Region and Contour Based Cell Cluster Segmentation
Algorithm for In-SituMicroscopy
A. Sheehy1, G. Martinez1, J.-G. Frerichs2, T. Scheper2
1Image Processing and Computer Vision Research Laboratory (IPCV-LAB), Universidad de Costa Rica
2Institute of Technical Chemistry (TCI), Leibniz Universität Hannover
IEEE CCE, Mexico City, Mexico, November 2008
Topics
• Introduction• Existing algorithms• Proposed algorithm• Experimental results• Summary and conclusions
Introduction
Bioreactor
Cell culture
microscope
CCD Camera
Capturedintensityimage
Imageanalysis
Sampling zone
Estimation of importantparameters of thecell culture
In-situ microscopy
IntroductionCell image segmentation
Binary image IB(x,y), x:1..L,y:1..M, with pixels belongingto the cell clusters in whiteand pixels belonging to thebackground in black
Segmentation
Intensity image I(x,y), x:1..L, y:1..M, captured by the in-situ microscope
Existing Algorithms
• Thresholding based segmentation algorithmsEspinoza’s algorithm: Latest contribution for segmenting BHK cells images captured by an in-situ microscope
• Contour based segmentation algorithms• Region based segmentation algorithms• Combinations of the previous
Cell cluster segmentation algorithms can be roughlydivided into 4 main groups:
Espinoza’s Algorithm
a. Estimate the local variance σ2(x,y) at eachimage position (x,y) using a 3x3 window:
1) Global thresholding of the local variance
3 3
3 3
1( , ) ( , )49 m n
m x y I x m y n=− =−
≅ ⋅ + +∑ ∑
[ ]3 3
22
3 3
1( , ) ( , ) ( , )50 m n
x y I x m y n m x yσ=− =−
≅ ⋅ + + −∑ ∑
2
2
0, ( , )( , )
1, ( , )g
Bg
if x y thI x y
if x y th
σ
σ
≤= >
b. Globally threshold the local variance:
thg is estimated applying a maximumlikelihood algorithm (Kittler & Illingworth)
c. Apply a 5x5 median filter to IB and eliminatesmall regions (<0.5%)
with:
a. Find for each segmented region r all the borderpixels Ig(r), g:0...G(r).
Espinoza’s Algorithm
( )
( ) ( )( )
1
1 rHr rl hr
hth I
H =
= ∑
2) Local thresholding of the intensity
c. Locally threshold the intensity in each region r :
d. Apply a 5x5 median filter to IB and eliminatesmall regions (<0.5%)
b. Select from all the border pixels Ig(r) only thoseborder pixels Ih(r), j:0...H(r) that have an intensity value similar to the background intensity value byapplying a RANSAC algorithm
with:
ryxthyxIifthyxIif
yxI r
r
B ∈∀
>≤
= ),(,),(,0),(,1
),( )(
)(
Espinoza’s Algorithm
Lowreliability in low contrastimages:
Lowprecision:
Problems
To segment the background using a region and contour based approachand then to invert the resulting binary image to get the cell cluster regions
Proposed Algorithm
we propose:
Because:
1. The background is homogeneous
2. The cell clusters are non homogeneous with darker intensityvalues and well defined contours
Proposed Algorithm1) Seed selection
a. Select as background seed candidates those pixles that meetthe following rules:
2 22( , ) 0.20x yσ σ< ⋅
2 2( , )I x y m σ> −
where, and are the mean and standard deviation of thepixels’ intensity values in the background
2σ2m
Proposed Algorithmm2 and σ2 , as well as the mean m1 and standard deviation σ1 of the pixels’ intensity values in the cell clusters are estimated modeling the probability density function of the intensity values p(I) as a sum of two weighted Gaussian density functions:
∑ ∑= =
−⋅⋅−⋅−⋅⋅=2
1
2
1
21 2
))(log()(2
)2log(2
))(log()()(j j
jjjjNkkcNNkckcNkL σπ
where N is the total number of pixels, h(I) is the number of pixels withintensity I and k is selected maximizing the following likelihood function:
10
( )k
I
h IcN=
= ∑ 101
1 ( )( )
k
I
h Im Ic k N=
= ⋅ ⋅∑ 2 21 1
01
1 ( )( )( )
k
I
h II mc k N
σ=
= ⋅ − ⋅∑255
21
( )I k
h IcN= +
= ∑255
212
1 ( )( ) I k
h Im Ic k N= +
= ⋅ ⋅∑255
2 22 2
12
1 ( )( )( ) I k
h II mc k N
σ= +
= ⋅ − ⋅∑
Proposed Algorithm
b. Reject those seed candidates that meet at least
one of the following rules:
-One of its neighbors is a contour pixel
-One of its neighbors has and also has a neighboring contour pixel
-One of its neighbors has , and a neighboring contour pixel
1( , )I x y m<
1 1 1( , )m I x y m σ< < +2 2
2( , )x yσ σ>
The contours in the image are obtained with the SUSAN algorithm:
Proposed Algorithm
The intensity values inside a circular mask are compared with the intensityvalue in the center (nucleus). If there isa low amount of pixels inside the maskwith an intensity value similar to theintensity value of the nucleus, then thenucleus is a contour pixel.
Proposed Algorithm2) Region growing
a. Add a pixel to a seed unless it meets at least one of thefollwing rules:
-It is a contour pixel.-It has and a neighboring contour pixel.-It has , and a neighboring contour pixel
1( , )I x y m<1 1 1( , )m I x y m σ< < + 2 2
2( , )x yσ σ>
Intensity image Binary image with segmented background
b. Repeat step a. until no more pixels can be added tothe background region
Proposed Algorithm3) Invert the binary image
Binary image withsegmented background
Inverted binary image
a. Invert the binary image with the segmented background.
b. Apply a 5x5 median filter to the inverted binary image.
Reliability
Experiments in 30 different BHK cell images:
*BHK: Baby Hamster Kidney cells
Experimetal Results
100%30Proposedalgorithm
66%20Espinoza’salgorithm
Percentage ofimages with
most cell clusterscorrectly
segmented
# images withmost cell clusters
correctlysegmented
Algorithm
Espinoza’s algorithm segmentation result
Proposed algorithm segmentation result
Low contrastintensity image
Precision
2 2 2( ) ( )i i id x y= ∆ + ∆
n
dMSE
n
ii∑
== 1
2Mean square positionerror of the boundaries:
Experimetal Results
55,10%1,8 pixel24,02 pixel2
Improvement
AverageMSE for
the proposedalgorithm
AverageMSE for
Espinoza’s algorithm
Manual segmentation Espinoza’s algorithm segmentation result
Proposed algorithm segmentation result
Distance from each boundary pixel i, i:1…n, of the automatically segmented regions to the closest boundary pixel of the manually segmented regions:
Conclusions
The proposed segmentation algorithm is
• 33% more reliable• 55% more accurate
than the Espinoza’s algorithm in low contrast images
Thank you very much for your attention
