Synechocystis sp. PCC 6803 CELL IMAGES Committee …cbs/projects/2004_report_ko... · ·...
Transcript of Synechocystis sp. PCC 6803 CELL IMAGES Committee …cbs/projects/2004_report_ko... · ·...
FEATURE EXTRACTION FROM
Synechocystis sp. PCC 6803 CELL IMAGES
by
Shylaja Kokoori
Committee Members
Dr. Robert Roberson
Associate Professor, School of Life Sciences
Arizona State University
Tempe, AZ 85287- 1804
Dr. Rosemary Renaut
Director, Computational Biosciences Program
Arizona State University
Tempe, AZ 85287- 1804
Dr. Kenneth Hoober
Professor, School of Life Sciences
Arizona State University
Tempe, AZ 85287- 1804
ARIZONA STATE UNIVERSITY
August 2004
ABSTRACT
Synechocystis sp. PCC 6803 has been one of the most popular organisms
for genetic and physiological studies of photosynthesis because of its
capability of growth heterotrophically at the expense of glucose and
because of the availability of its entire genome sequence. Dr. Roberson’s
laboratory in the School of Life Sciences is conducting electro microscopy
studies using three- dimensional (3D) reconstruction of electron
tomographic data of wild type Synechocystis cells and selected mutants.
Their goal is to identify mechanisms of photosynthesis and thylakoid
membrane biogenesis in this cyanobacterium. Accurate extraction of
important inclusions from electron tomographic data, by segmenting
images and performing 3D analysis and modeling, will help in achieving
the goals. Currently, however segmentation in electron tomography is
almost exclusively a manual operation. As a result, it is often the most
time consuming and subjective step in the data analysis process. This
project is an attempt to automatically segment features such as ribosomes,
thylakoid membranes and cytoplasmic filaments from tomographic data of
Synechocystis cells. Morphological operations and mathematical methods
are used in order to provide biologists information in a timely manner.
II
TABLE OF CONTENTS
PAGE
LIST OF FIGURES IVChapter
1
GOALS OF PROJECT 1
Chapter
2
ABOUT Synechocystis 2
Chapter
3
INTRODUCTION AND OVERVIEW 3
Chapter
4
ELECTRON TOMOGRAPHY 7
Chapter
5
DATA PREPARATION 7
Chapter
6
DATA 8
Chapter
7
METHODS 11
7.1 First Order Differential Method 127.2 Convolution 137.3 Gaussian Smoothing 157.4 Seeded Region Growing 157.5 Watershed Segmentation 167.6 Eigen Value Analysis of Hessian
Matrix
18
Chapter
8
ALGORITHM 19
Chapter
9
TOOL DESCRIPTION 21
Chapter
10
IMPLEMENTATION 23
Chapter
11
RESULTS 25
III
Chapter
12
CONCLUSIONS AND FUTURE WORK 30
REFERENCES 32
IV
LIST OF FIGURES
PAGESFigure 1: Electron tomographic slice through the
Synechocystis cell
2
Figure 2: Histogram thresholding. (a)original image (b)thresholded image t=125 (using Vigra) and (c)Histogram bar plot of original image (using Matlab)
4
Figure 3: Edge detection (a) original image (b) after applying
Canny edge detector σ = 3, gradient threshold = 3 (using Vigra)
6
Figure 4: High magnification showing ribosomes 9Figure 5: High magnification showing thylakoid membranes 9Figure 6: High magnification showing filaments 10Figure 7: 1st and 2nd derivative of an edge illustrated in one
dimension
11
Figure 8: Synechocystis cell image after Gaussian smoothing
operation, σ = 3
13
Figure 9: Synechocystis cell image gradient after Gaussian
smoothing operation, σ = 3
13
Figure 10a: Example image and kernel to illustrate
convolution
14
Figure 10b: Image overlayed with kernel 14Figure 11: Watershed segmentation. (a) Gray level gradient of image data (b) Displays local minima and watershed region in the image data
16
Figure 12: Watershed segmentation on Synechocystis cell. (a) Original image (b) edge image after performing watershed transformation (c) after region growing (enlarged image)
17
Figure 13: Labeled region after watershed transformation 19
Figure 14: Thresholded image after watershed segmentation 20Figure 15: Extracted thylakoid membranes 21
V
Figure 16: Tool user interface (a) Displays dialog box to set parameters(b) Window displaying opened image file
22
Figure 17: (a) Original Synechocystis cell image (b) extracted ribosomes
25
Figure 18a: Manually- segmented model of a zoomed - in portion of the image
25
Figure 18b: Ribosomes detected using the tool overlayed on the real image
26
Figure 19: (a) Original Synechocystis cell image and (b) extracted thylakoid membranes overlayed on the original image
28
Figure 20: (a) Original Synechocystis cell image and (b) extracted filaments overlayed on the original image
29
VI
1
GOAL OF PROJECT
The goal of this project is to help understand the basic cell biology of
photosynthesis of the unicellular cyanobacterium Synechocystis sp. PCC
6803 by providing algorithms that enable automatic extraction of
inclusions such as ribosomes, thylakoid membranes and cytoplasmic
filaments from tomograms. Currently, segmentation is the bottleneck in
electron tomographic data analysis because it is almost entirely a manual
operation. This project is an attempt to design and develop a tool that
automatically segments inclusions in the cell by making use of
morphological operations and mathematical methods. Extracting features
from the cell in an accurate and timely manner for 3D analysis and
modeling, will aid with understanding the basic cell biology of this
organism and, specifically, the mechanisms of photosynthesis and
thylakoid membrane biogenesis. Second, the tool has to provide a friendly
user interface that will help the user to use the tool efficiently. The main
goal here is to develop software, which is reliable and expandable so that
in the future new features can be added effortlessly.
2
ABOUT Synechocystis :
Cyanobacteria are photoautot rophic organisms, which are capable of
performing oxygen- producing photosynthesis similar to plants [18].
Synechocystis sp. PCC 6803 is one such unicellular non- nitrogen - fixing
cyanobacterium, which is an inhabitant of fresh water [19]. This
cyanobacterium is capable of growing heterotrophically at the expense of
glucose, which has made it a desirable model organism for genetic and
physiological studies of photosynthesis. This organism displays a unique
combination of highly desirable molecular - genetic, physiological, and
morphological characteristics: it is spontaneously transformable,
incorporates foreign DNA into its genome by double- homologous
recombination (making gene knock- outs and replacements clear- cut), can
grow under many different physiological conditions (such as photoauto /
mixo/ heterotrophically), is small (~ 1.5 mm in diameter) and is suitable
for quantitative 3D ultra structural analysis. This, coupled with the fact
that cyanobacteria are closely related to the ancestors of chloroplasts,
makes Synechocystis an ideal experimental system [16]. Fig. 1 displays a
tomographic slice of Synechocystis cell containing inclusions such as
ribosomes, thylakoid membranes and cytoplasmic filaments.
3
Figure 1: Electron Tomographic slice through the Synechocystis cell. Scale bar = 100 nm
INTRODUCTION AND OVERVIEW:
Image segmentation is an important step in image analysis. It is, however,
a difficult step because it depends largely on the data and the application.
For a biomedical application segmentation may involve identifying the
shape and size of a tumor. For a geographic application the task may
include identifying the location of roads in an area or the location of weeds
in a lake. For tomographic data segmentation might involve extracting
inclusions for further analysis. The ultimate goal, generally, is to reduce
the given image to non- intersecting regions of interest [3].
The most common methods used to segment images are (a) histogram
thresholding, (b) edge based methods and (c) region growing methods [4].
HISTOGRAM THRESHOLDING:
This is one of the easiest methods of segmentation and is the ideal
method to segment objects from a distinct background. There are different
types of adaptive- thresholding methods available which work based on
4
maxima or minima of the histogram, or by making use of other histogram -
related functions. Since thresholding does not consider any other factor
like shape or position it is suitable only for simple images [4]. Fig. 2a, Fig.
2b and Fig. 2c show the original image, thresholded image at a grayscale
value of 125 and a histogram bar plot of the original image respectively.
Thresholding could be used to extract the ribosomes but largely depends
on the threshold value being selected. A high threshold value results in
eliminating some of the ribosomes from the image and a low threshold
value results in extracting noise and unwanted structures from the image.
It is thus not an appropriate choice for this project. Moreover, it does not
help in detecting thylakoid membranes because of the difficulty to identify
thylakoid membranes from the background.
( a) (b)
5
( c)
Figure 2: Histogram thresholding. (a)original image. Scale bar=100 nm (b)thresholded image t=125 (using Vigra) and (c)Histogram bar plot of original image (using Matlab)
EDGE BASED SEGMENTATION METHODS
Edges in an image are detected by making use of a combination of
methods such as convolution matrix- based operators and the Hough
transform. The methods section in this report (page 9) explains the
convolution operation in further detail. The Hough transform is a method
used to detect features such as lines, curves and ellipses, where the
desired feature can be expressed in a parametric form. The main advantage
of using the Hough transform is that it is relatively unaffected by either
the presence of gaps in the feature being identified or by the image noise
[4]. Thus, edge detection helps to identify boundaries of the primary
elements in an image. However, the presence of noise in an image may lead
to over fragmentation and some pale edges not being detected [4]. An
6
example of a classic edge detector is Canny edge detector. Steps in this
algorithm include
• convolving image with Gaussian of scale σ,
• computing x and y gradients
• finding peaks in the image gradient
• performing a threshold operation to remove unwanted
responses.
Fig. 3a and Fig. 3b show the original image and edge image after applying
Canny edge detection to the real image respectively.
( a) (b)
Figure 3: Edge detection (a) original image. Scale bar = 100 nm (b) after applying Canny
edge detector σ = 3, gradient threshold = 3 (using Vigra)
REGION GROWING
7
This is a very popular segmentation method and it gives good edge
detections in the case of noisy images. Initially the image is split to a large
number of smaller regions which are merged recursively into larger regions
based on different criteria such as homogeneity, or merging nearest
neighbors [6]. In this project, watershed segmentation, a special case of a
seeded - region growing, has been used to extract ribosomes from the
images. This method is explained in detail in the methods section (page
12).
In many of the low contrast, noisy images, one or more of these methods
need to be used to obtain good results. Also, in the case of noisy images,
performing preprocessing operations such as image smoothing before the
feature extraction process reduces the noise present in the image leading
to better results.
ELECTRON TOMOGRAPHY
Electron tomography is a method that helps determine the 3D structure of
cells and tissues at high resolution. Here, a series of images of an object
are taken at various tilt angles and they are combined together using the
back projection algorithm to produce a 3D structure of the object [17]. A
very useful feature of electron tomography is that it allows viewing the
structure in situ, which aids study of the structural arrangement. Electron
tomography is an important tool in studying macromolecules and cellular
complexes [11].
8
One of the main issues in biological electron tomography is that the
images obtained have a low signal - to- noise ratio mainly due to complexity
of the biological specimen. For example, a cell can contain numerous
organelles of different shapes and sizes. Another issue arises from the fact
that electrons have the potential to damage the biological specimen. An
idea, thus, is to use an electron dose such that it provides adequate
contrast to study the specimen without damaging the structure [2].
Improvements in instrumentation, and using better techniques to improve
data quality, have led to production of tremendous amounts of data
making it very challenging to mine the vast amounts of data produced
using electron tomograms. Therefore, developing tools that can obtain
information with minimal human intervention is becoming essential.
DATA PREPARATION:
Serial thick sections of 100- 300 nm of the Synechocystis cell are cut and
post stained using uranyl acetate and Reynold's lead citrate. Colloidal gold
particles are attached to the surface of the section and it is viewed using
an electron microscope. Tomograms are produced by taking images at
every 1.5 o at a range of tilt angles from - 60 o to 60 o. IMOD software
developed at the University of Colorado, Boulder, tracks the position of the
gold particles on the section to align and merge the images to create a 3D
model [18].
DATA:
9
IMOD is used to split the 3D model into a stack of two- dimensional (2D)
tiff images which serve as the input data for the program. Each two-
dimensional image is an RGB image of dimension 768*768 pixels and
having a size of 1.69 MB. The images contain inclusions such as ribosomes,
thylakoid membranes, and cytoplasmic filaments, along with various other
cell organelles. The task here is to extract the ribosomes, thylakoid
membranes and filaments.
Ribosomes are the prominent black structures, which appear to have a
circular shape in the image. Since they inherently have a 3D structure they
have varying diameters on a 2D image depending on which slice of the
section is being viewed. On an average, however, they have a diameter of
approximately 8- 10 pixels (Fig. 4).
Thylakoid membranes are curvilinear structures that are located nearer to
the cell wall having a width of approximately 12- 15 pixels (Fig. 5).
Filaments are also curvilinear structures which, however, appear
throughout most cell. They are of very small diameter and form a mesh
structure in the cell (Fig. 6).
10
Figure 4: High magnification showing ribosomes. Scale bar = 50 nm
Figure 5: High magnification showing thylakoid membranes. Scale bar = 100 nm
11
Figure 6: High magnification showing filaments. Scale bar = 50 nm
12
METHODS:
A major task in feature extraction process is edge detection. Most edge
detection methods work based on the assumption that edges occur where
there is an intensity gradient in the image. The strength of an edge
depends on the speed at which the intensity changes, the faster the change
in intensity the stronger the edge. Therefore, one of the methods to detect
edges is to identify step discontinuities in the image [7]. Identifying the
discontinuities by finding local maxima or minima from the first
derivative, or by finding zero crossings from the second derivative of the
image, are the most commonly used methods for edge detection (Fig. 7).
Enhancement and smoothing operations applied to the image can make the
edges more noticeable.
Function representing an edge
1 st derivative
Zero Crossing
2 nd derivative
Figure 7: 1st and 2nd derivative of an edge illustrated in one dimension (http:/ /ari.cankaya.edu.tr/ ~reza /ImLab4.htm)
13
First order differential method:
In a discrete image the gradient can be calculated by taking the difference
in the gray scale values of adjacent pixels [7].
The gradient of a two dimensional image I(x ,y) is given by the vector ∇Ι= [δ
Ι/δ x,δΙ/δ y] , which has magnitude √(δΙ/δ x) 2+(δΙ/δ y) 2 ) and
direction
tan - 1[(δΙ/δ y) / (δΙ/δ x)] .
One of the main issues associated with this method of edge detection is
the presence of noise. For most noise models large derivatives due to noise
are local events, however large derivatives of the signal can be present over
a larger area. This property can be used to reduce the noise. Thus, an
alternative to overcome this problem is to use an image smoothing
algorithm.
A two- step approach has been adapted for the images in the project:
• Convolve the image with a Gaussian mask,
Gs(x,y)= 1 exp[- (x2+y 2)/2 σ2]
2πσ 2
to smooth it.
• Calculate the derivative of the smoothed image.
Smoothing an image and then differentiating it is the same as convolving it
with derivative of a smoothing kernel [7].
14
Figure 8: Synechocystis cell image after
Gaussian smoothing operation, σ =3
Figure 9: Synechocystis cell image gradient after
Gaussian smoothing operation, σ =3
Convolution:
Convolution operations are very useful for a large number of image
processing operations like image smoothing and image enhancement. It is
15
a multi pixel operation where each output pixel is altered based on the
values of a set of the adjoining pixels, using a mask:
Figure 10a: Example image and kernel to illustrate convolution
Figure 10b:
Image
overlayed with kernel
The output pixel value is a linear combination of certain input pixel values.
The value of the bottom right pixel in the output image will be given by:
Kernel maskDigital Image
I11 I12 I13 I14 I15 I16 I17 I18 I19
I21 I22 I23 I24 I25 I26 I27 I28 I29
I31 I32 I33 I34 I35 I36 I37 I38 I39
I41 I42 I43 I44 I45 I46 I47 I48 I49
I51 I52 I53 I54 I55 I56 I57 I58 I59
I61 I62 I63 I64 I65 I66 I67 I68 I69
I11 I12 I13 I14 I15 I16 I17 I18 I19
I21 I22 I23 I24 I25 I26 I27 I28 I29
I31 I32 I33 I34 I35 I36 I37 I38 I39
I41 I42 I43 I44 I45
K1
1
I46
K12
I47
K13
I48
I49
I51 I52 I53 I54 I55
K2
1
I56
K22
I57
K23
I58
I59
I61 I62 I63 I64 I65
K3
1
I66
K32
I67
K33
I68
I69
K11 K12 K13
K21 K22 K23
K31 K32 K33
16
O57= I46 K11 + I47 K12+ I48 K13 + I56 K21 + I57 K22 + I58 K23 + I66 K31+ I67 K32+I 68
K33.
Convolution is basically a process where each pixel is replaced with a
weighted average of values of the neighboring pixels as defined by the
mask [8]. Convolving with a Gaussian mask can help filter out the noise.
Gaussian Smoothing:
Smoothing using a Gaussian kernel is easy because of the following
properties:
- Most of the optimal filters have a Gaussian like profile because a
smoothing filter must place a stronger weight on the pixels in the
center of the filter and lesser weight on those that are distant.
- Convolving a Gaussian with a Gaussian results in a Gaussian:
Gs1** Gs2=G √(σ1)2
+(σ2)2
)
Because of this re- smoothing a smoothed image will still result in a
smoother image.
- Gaussian kernel is separable:
Gs(x,y)= 1 exp[- (x2+y 2)/2 σ2]
2πσ 2
= [ 1 exp (- x2/2 σ2) * 1 exp (y2/2 σ2) ] ,
√ 2πσ √ 2πσ
- A product of two 1D Gaussians [10]. Thus convolving with a 2D
Gaussian kernel is equivalent to convolving with two 1D kernels.
This is important because separable kernels are very helpful in
reducing the computation cost. Fig. 8 displays the resulting image
after performing Gaussian smoothing on Fig.1.
Seeded Region Growing:
17
The step prior to seeded region growing segments the image to uniquely
labeled seed pixels and unlabeled pixels. The unlabeled pixels are then
assigned to closely matching seeded regions based on some fitness criteria.
Possible fitness functions are
- Fitness of the local gradients, so that the regions meet at the local
maxima
- The difference between the gray level of the candidate pixel and the
mean gray level of the region [1].
Watershed Segmentation:
This is a special case of seeded region growing where two neighboring
regions meet at local maxima. Here the pixels in the image are sorted
based on the grayscale value or on intensities. Local minima form the
catchment basins. This can be theoretically compared to a dam where
water starts filling in the basins and the water level rises. The point where
two catchment basins tend to intersect forms the boundary of the region
[1].
boundary
regions
18
(a) (b)
catchment basins(local
minima)
Figure 11: Watershed segmentation. (a) Gray level gradient of image data (b) Displays local minima and watershed region in the image data
One of the disadvantages of this method is that it can lead to over
segmentation. Fig.11 helps in explaining this issue. Oversegmentation
occurs because every local minima however small, forms a catchment
basin. A solution here is to ignore the catchment basins which are too
shallow.
(a) (b)
19
(c)
Figure 12: Watershed segmentation on Synechocystis cell. (a) Original image (b) Edge
image after performing watershed transformation (c) After region growing (enlarged
image)
Eigen value analysis of Hessian matrix:
The Hessian matrix is a matrix built from the second order partial
derivatives of the image [3] and contains information about the image
curvature, which is useful in identifying shapes in an image.
For a 2D image it is given as
H = Ixx Ixy Iyx Iyy where Ixy = ∂ 2 I , Ixx = ∂ 2 I and Iyy =
∂ 2 I . x y x ∂ ∂ ∂ 2 y ∂ 2
Note, Ixy = Iyx
For a 3D image it is given as
20
Ixx Ixy IxzH = Iyx Iyy Iyz
Izx Izy Izz
Here I represents a volume function I(x, y, z) [3].
Convolving the image data with derivatives of Gaussian kernel can be used
to perform computation of H, but this is a computationally intensive
process. The complexity can be reduced by making use of separable
Gaussian kernels (page 16) [10].
Given H for each pixel in the image, eigen values and eigen vectors of each
H can be calculated. Two eigen values λ1 & λ2 are obtained for 2D images
and three eigen values λ1, λ2 and λ3 are obtained for a 3D image. The
eigen vector corresponding to the largest eigen value represents the
direction of the curvature. In addition, if λ1 & λ2 are big then that pixel
represents a corner.
Algorithms given in the next chapter makes use of these methods to
extract the inclusions ribosomes, thyalkoid membranes and filaments.
21
ALGORITHM:
Ribosome segmentation:
An approach based on watershed transformation, a special case of a
seeded - region growing algorithm, has been used to segment ribosomes
which are the most prominent structures in the image.
Steps:
• Perform smooth operation with Gaussian kernel(σ =3)
• Find x and y components of the image gradient
• Transform components to gradient magnitude
• Find local minima of gradient magnitude
• Label the minima found
• Perform region growing using the minima points as seed points.
• Threshold the resulting image(t=125)
Figure 13 : Labeled region after watershed transformation
22
Figure 14: Thresholded image after watershed segmentation (t=125)
Thylakoid membrane & filament segmentation:
In order to improve the contrast between the curvilinear structures
(thylakoid membranes or filaments) from the rest of the image, local
geometric properties of curvilinear structures have been exploited, based
on eigen value and eigen vector analysis of the Hessian matrix.
Steps:
• Calculate second order spatial derivative Ixx, Ixy, Iyx and Iyy of the
image I(x,y), where Ixy = Iyx, by convolving the image with
derivatives of a Gaussian.
• Determine the hessian matrix, H, at each pixel.
• Find the eigen values and eigen vectors of H.,
• Largest eigen value and its corresponding eigen vector indicate the
strength and direction of the curve.
• Filter the pixels which meet the curvature property.
23
Figure 15: Extracted thylakoid membranes
TOOL DESCRIPTION:
The end product is an executable file which runs on a windows system via
cygwin[19].
Following tool(s) and libraries are essential for the application to function
effectively on a windows system
- Cygwin, provides a Linux- like environment for windows. It also
provides an x- window server which helps open the application
window [15].
- VIGRA computer vision Library, version 1.2.0 implemented by Ullrich
Köthe [12]
- Linear algebra library newmat, version 11, implemented by Robert
Davies [13]
The user interface for the tool was developed using Qt, a cross platform
development tool. Therefore, the source code should compile on UNIX
systems too but has not been tested.
The user interface provides menu options to Open, Close and Save Files.
Another menu item is Segment, which provides an option to segment
24
ribosomes, thylakoid membranes and filaments. In addition, a dialog box is
provided in which the user can set parameters such as the threshold if
they are not satisfied with the default results obtained.
Figure 16a: Tool user interface - Displays dialog box to set parameters
Figure 16b: Tool user interface - Window displaying opened image file
25
The tool also provides tool tips and shortcut keys, which helps the end
user in using the menu options easily.
26
IMPLEMENTATION:
Three programming languages were considered for this tool development.
- Matlab
- Python
- C++
A semi- automatic feature extraction tool was implemented initially in
Matlab. Matlab was chosen because it is a matrix oriented programming
language and it provides an image processing toolbox, which could reduce
the development time significantly. However, licensing Matlab is expensive;
moreover usage of control loops slows down the program significantly,
therefore, I decided to use Python.
Python is an interpretive object oriented scripting language[9]. It is a
programming language, which is freely available, and it provides many
extension modules like PIL (python imaging library) and
Numeric/numarray, which makes it a very desirable tool. Therefore, I used
Python to develop the code and the SDC morphology toolbox for the
morphological operations such as open, close, erode and dilate. However,
processing large images (of size 768*768 pixels) was computationally
intensive since Python is an interpretive language. For example, analysis
based on Hessian matrix to identify curvilinear structures took
approximately 10- 15 minutes per image on a normal Pentium 4 PC. The
solution here was to write parts of computation intensive code in C.
Qt a tool developed by Trolltech[14], was chosen to develop the user
interface for the program. Qt is a C++ application development
framework, which includes a class library and tools for cross - platform
development [14]. In addition, it is a sophisticated toolkit that is very
helpful in development of efficient user interfaces. However interfacing
Python with Qt was not very easy. Therefore, the entire code was ported to
C++ using libraries Vigra and Newmat. Both the libraries are freely
27
available, and both make use of the C++ standard template library, making
it generic and easy to use.
The final version of the project is written using C++. An image processing
library Vigra 1.2.0 has been used to provide the necessary image
processing functions and a library Newmat has been used to provide the
necessary Linear algebra support for the tool. On windows operating
system the tool is executed using cygwin, which provides a Unix like
environment on windows.
28
RESULTS
(a) (b)
Figure 17: (a) Original Synechocystis cell image. Scale bar = 100 nm (b) Extracted
Ribosomes
Figure 18a: Manually segmented model of a zoomed- in portion of the image (white circles represent ribosomes)
29
Figure 18b: Ribosomes detected using the tool overlayed on the real image
Fig. 17a and Fig. 17b show Synechocystis cell image and ribosomes
segmented from the image using the tool respectively. Verifying the
authenticity, by overlaying one over other using Adobe Photoshop, showed
that large percent of ribosomes present in the cell were correctly
identified. This fact was further confirmed by comparing the ribosomes
identified by the tool against manually segmented images as shown in Fig.
18a and Fig. 18b.
The table given below provides test results for some of the images tested
using the program. Various reasons are responsible for the difference in
the number of manually segmented ribosomes and those detected by the
program. The reasons could be classified into one of these categories:
30
- False Positives, result from wrongly identifying features, which are
not ribosomes, but have similar characteristics, such as intensity
value, like that of ribosomes in the image. Another reason could be
that they are part of ribosomes, which have not been identified in
the current slice, and will be observed in the subsequent slices.
- False Negatives
File name Number from manual segmentation
Number correctly identified by the program
False Positives
False Negatives
Zap010.tif(892 x 759 Pixels)
35 33 14 2
Slice1.tif(682 x 682 Pixels)
181 137 2 44
Slice2.tif(682 x 682 Pixels)
70 66 17 4
The tool was tested against many other slices and compared visually,
similar results were obtained.
31
(j) (j)
Figure 19: (a) Original Synechocystis cell image and (b) Extracted thylakoid membranes
overlayed on the original image
Fig. 16a and Fig. 16b show Synechocystis cell image and thylakoid
membranes identified by the tool superimposed over the original image,
respectively. As observed from the images most of the curvilinear
structures representing thylakoid membranes have been identified.
However some of the segments inside the cell have been wrongly
identified. In addition, there are discontinuities in some areas were the
structure is slightly faded. The algorithm needs to be modified to
accommodate this issue.
Figure (19)a and (19)b shows the segmentation of filaments. More work
needs to be done for filament extraction.
32
(k)
(b)
Figure 20: (a) Original Synechocystis cell image and (b) Extracted filaments overlayed
on the original image
33
CONCLUSIONS AND FUTURE WORK
This project focuses on ways to extract features from Synechocystis
bacterial cells. The dataset used here are tomograms of the bacterial cell
taken using an Electron Microscope. Two different approaches have been
used to detect the presence of ribosomes, thylakoid membranes and
filaments. A method based on watershed segmentation, which is a special
case of a region growing algorithm, has been used to detect ribosomes. A
comparison of segmentation results produced by the program against
results obtained from manual segmentation shows that the program has
identified a large percentage of the manually detected ribosomes.
Nevertheless, watershed segmentation is not very useful in identifying
curvilinear structures like thylakoid membranes and filaments because
these structures are of low contrast with respect to the background and it
is difficult to identify the structures from its boundaries. Therefore,
geometric properties of the curvilinear structures have been used to
extract them from the surroundings. However, while considering 2D
images there are areas where the curvilinear structure tends to fade out, in
such cases a 3D view might be of more help.
The program is capable of handling different file formats like jpg, png, tiff
depending on the availability of the corresponding libraries to read/write
these file formats. This program, however, has been tested only with tiff
image files. In addition, it has been tested only with 3- 4 files at one point
of time due to memory limitations.
The algorithms are very flexible and easily extendable to support other
requirements such as extracting other features or using for similar
applications. In addition, the source code for the image- processing library
and linear algebra library are available and they can be easily modified to
support additional requirements that are currently not provided by the
library.
34
This project addresses only segmenting ribosomes, thylakoid membranes
and filaments. Future research work should focus on identifying other
inclusions present in the cell besides these features. Developing a full-
fledged version of the program can significantly cut down the time
biologists spend on manual segmentation. One of the methods, which
could be used to detect other features present in the cell, will be to
perform a hierarchical segmentation by masking the segments that have
already been identified and try to extract new ones. More research needs to
be performed to eliminate False Positives as much as possible in the case
of filaments. Providing an option in which the user can edit the segmented
image and manually correct the errors would be very useful.
35
References
1) Ullrich Köthe , 1995. Primary Image Segmentation. Proc.17. DAGM-Symposium, Springer 1995
2) A. Bartesaghi, G. Sapiro, S. Lee, J. Lefman, and S. Subramania. 2003. A new approach for 3D segmentation of cellular tomograms obtained using three - dimensional electron microscopy. Institute for Mathematics and its Applications, December 2003 Preprints #1950
3) Adam Huang. 2003. Three- Dimensional Biomedical Image Segmentation And Visualization: A Shape- Based Approach. Ph.D. thesis. Arizona State University.
4) David Neary. 2000. Fractal methods in Image Analysis and Coding. Ph.D. thesis. Dublin City University School of Electronic Engineering.
5) D. Ballard and C. Brown. 1982. Computer Vision. Prentice- Hall.
6) Rolf Adams, Leanne Bischof. 1994.Seeded region growing. IEEE Trans. on PAMI. Vol.16: 641 - 647
7) D. Marr and E. C. Hildreth. 1980. Theory of edge detection. Proceedings of the Royal Society, London B. 207:187- 217
8) Rafael C. Gonzalez and Richard E. Woods. 1992. Digital Image Processing. Addison- Wesley Publishing Company.
9) Alex Martelli . 2003. Python in a Nutshell.
10)Karank. 2002. Edge detection. chapter 9.
11)A.J. Koster and J. Klumperman. 2003. Electron microscopy in cell biology: an integrated view on structure and function. Supplement Nature Reviews Molecular Cell Biology 4:SS6–SS10.
12)Vigra. Home page. <http: / / kogs - www.informatik.uni -hamburg.de / ~ koethe /vigra / >
13)Davies, Robert. Home page. <http: / /www.robertnz.net / index.html>
14)Qt. Home page. <www.trolltech.com/ >
15)Cygwin. Home page. <http: / /www.cygwin.com / >
36
16)Dr.Robert Roberson, E- mail to author, 16 April, 2004.17)O'Toole, E.T., Winey, M. J., McIntosh, J.R., Mastronarde, D.N. 2002.
Electron Tomography of Yeast Cells. Meth. Enzymol. 351: 81- 95
18)Kaneko T, Tabata S., Whitmarsh J. 1997. Complete genome structure of the unicellular cyanobacterium Synechocystis sp. PCC6803. Plant Cell Physiol. 38(11):1171- 6.
19)Flaubert Mbeunkui. 2003. The effects of low nitrate levels on the freshwater cyanobacterium Synechocystis sp. strain PCC 6803: Construction of a bioreporter assay and molecular characterization by transcriptome and proteome analysis. Ph.D. thesis. University of Stuttgart.