Data Medical

Segmentation and Tracking of *D Neuron Microscopy Images Using a

PDE Based Method and Connected Component Labeling algorithm

Illhwan Jol

Yalin Wang Stephen Wong2 Shing-Tung Yau3 Tony F. Chan1

'Department of Mathematics 2Center for Bioinformatics 3Department of Mathematics

University of California, Los Angeles Harvard University

Harvard Medical School

{ylwang,ilhwanjo,chan}@math.ucla.edu ************@***.*******.*** ***@****.*******.***

Its initial contour can be placed anywhere in the image and

Abstract- In this paper we introduce our preliminary re-

search results for segmentation and labeling of 3-dimensional topological change of contour is allowed. It is also robust to

microscopy neuron image. We segment each of stacked 2- noise.

dimensional image slices using a Partial Differential Equation

B. Tracking objects with the CV Algorithm

(PDE) based algorithm and project previous slice segmentation

result to the next slide as an initialization condition. Then we Moelich and Chan proposed an algorithm for tracking ob-

label neurons using an efficient connected component labeling

jects in video sequences [5]. The algorithm can be described

algorithm. We show sample results obtained from real neuron

as a sequential segmentation in which the final segmentation

image data.

contour of a frame is used as the initial contour for the

I. INTRODUCTION segmentation of the next frame. In order for the tracking

with CV algorithm to work properly, it is required that

To study morphology of neurons in three dimensions can

initial contour be in contact with the object to be detected

help neuroscientists understand neuronal development [1].

[5]. Moelich et. al. added a modification based on target

Three dimensional image stacks of neurons such as motor

intensities to overcome this problem. In our work we only

neurons can be obtained by confocal microscopy [2]. In a

employ the following algorithm.

recent paper [3], Cai et. al. use slice-wise segmentation of

3D microscopy image stacks. Their work is based on GVF

snake model for segmentation of 2D images. In this paper we C0 initial contour

=

propose a segmentation-labeling model for 3D neuron image

based on tracking with Chan-Vese algorithm and connected for I= I to N

component labeling algorithm. Some preliminary results are

presented. Ck = Chan-Vese(Ck 1, Ik)

draw contour on image

II. TRACKING WITH CHAN-VESE ALGORITHM

AND CONNECTED COMPONENTS LABELING output frame

}

A. Chan-Vese Segmentation Algorithm

The Chan-Vese algorithm is a region-based segmentation

Fig. 1. The sequential tracking algorithm.

model which is based on the active contour model, the

Mumford-Shah functional and the Osher-Sethian level set

C. Connected Components Algorithm

method [4]. Given a grayscale image I: Q C RP -> R+

(p = 2, 3), the Chan-Vese algorithm finds a curve C that Once a binary image is obtained by segmentation, pixels in

represents a partition of Q into two regions Qin and Qout so the image can be grouped based on maximal connectivity by

that they give an optimal piecewise constant approximation applying a connected component operator [6]. We extended

of the image. The contour C minimizes the following energy the Lumia, Shapiro and Zuniga algorithm to 3D volume

image. With defined "neighbors" for each pixel, the original

(I c)2dx

E(C, Cl, C2) = A algorithm can label all the connected components in a 2D

j

image by finding all the equivalent classes in two passes. We

in

+,ulength(C)

(I(X) -C2)2d

+A2 extend this algorithm into 3D volume data. Specifically, we

label all the connected components in each 2D slice. With 3D

Qout

where C1, C2 are the average intensities in Qin and Qout neighborhood definition, we find all equivalent classes of the

respectively and Ai,,u are parameters. foreground pixels in two passes along the z direction (here

we assume each slice has x and y directions). The algorithm

This model can detect objects with edges that are not

necessarily defined by gradient or with smooth boundaries. is very efficient to label 3D neuron data.

1-4244-0278-6/06/$20.00 2006 IEEE

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 25, 2008 at 12:39 from IEEE Xplore. Restrictions apply.

D. Neuron Segmentation and Labeling

We segment the first 2D image slice using the Chan-Vese

algorithm, apply tracking with CV algorithm to the image

stack. We use the segmentation result as an initial contour

for the next 2D image and sequentially apply this process

to the whole stacked 2D images. In practice, we can use

a manual segmentation by medical experts for the first 2D

image and we can also start with any 2D image in the stack.

Once all the 2D images of the stack is segmented, we apply

connected component operator to the stack to obtain labeling

of objects.

III. SAMPLE RESULTS

We tested our method on a 3D real neuron image data. The

tested image is of size 512 x 42 x 256 and contains seven

neurons. The results are shown in Figure 2 and 3. As shown

in Figure 2, With previous segmentation result (Figure 2 (a))

as the initial condition for the segmentation in the next slide.

(a)

Figure 2 shows the sequence of the segmentation on the next

slide. Figure 2 demonstrates that our algorithm can segment

Fig. 3. 3D rendering of tracked neuron data compared with ground truth

the 3D neuron image volume with Chan-Vese model. data. (a) 3D rendering of tracked neuron data (in three different view

Figure 3 shows 3D rendering segmentation results and the directions with original intensity values); (b) ground truth data of the given

neuron image in which 7 neurons are labeled in different colors.

ground truth data. Figure 3 (a) show three different views of

tracked neurons with original image intensity values. (b) is

the ground truth data of the given image. Compare (a) and

focus on new algorithm development to distinguish close

(b), we can see most of neurons are correctly segmented and

neurons.

labeled.

REFERENCES

[1] N. Kasthuri and J. W. Lichtman, "The role of neuronal identify in

synaptic competition", Nature, vol. 424, no. 6974, pp. 426-430, 2003.

[2] G. Feng, R. H. Mellor, M. Bernstein, C. Keller-Peck, Q. T. Nguyen,

M. Wallace, J. M. Nerbonne, J. W. Lichtman and J. R. Sanes,

"Imaging neuronal subsets in transgenic mice expressing multiple

spectral variants of GFP", Neuron, vol. 28, no. 1, pp. 41-51, Oct.

2000

[3] H. Cai, X. Xu, J. Lu, J. Lichtman, S.P. Yung, and S.T.C Wong, "Shape-

constrained repulsive snake method to segment and track neurons in

3D microscopy images", ISBI 2006, pp. 538-541.

[4] T. Chan and L. Vese, "An active contour model without edges, "Int.

Conf Scale-Space Theories in Computer Vision, 16(2):266-277, 1999.

[5] M. Moelich and T. Chan, "Tracking Objects with the Chan-Vese

Algorithm, " UCLA CAM Report 03-14.

[6] R. M. Haralick and L. G. Shapiro, "Computer and Robot Vision

Volume I, " Addison Wesley, vol 1, 1992.

(c)

Fig. 2. Final segmentation contour (a) of a 2D image slice is used as the

initial contour for the next, (b). (c) shows the segmentation of the next 2D

image.

IV. CONCLUSIONS AND FUTURE WORK

One drawback of the tracking with CV algorithm is the

inability to distinguish between objects with similar intensi-

ties that are close to each other [5]. Neurons in microscopy

data have similar intensities and if two neurons are close

enough to each other or the boundary between them is weak,

contours may merge with each other. Our future work will

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 25, 2008 at 12:39 from IEEE Xplore. Restrictions apply.

Contact this candidate