red blood cell tracking using optical flow methods

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014 991

Red Blood Cell Tracking Using OpticalFlow Methods

Dongmin Guo, Anne L. van de Ven, and Xiaobo Zhou

Abstract—The investigation of microcirculation is an importanttask in biomedical and physiological research because the micro-circulation information, such as flow velocity and vessel density, iscritical to monitor human conditions and develop effective thera-pies of some diseases. As one of the tasks of the microcirculationstudy, red blood cell (RBC) tracking presents an effective approachto estimate some parameters in microcirculation. The commonmethod for RBC tracking is based on spatiotemporal image anal-ysis, which requires the image to have high qualification and cellsshould have fixed velocity. Besides, for in vivo cell tracking, cellsmay disappear in some frames, image series may have spatial andtemporal distortions, and vessel distribution can be complex, whichincrease the difficulties of RBC tracking. In this paper, we proposean optical flow method to track RBCs. It attempts to describe thelocal motion for each visible point in the frames using a local dis-placement vector field. We utilize it to calculate the displacementof a cell in two adjacent frames. Additionally, another optical flow-based method, scale invariant feature transform (SIFT) flow, isalso presented. The experimental results show that optical flow isquite robust to the case where the velocity of cell is unstable, whileSIFT flow works well when there is a large displacement of thecell between two adjacent frames. Our proposed methods outper-form other methods when doing in vivo cell tracking, which can beused to estimate the blood flow directly and help to evaluate otherparameters in microcirculation.

Index Terms—In vivo cell tracking, intravital microscopy imag-ing, optical flow, red blood cells, scale invariant feature transform(SIFT) flow.

I. INTRODUCTION

M ICROCIRCULATION is pivotal for the maintenance ofhuman’s tissues and organs since the exchange of oxygen

and nutrients between blood and tissue occurs at this level. Dis-eases such as diabetes [1], hypertension [2], and cancer [3], [4]are associated with changes in microcirculation. Changes at

Manuscript received April 4, 2013; revised August 2, 2013; accepted Septem-ber 5, 2013. Date of publication September 16, 2013; date of current versionMay 1, 2014. This work was supported by Funding: NIH U01HL111560-01(Zhou), NIH 1U01CA166886-01 (Zhou), NIH 1R01DE022676-01 (Zhou) andin part by a NIH/NCI Physical Sciences Oncology Center (PS-OC) Pilot ProjectAward to AV and ML as part of the Methodist (U54CA143837) PS-OC Center.

D. Guo and X. Zhou are with Department of Radiology, Wake For-est University Health Sciences, Winston-Salem, NC 27157 USA (e-mail:[email protected]; [email protected]).

A. L. van de Ven is with Department of Physics, Northeastern University,360 Huntington Ave, Boston MA 02115 USA (e-mail: [email protected]).

This paper has supplementary downloadable material available athttp://ieeexplore.ieee.org.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/JBHI.2013.2281915

the level of the microcirculation can significantly impact tissuestructure and function [5]. Therefore, monitoring microcircu-lation information can help to detect potential diseases at anearly stage. Thanks to the use of cellular markers tagged withfluorescent protein, some microcirculation information, such asoxygen saturation, vessel diameter, flow velocity, vessel density,and permeability, is easy to obtain by a noninvasive way, whichis based on analyzing imaging of the microvessels by using amicroscopy [6]–[9]. The analysis of these images allows us toobtain multiple microcirculation parameters. RBCs take a crit-ical role in oxygen transport through microcirculation. There-fore, tracking RBCs is crucial to measure plasma flow velocity,RBC passage velocity, and even oxygen saturation [8], [10].

There are several methods used for RBC tracking; mostof them are based on the analysis of a spatiotemporal im-age [8], [11], [12]. However, these methods require that cellsshould move with stable velocity and the trajectory of cellsshould be very clear in blood vessels. Additionally, for in vivocell tracking, there are some challenging problems we cannotavoid. First, cells may move in and out of the plane-of-focusin large vessels, which causes cells disappear in some frames.Second, the imaging area is rarely stationary due to the heart-beat and movement of subject. Thus, the image series may haveboth spatial and temporal distortions. Third, blood flow can behighly complex in cases such as cancer, with blood cells simul-taneously flowing in multiple directions. The common methodsaforementioned cannot work well in these cases. More effectivemethods are highly required.

In this paper, we obtain a time series of in vivo blood flowimages in tumor-bearing mice by high-resolution intravital mi-croscopy (IVM) of fluorescently labeled RBCs. We proposetwo methods for automatic tracking and measurement of themotion of RBCs in the microvessels based on the optical flowalgorithm [13], [14], which attempts to describe the local mo-tion for each visible point in an image sequence using a localdisplacement vector field. Here, we use the algorithm to cal-culate the displacement of a cell in two adjacent frames in theblood cell movement video. In our method, we first use a math-ematical morphology method to obtain the location of a cell inthe first frame. Then, in the following frames, we calculate thedisplacement of the cell between two adjacent frames. Hence,we can obtain a cell’s location in every frame in the video andthe trajectory of cell movement is therefore recorded. With theinformation, we can easily calculate some microcirculation pa-rameters, such as blood flow velocity and oxygen saturation.Considering that the method cannot solve the problems in allcases of cell movement, another optical flow method, SIFT flow,which tries to align an image to its nearest neighbors in a large

2168-2194 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

992 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 3, MAY 2014

Fig. 1. Representative video frame of RBCs (blue) in the vasculature (green)of a BCM-4195 tumor.

number of images by extracting keypoints of an image, is alsopresented. The experimental results show that both methods cantrack cells accurately. Optical flow is especially robust in caseswhere the velocity of the cell is unstable, while SIFT flow workswell when there is a large cell displacement between adjacentframes. In contrast, common cell tracking methods which areall based on spatiotemporal images, cannot conduct effectivecell tracking in these cases. Therefore, our proposed methodsoutperform other methods for in vivo cell tracking.

The remainder of this paper is organized as follows. Section IIdescribes how the RBC tracking data are obtained. Section IIIintroduces our proposed methods. The design of experimentsis presented in Section IV. Section V describes experimentalresults and the corresponding discussions. Finally, Section VIoffers our conclusions.

II. IMAGE ACQUISITION

IVM is a powerful optical imaging technique for collectingdynamic data in vivo with high spatial and temporal resolu-tion [15]–[19]. Here, live tumor-bearing mice were imaged us-ing an upright Nikon A1R laser-scanning confocal microscopeequipped with a resonant scanner for high-speed imaging [20].BCM-4195 tumor fragments were grown on the mammary padof SCID/beige mice (Harlan Laboratories, Wilmington, MA,USA) [21], [22] and exposed for IVM using an aseptic skin-flapprocedure [23]. Mice received an one-time intravenous injec-tion of DiD-labeled autologous red blood cells 1–3 days priorto imaging [23] and a bolus injection of 40 kDa FITC-dextrantracer (1 μM, Invitrogen) immediately prior to imaging [20].Videos were acquired at 30 frames-per-second (fps) during si-multaneous excitation with 488 and 640 nm lasers. Emissionswere collected by photomultiplier tubes with band-pass filtersof ∼ 30–35 nm width centered at 525 nm (FITC-dextran tracer)and 670 nm (RBCs). All animal protocols were reviewed andapproved by the IACUC [24] at The Methodist Hospital Re-search Institute (TMHRI).

Flow within BCM-4195 tumor was visualized by trackingthe movement of RBCs. Since only a small number of RBCswere labeled (∼ 3%), individual cells could be resolved. Thewalls of the tumor vasculature were delineated using FITC-dextran. Fig. 1 shows an example video frame, captured overa 1000 × 500 μm field of view (FOV). Flow through these andother BCM-4195 tumor vessels was observed to be irregular,

characterized by pulsatile RBC movement and simultaneousRBC movement in multiple directions. Quantitative data wereextrapolated using the analysis techniques described next.

III. METHOD

Two cell tracking methods, which are based on optical flowand SIFT flow, are introduced in this section. These methods arecommonly used in vehicle tracking and robot navigation [14],[25]. Considering that RBCs moving in blood vessels have thesimilar movement styles as vehicles and robots, we can applythe algorithms to the tracking of RBCs.

A. Optical Flow

The study of optical flow has a long history in visual per-ception and computer vision. Optical flow describes the localmotion for each visible point in an image sequence by the localdisplacement vector field [26]. For digital image sequences, itspecifies how much each image pixel moves between adjacentimages and can be efficiently computed using successive imagematching. Many algorithms exist to solve or optimize opticalflow methods [13], [14], [27]–[31], developed based on the as-sumption that the gray value of a pixel is not changed by thedisplacement from time t to time t + 1, which is represented by

I(x + u, y + v, t + 1) = I(x, y, t) (1)

where I(x, y, t) denotes a rectangular image sequence, andw := (u, v, 1)T is the searched displacement vector betweenan image at time t and another image at time t + 1. u and v arethe horizontal and vertical components of the flow field.

However, not all optical methods are suitable for the cell-tracking problem discussed in this paper. In the traditional opti-cal flow methods, i.e., Lucas–Kanade [27] and Horn–Schunck[32], the vector field extracted will not be dense or will losein discontinuities. Also, in the recent variational optical flowmethod such as [29] and [13], the discontinuities in the vectorfield are well preserved by introducing sophisticated measuresfor edge regions. The performance of the matching (1) is lim-ited due to noise and the relatively small size of cells’ objectsappeared in the videos. In this paper, we recommend using thealgorithm of Liu [33] to obtain the vector fields u and v. Liu’salgorithm is also under the brightness constancy assumption.

Unlike previous methods, which consider the matching ofimage pixels, Liu’s method considers the matching of SIFT-transformed images. The advantages of using SIFT transformin optical flow calculation are as follows: 1) significant imagefeatures are well encoded in the SIFT feature vectors; 2) largedistortions due to cell motion or noise are insensitive to SIFTfeatures; 3) long-term cell motion can still be recovered by SIFT-matching; and 4) the SIFT-based optical flow can be computedin the fast discrete optimization method.

B. SIFT Flow

SIFT flow exploits the SIFT feature descriptors that are rea-sonably invariant to changes in illumination, scaling, rotation,image noise, and small changes in viewpoint. The SIFT descrip-

GUO et al.: RED BLOOD CELL TRACKING USING OPTICAL FLOW METHODS 993

Fig. 2. SIFT feature histogram. (Left) Each pixel in a local area is assignedwith a specific orientation, and (right) the orientations are then accumulated intoa 4-bin directional histogram. The histogram vectors can be directly used forkeypoint matching.

tor was first proposed by Lowe [34], and then further developedand improved by Bay [35] and Dalal [36]. The SIFT descriptor,as its followers, is a vector encoding the local orientation infor-mation such as edges and corners of a local area. The orientationinformation is invariant to illumination and change of scale.Thus, SIFT is insensitive to light, shadow, and even most affinetransforms encountered in natural images. SIFT flow matchingcan avoid the noise and change of cell scales in image sequence.The SIFT flow method includes two major steps: 1) SIFT ex-traction and 2) belief propagation optimization. This paper willbriefly introduce them in the following sections.

1) Computation of Scale Invariant Feature Transform: Thegoal of SIFT computation is to extract the local orientationhistogram. It generally takes four steps in SIFT computation:extrema detection, keypoint localization, edge elimination, andorientation assignment. The first three steps attempt to extractsignificant pixel candidates which are robust to scale changes.The last step extracts the local orientation information and storethe orientation into histogram vectors as Fig. 2 shows [37],where each pixel in a local area is assigned with a specificorientation. The orientations are then accumulated into a four-bin directional histogram. In SIFT flow computation, the majorsteps in SIFT transform are simplified due to computational ef-ficiency. The orientation histograms are directly extracted fromthe edge images obtained from the consecutive difference ofGaussian convolutions.

2) Formulation of SIFT Computation: SIFT flow com-putation is formulated as the minimization over the SIFT-transformed image pair. For a given image pair, the objectivefunction is written as

E(w) =∑

p

min(‖s1(p) − s2(p + w(p))‖1 , t)

+∑

p

η(|u(p)| + |v(p)|)

+∑

(p,q)∈ε

min(α |u(p)| − |u(q)| , d)

+ min(α |v(p)| − |v(q)| , d), (2)

where s1 is the SIFT descriptor over pixel p in the first frameand s2 is the SIFT descriptor on pixel q in the second frame. tand d are thresholds. w(p) = (u(p), v(p)) is the flow vector atp which represents the warping that aligns both of the frames.

Fig. 3. Structure of the proposed method.

Fig. 4. Locating cells using the mathematical morphology method.

Unlike the general variational approaches such as [29] and[13], the functional (2) cannot be directly solved by the Euler–Lagrange equation. Instead, [14] applies a special model ofbelief-propagation (BP) for solving (2). The BP model does notrequire the calculation of function gradient, and the resulting wis iteratively updated by the message-passing algorithm in theBP model.

IV. EXPERIMENTS

The experiments were conducted in MATLAB7.1 on an IntelCore2 Quad 3.0 GHz CPU, 8 GB Memory desktop. The basicstructure of our proposed method is presented in Fig. 3. Inour method, for a specific cell, we first use the mathematicalmorphology method to obtain the cell’s location in the firstframe. Then, we find the displacement of the cell between thefirst and the second frame using the optical flow and SIFT flowalgorithms, separately. Given a cell’s location in the first frameand the displacement between the first and the second frame, itis easy to calculate the cell’s location in the second frame. Thisprocess is iterated until the location of cell in the last frame isfound.

A. Locating Cells

We first used mathematical morphology to obtain the locationof cells in the first frame. In detail, for an input image I , it firstfinds the regional maxima of I . The output J , which has thesame size of I , identifies the locations of the regional maximain I . In J , pixels that are set to 1 identify regional maxima;all other pixels are set to 0. Then, we perform morphologicaloperations on J to shrink an object to a point.

Fig. 4 shows the location of cells identified using this method.Multiple RBCs moving through a blood vessel were visualizedby blue fluorescence, shown here using red circles to mark their


locations. Randomly selected cells were subjected to opticalflow and SIFT flow algorithms to obtain their location in thenext frame.

B. Displacement Calculation Using Optical Fow

The algorithm is used to calculate the displacement of a givencell between two frames. The MATLAB package for this algo-rithm is provided in [38]. According to our previous experi-mental results, the following parameters were set to obtain thebest tracking results. The regularization weight (alpha) was setto 0.04, the downsample ratio was set to 0.55, the width of thecoarsest level (minWidth) was set to 5, the number of outer fixedpoint iterations (nOuterFPIterations) was set to 7, the numberof inner fixed point iterations (nInnerFPIterations) was set to1, and finally, the number of SOR iterations (nSORIterations)was set to 30 when we preferred accuracy over training time.However, if we want to get a faster estimated result, (nSORIter-ations) can be set to 15. Eventually, the horizontal and verticalcomponents of the flow field were computed using the opticalflow algorithm, and locations of the cell in the whole image setsor videos were therefore computed.

C. Displacement calculation using SIFT flow

The SIFT flow algorithm is provided in [39]. The followingparameters were set to get good estimated results. The weightof the truncated L1-norm regularization on the flow (alpha) wasset to 5; the threshold of the truncation d was set to 20; theweight of the magnitude of the flow (gamma) was set to 0.01;the number of iterations (nIterations) was set to 30; the numberof hierarchies in the efficient BP implementation (nHierarchy)was set to 2; and finally, the half size of the search window(wsize) was set to 5. Same as optical flow, we computed thehorizontal and vertical components of the flow field and thenfound the locations of the cell frame by frame.

D. Validation

To test if the estimated locations by using the above algo-rithms are good enough, a validation experiment was proposed.We would like to find out the observed value of cell locationin each frame and compare it with the estimated value. In ourvalidation experiments, ten people (three biologists, three math-ematicians, and four specialists in medical image processing)were asked to run experiments and locate the cells’ locationsframe by frame by eyes. Therefore, for a cell in a video, wehave such an observation matrix as follows:

A =

⎡

⎢⎢⎢⎢⎣

x1,1 , x1,2 , . . . , x1,n

x2,1 , x2,2 , . . . , x2,n

...

x10,1 , x10,2 , . . . , x10,n

⎤

⎥⎥⎥⎥⎦. (3)

Assume that the ten persons’ observations to the same frameare independent. Therefore, given a fixed frame j, xi,j (i =1, . . . , 10, indicates the number of persons) are with the normaldistribution. We use p-value (p = 0.05) to reject the invalid

Fig. 5. Tracking singular cell in complicated background. (a) Optical flow,(b) SIFT flow.

observations in xi,j (i = 1, . . . , 10) and average the rest validvalues as the true location of a cell in frame j.

V. RESULTS AND DISCUSSION

This section describes three independent experiments inwhich optical flow and SIFT flow predictions were compared tomanual measurements.

A. Experiment 1: Track a Single Cell in aComplicated Background

This experiment shows the important cell tracking in poorlydistinguished backgrounds. IVM images are subject to a highdegree of noise, which can result in poor contrast. Effective de-lineation of the vessel walls is dependent on the quality of tracerused and can vary with how the tissue of interest is mounted forimaging. Additionally, IVM techniques generally capture im-ages over a relatively narrow axial plane, such that objects canbe lost if they move out of the plane of focus. This is particularlya problem with large dilated vessels, such as those observed intumors. In this case, some common methods, like Gabor filterfor the spatiotemporal image analysis method, will fail to getthe missed locations of the cells in the frames.

We hence use our proposed methods to track cells in thiscase. Fig. 5(a) and (b) shows the optical flow and SIFT flowtracking results overlaid on the last video frame. Both methodstracked the cell’s movement very well. However, only observingthe two figures, it is hard to see their performance clearly. So,we conducted the validation experiment by recording the cell’slocations in each frame obtained by optical flow and SIFT flowat the same time. The validation results are shown in Fig. 6,where the horizontal axis and vertical axis represent the cell’slocation in a frame. Ninety frames were processed in total. Theblack stars present the cell’s locations computed by SIFT flow,and the blue circles stand for the same cell’s locations computedby optical flow, while the red dots show the observed locations(used as true values of these locations) obtained by ten observersand processed as introduced in Section IV-D. To see the resultsclearly, we zoomed in the two parts of the original curve, as


Fig. 6. Validation of optical flow and SIFT flow.

TABLE IMSE BETWEEN PREDICTED VALUES AND OBSERVED VALUES

ellipses A and B show. Ellipse A includes five frames. From thefigure, we can see that a red dot is missing in one of the framesin ellipse A. It means that all of the ten observers cannot tellthe location of the cell in that frame. In fact, the cell disappearsin that frame and appears again in the next frame. However,both SIFT flow and optical flow can predict the locations of thecell in each of the five frames. Ellipse B represents the samecase, demonstrating that in some frames, our proposed methodcan compute a cell’s location accurately when cells cannot bevisualized by eyes.

We used mean squared error (MSE) to calculate the errorbetween predicted values and observed values, as Table I shows.Specially, we calculated the Euclidean distance between thepredicted location and the observed location in each frame.In the same video, we randomly tracked ten cells’ locationsusing the two methods. It should be noted that some cells aremissed in some frames, we do not have observed values ofthe cells’ locations and cannot judge whether the predictionsof the two methods are good or not. We therefore deleted thepredicted values in these frames when calculating the predictionaccuracy. In the table, both methods have the same predictivecapacity for predicting the locations of eight of the ten cells.However, optical flow outperforms SIFT flow when predictingthe locations of cells 6 and 7, and SIFT flow shows slightly betterperformance when predicting the locations of cell 9. Optical flowworks better than SIFT flow in the case when the cells move ina complicated environment. This is because the characteristicsof optical flow are quite robust to abrupt movement of objective

Fig. 7. Tracking cells with two different movement tendencies. (a) SIFT flowand (b) optical flow.

since pixel-wise correspondence has advantages when handlingcomplicated and unreliable scene alignment.

B. Experiment 2: Tack Cells Fowing in TwoDifferent Directions

This experiment shows the case when two cells in the samevideo flow in different directions. Such a case is very typical intumors, where microvessels are generally highly branched andtortuous [40]. Common cell tracking methods cannot work wellin this case since they assume that all particles in the flow shouldhave the same velocity and require all particles should have thesimilar movement tendency.

Fig. 7 shows an example in which multiple vessels are con-tained in a single FOV and each vessel has its own associateddirection of flow. Two vessels were analyzed: Fig. 7(a) is the ex-perimental result using the SIFT flow algorithm, while Fig. 7(b)is the result of optical flow with the same video. The arrowlines in both figures show the direction of blood flows. Fromthe two figures, we can see that both SIFT flow and optical flowcan track the movement of cells accurately, even though theirdirections of movement are quite different.

The predictive ability of the two algorithms is shown in Figs. 8and 9, respectively. Fig. 8 shows the cell tracking result of cellA and Fig. 9 the tracking results of cell B in Fig. 7, respectively.For both of the two cell movement trajectories, we zoomed in apart of them to see the detailed tracking results. We can see thatthe two cells’ movements are abrupt in some places (see Figs.8 and 9). Different performances of SIFT flow and optical flowalgorithms are observed, particularly when cells make abruptchanges in velocity or direction. The SIFT flow method appearsto show better tracking than optical flow because the opticalflow method produces pixel-to-pixel correspondences betweentwo images, while SIFT flow creates SIFT descriptors instead


Fig. 8. Tracking cell A: results and validation.

Fig. 9. Tracking cell B: results and validation.

TABLE IIMSE BETWEEN PREDICTED VALUES AND OBSERVED VALUES

of raw pixels. So SIFT flow is more robust to the large changesbetween two adjacent frames.

Table II presents the MSE of the two methods compared withthe observed values obtained in Section IV-D. Ten cells wererandomly selected from the vessels in the video. The predictederrors of the two methods are listed in Table II. Two points shouldbe noted in this experiment: First, both algorithms perform wellwhen the trajectory of the cell is quite simple (see Fig. 8), andSIFT flow shows especially good performance when the cellroute is subject to large changes (see Fig. 9). Second, bothalgorithms have quite high performance to track cells when aseries of cells move toward different directions.

Fig. 10. Tracking cells through heartrate-induced abnormalities. (a) SIFT flowand (b) optical flow.

Fig. 11. Tracking cell B: results and validation.

C. Experiment 3: Track Cells Through the Motionof the Subject

This experiment shows the importance of tracking large trackdeviations between frames. Animal motion due to heart beat,breathing, and twitching can induce artifacts in cell tracking,which appears as a large deviation in trajectory from one frameto the next.

Fig. 10 shows the experimental results using both algorithms.In this video, the location of the cell in some frames is distorteddue to animal heartbeat. Normally, other methods will fail totrack the cell in successive frames since the location of the cellin the distorted frame is wrong. However, from the two figures,we can see that both of our proposed methods can track the cellvery well. To compare them clearly, we conducted validationexperiments, as shown in Fig. 11. To see the two methods’performance accurately, we randomly selected ten cells to trackand the accuracy of the two methods are listed in Table III. Wecan see in the table that SIFT flow outperforms optical flow in


TABLE IIIMSE BETWEEN PREDICTED VALUES AND OBSERVED VALUES

this case. This is because SIFT flow is much more robust tothe large change of scene between two adjacent frames sinceSIFT flow creates SIFT descriptors for matching rather thanpixel-to-pixel correspondences.

VI. CONCLUSION

This paper proposed two objective tracking methods, opticalflow and SIFT flow, to solve the difficulties for in vivo RBCtracking. There are some challenging problems we cannot avoidduring in vivo cell tracking, such as cells may disappear in someframes, imaging can be disturbed by heartbeat and movement ofsubject, and vessels can be highly complex, and therefore cellsin the same imaging area may flow in multiple directions. Theexperiments in this paper especially selected three difficult invivo tracking cases to test our proposed methods. The first oneis to track a single cell in a complicated background, the secondone is to track cells flowing in two different directions, and thethird one is to track cells though motion of the subject. Theexperimental results showed that both of the methods can workwell in all of the cases. However, optical flow is much morerobust to the case where the velocity of cell is unstable, whileSIFT flow works better when there is a large displacement of thecell between two adjacent frames. Therefore, in the applicationof the two methods, we can select to use them according to dif-ferent cases. Because both of the methods have to calculate thedisplacements of each pixel, the calculation is time consuming.In future, we will improve the algorithms in the aspect of speedand accuracy.

ACKNOWLEDGMENT

Special thanks to Melissa D. Landis for providing the animalmodel.

REFERENCES

[1] C. Chang, R. Tsai, W. Wu, S. Kuo, and H. Yu, “Use of dynamic capil-laroscopy for studying cutaneous microcirculation in patients with dia-betes mellitus,” Microvasc. Res., vol. 53, no. 2, pp. 121–127, 1997.

[2] B. Levy, G. Ambrosio, A. Pries, and H. Struijker-Boudier, “Microcircu-lation in hypertension,” Circulation, vol. 104, no. 6, pp. 735–740, 2001.

[3] L. Weiss, U. Nannmark, B. Johansson, and U. Bagge, “Lethal deforma-tion of cancer cells in the microcirculation: A potential rate regulator ofhematogenous metastasis,” Int. J. Cancer, vol. 50, no. 1, pp. 103–107,2006.

[4] Q. G. de Lussanet, W. H. Backes, A. W. Griffioen, A. R. Padhani,C. I. Baeten, A. van Baardwijk, P. Lambin, G. L. Beets, J. van

Engelshoven, and R. G. Beets-Tan, “Dynamic contrast-enhanced mag-netic resonance imaging of radiation therapy-induced microcirculationchanges in rectal cancer,” Int. J. Radiat. Oncol. Biol. Phys., vol. 63, no. 5,pp. 1309–1315, 2005.

[5] G. Satiukova and O. Kurguzov, “Changes in microcirculation and theircorrection in patients with varicose veins and post-thrombotic syndrome,”Morfologiia (Saint Petersburg, Russia), vol. 118, no. 5, pp. 29–35, 2000.

[6] R. T. Ullrich, J. F. Jikeli, M. Diedenhofen, P. Bohm-Sturm, M. Unruh,S. Vollmar, and M. Hoehn, “In vivo visualization of tumor microvesseldensity and response to anti-angiogenic treatment by high resolution mriin mice,” PloS one, vol. 6, no. 5, p. e19592, 2011.

[7] C. Xu, W. U. Schmidt, K. Villringer, P. Brunecker, V. Kiselev, P. Gall,and J. B. Fiebach, “Vessel size imaging reveals pathological changes ofmicrovessel density and size in acute ischemia,” J. Cerebral Blood FlowMetabol., vol. 31, no. 8, pp. 1687–1695, 2011.

[8] S. A. Japee, R. N. Pittman, and C. G. Ellis, “Automated method for track-ing individual red blood cells within capillaries to compute velocity andoxygen saturation,” Microcirculation, vol. 12, no. 6, pp. 507–515, 2005.

[9] M. Tomita, I. Schiszler, H. Takeda, Y. Tomita, T. Osada, M. Unekawa,N. Tanahashi, and N. Suzuki, “Spatial and temporal heterogeneity ofsingle capillary plasma flow and RBC tracking in the rat cerebral cortex,”J. Cerebral Blood Flow Metabol., vol. 25, pp. S164–S164, 2005.

[10] M. Unekawa, M. Tomita, Y. Tomita, H. Toriumi, K. Miyaki, and N. Suzuki,“RBC velocities in single capillaries of mouse and rat brains are the same,despite 10-fold difference in body size,” Brain Res., vol. 1320, pp. 69–73,2010.

[11] S. Yamamoto, Y. Nakajima, S. Tamura, Y. Sato, and S. Harino, “Extrac-tion of fluorescent dot traces from a scanning laser ophthalmoscope imagesequence by spatio-temporal image analysis: Gabor filter and radon trans-form filtering,” IEEE Trans. Biomed. Eng., vol. 46, no. 11, pp. 1357–1363,Nov. 1999.

[12] Y. Chen, Z. Zhao, L. Liu, and H. Li, “Automatic tracking and measurementof the motion of blood cells in microvessels based on analysis of multiplespatiotemporal images,” Meas. Sci. Technol., vol. 22, no. 4, p. 045803,2011.

[13] T. Brox, A. Bruhn, N. Papenberg, and J. Weickert, “High accuracy opticalflow estimation based on a theory for warping,” in Proc. 8th Eur. Conf.Comput. Vis., 2004, pp. 25–36.

[14] C. Liu, J. Yuen, and A. Torralba, “Sift flow: Dense correspondence acrossscenes and its applications,” IEEE Trans. Pattern Anal. Mach. Intell.,vol. 33, no. 5, pp. 978–994, May 2011.

[15] M. J. Pittet and R. Weissleder, “Intravital imaging,” Cell, vol. 147, no. 5,pp. 983–991, 2011.

[16] R. N. Germain, M. J. Miller, M. L. Dustin, and M. C. Nussenzweig, “Dy-namic imaging of the immune system: Progress, pitfalls and promise,”Nature Rev. Immunol., vol. 6, no. 7, pp. 497–507, 2006.

[17] J. N. Kerr and W. Denk, “Imaging in vivo: Watching the brain in action,”Nature Rev. Neurosci., vol. 9, no. 3, pp. 195–205, 2008.

[18] G. M. Tozer, S. M. Ameer-Beg, J. Baker, P. R. Barber, S. A. Hill,R. J. Hodgkiss, R. Locke, V. E. Prise, I. Wilson, and B. Vojnovic, “Intravi-tal imaging of tumour vascular networks using multi-photon fluorescencemicroscopy,” Adv. Drug Del. Rev., vol. 57, no. 1, pp. 135–152, 2005.

[19] R. K. Jain, L. L. Munn, D. Fukumura et al., “Dissecting tumour patho-physiology using intravital microscopy,” Nature Rev. Cancer, vol. 2, no. 4,pp. 266–276, 2002.

[20] A. L. van de Ven, M. Wu, J. Lowengrub, S. R. McDougall, M. A. Chaplain,V. Cristini, M. Ferrari, and H. B. Frieboes, “Integrated intravital mi-croscopy and mathematical modeling to optimize nanotherapeutics de-livery to tumors,” AIP Adv., vol. 2, no. 1, p. 011208, 2012.

[21] M. D. Landis, B. D. Lehmann, J. A. Pietenpol, and J. C. Chang, “Patient-derived breast tumor xenografts facilitating personalized cancer therapy,”Breast Cancer Res., vol. 15, no. 1, pp. 1–9, 2013.

[22] A. F. Schott, M. D. Landis, G. Dontu, K. A. Griffith, R. M. Layman,I. Krop, L. A. Paskett, H. Wong, L. E. Dobrolecki, M. T. Lewis et al., “Pre-clinical and clinical studies of gamma secretase inhibitors with docetaxelon human breast tumors,” Clin. Cancer Res., vol. 19, no. 6, pp. 1512–1524,2013.

[23] A. L. van de Ven, P. Kim, O. Haley, J. R. Fakhoury, G. Adriani,J. Schmulen, P. Moloney, F. Hussain, M. Ferrari, X. Liu et al., “Rapidtumoritropic accumulation of systemically injected plateloid particles andtheir biodistribution,” J. Controll. Release, vol. 158, no. 1, pp. 148–155,2011.

[24] [Online]. Available:http://www.iacuc.org/aboutus.htm[25] K. R. Aires, A. M. Santana, and A. A. Medeiros, “Optical flow using color

information: Preliminary results,” in Proc. ACM Symp. Appl. Comput.,2008, pp. 1607–1611.


[26] J. J. Gibson, The perception of the visual world. Oxford, England:Houghton Mifflin, 1950.

[27] B. D. Lucas, T. Kanade et al., “An iterative image registration techniquewith an application to stereo vision,” in Proc. 7th Int. Joint Conf. Artif.Intell., 1981, pp. 674–679.

[28] D. J. Heeger, “Optical flow using spatiotemporal filters,” Int. J. Comput.Vis., vol. 1, no. 4, pp. 279–302, 1988.

[29] G. Aubert, R. Deriche, and P. Kornprobst, “Computing optical flow viavariational techniques,” SIAM J. Appl. Math., vol. 60, no. 1, pp. 156–182,1999.

[30] J. Barron and N. Thacker, “Tutorial: Computing 2d and 3d optical flow,”Imag. Sci. Biomed. Eng. Divis., Medical School, Univ. Manchester, 2005.

[31] A. Bruhn, J. Weickert, and C. Schnorr, “Lucas/kanade meetshorn/schunck: Combining local and global optic flow methods,” Int. J.Comput. Vis., vol. 61, no. 3, pp. 211–231, 2005.

[32] B. K. Horn and B. G. Schunck, “Determining optical flow,” Artif. Intell.,vol. 17, no. 1, pp. 185–203, 1981.

[33] C. Li et al., “Beyond pixels: Exploring new representations and applica-tions for motion analysis,” Ph.D. dissertation, Massachusetts Institute ofTechnology, Cambridge, MA, USA, 2009.

[34] D. Lowe, “Object recognition from local scale-invariant features,” in Proc.IEEE Comput. Vis., Proc. 7th IEEE Int. Conf., 1999, vol. 2, pp. 1150–1157.

[35] H. Bay, T. Tuytelaars, and L. Van Gool, “Surf: Speeded up robust features,”in Proc. Eur. Conf. Comput. Vis., 2006, pp. 404–417.

[36] N. Dalal and B. Triggs, “Histograms of oriented gradients for humandetection,” in Proc. IEEE Comput. Vis. Pattern Recognit., IEEE Comput.Soc. Conf., 2005, vol. 1, pp. 886–893.

[37] C. Zhu, “Video object tracking using sift and mean shift,” M.S. Thesis,Dept. Signals and Syst., Chalmers University OF Technology Goteborg,Sweden, 2011.

[38] C. Liu. (2009). [Online]. Available:http://people.csail.mit.edu/celiu/OpticalFlow/.

[39] C. Liu. (2010). [Online]. Available:http://people.csail.mit.edu/celiu/SIFTflow/

[40] H. Gerhardt and H. Semb, “Pericytes: Gatekeepers in tumour cell metas-tasis?,” J. Molecul. Med., vol. 86, no. 2, pp. 135–144, 2008.

Authors’ photographs and biographies not available at the time of publication.

red blood cell tracking using optical flow methods

Documents