipartite ...
Automated visual tracking of cells from video microscopy has many important biomedical applications. In this paper, we model the problem of cell tracking over ...

Video Stabilization and Completion Using Two Cameras - IEEE Xplore
Abstract—Video stabilization is important in many application fields, such as visual surveillance. Video stabilization and com- pletion based on a single camera ...

Tangible Video Teleconference System Using Real ... - IEEE Xplore
Jul 15, 2009 - Abstract — This paper deals with a real-time image based relighting system for tangible video teleconference. The proposed image based relighting system renders the extracted human object using the virtual environmental images. The p

Improved Hand Tracking System - IEEE Xplore
May 1, 2012 - training time by a factor of at least 1440 compared to the ... Taiwan University of Science and Technology, Taipei 106, Taiwan (e-mail:.

Speckle Tracking in 3D Echocardiography with Motion ... - IEEE Xplore
tracking in 3D echocardiography. Instead of tracking each speckle independently, we enforce a motion coherence con- straint, in conjunction with a dynamic ...

Fundamentals of Inter-Cell Overhead Signaling in ... - IEEE Xplore
Index Terms—Heterogeneous cellular network, inter-cell coor- dination ...... Overhead outage p versus average packet service rate =B in the three scenarios.

cooperative target tracking in vehicular sensor networks - IEEE Xplore
This subject has been studied in fields such as airborne traffic, computer vision, and wireless sensor networks. A VANET brings out new challenges that should ...

Robust Multitask Multiview Tracking in Videos - IEEE Xplore
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS. 1. Robust Multitask Multiview Tracking in Videos. Xue Mei*, Senior Member, ...

mpeg-2 video encoder chip - IEEE Xplore
Thanks to increased market accep- tance of applications such as digital versatile disks (DVDs), HDTV, and digital satellite broadcasting, the MPEG-2 (Moving ...

DISCOV: A Framework for Discovering Objects in Video - IEEE Xplore
ance model exploits the consistency of object parts in appearance across frames. We use maximally stable extremal regions as obser- vations in the model and ...

Maximizing user utility in video streaming applications - IEEE Xplore

Multi-View Video Summarization Using Bipartite ...
Then the local edge histograms for each of these blocks are obtained. Edges are broadly grouped into five bins: vertical, horizontal, 45 diagonal, 135 diagonal, and isotropic. So, texture of a frame is represented by an 80-dimensional(16 blocks. 5 bi

Distributed Adaptive Learning of Graph Signals - IEEE Xplore
Abstract—The aim of this paper is to propose distributed strate- gies for adaptive learning of signals defined over graphs. Assuming the graph signal to be ...

Video-to-Shot Tag Propagation by Graph Sparse Group ... - IEEE Xplore
proposes a novel Graph Sparse Group Lasso (shorted for GSGL) model to linearly reconstruct the visual feature of the test shot with the visual features of the ...

Graph-Based Multiprototype Competitive Learning and ... - IEEE Xplore
Oct 12, 2012 - to deal with high-dimensional data clustering, i.e., the fast graph- based multiprototype competitive learning (FGMPCL) algorithm.

Efficient Multiple Hypothesis Tracking by Track Segment ... - IEEE Xplore
Burlington, MA, USA. {chee.chong, greg.castanon, nathan.cooprider, shozo.mori balasubramaniam.ravichandran}@baesystems.com. Robert Macior. Air Force ...

Leakage Power Reduction in Flip-Flops by Using ... - IEEE Xplore
Leakage Power Reduction in Flip Flops by Using. MTCMOS and ULP Switch. Saihua Lin. Department of Electronic Engineering. Tsinghua University. Beijing ...

Modelling of Wave Propagation in Wire Media Using ... - IEEE Xplore
Abstract—The finite-difference time-domain (FDTD) method is applied for modelling of wire media as artificial dielectrics. Both frequency dispersion and spatial ...

segmentation and tracking of static and moving objects ... - IEEE Xplore
ABSTRACT. In this paper we present a real-time object tracking system for monocular video sequences with static camera. The work flow is based on a pixel-based foreground detection system followed by foreground object tracking. The foreground detecti

Posterior Cramer-Rao Bounds for Multi-Target Tracking - IEEE Xplore
Jan 1, 2006 - Though basically defined for estimation of deterministic parameters, it has been extended to stochastic ones in a Bayesian setting. In the target ...

Adaptive Output-Feedback Fuzzy Tracking Control for a ... - IEEE Xplore
Oct 10, 2011 - Adaptive Output-Feedback Fuzzy Tracking Control for a Class of Nonlinear Systems. Qi Zhou, Peng Shi, Senior Member, IEEE, Jinjun Lu, and ...

IEEE Photonics Technology - IEEE Xplore
Abstract—Due to the high beam divergence of standard laser diodes (LDs), these are not suitable for wavelength-selective feed- back without extra optical ...

wright layout - IEEE Xplore
tive specifications for voice over asynchronous transfer mode (VoATM) [2], voice over IP. (VoIP), and voice over frame relay (VoFR) [3]. Much has been written ...