SGN

9406 Signal Processing Graduate Seminar IV
Image segmentation and dynamic lineage
analysis in single

cell fluorescence microscopy
Quanli Wang, Jarad Niemi, Chee

Meng Tan, Lingchong You and Mike West
Muhammad Farhan (210911)
Outline
Introduction
Image preprocessing and hybrid image generation
Filters for hybrid images
Iterative mask

based cell segmentation
Cell tracking and lineage reconstruction
Results
Comparison and discussion
Conclusion
2
SGN

9406 Signal Processing Graduate
Seminar IV
Introduction
Time

lapse
movies
and
analysis
of
dynamics
of
gene
expression
at
single

cell
level
.
Challenge
is
automatic
segmentation
and
tracking
of
individual
cells
.
No
universal
algorithm
applicable
across
cell
types
.
Major
goal
in
algorithm
design
and
s/w
development
:
Set
of
tools
for
extracting
cell
like
objects
from
many
different
types
of
images
.
3
SGN

9406 Signal Processing Graduate
Seminar IV
Introduction (Contd..)
Problems with existing algorithms:
◦
Initial
cell
identity
and
location
marker
needed
for
segmentation!!
◦
Tracking
by
minimizing
global
energy
functions
typically
don’t
consider
locally
clustered
cells
.
Solutions to problems
◦
Iterative
extraction
of
cells
by
using
hybrid
grey

scale/black

white
images
i
.
e
gradual
conversion
of
grey

scale
image
to
black

white
mask
.
◦
Using
neighborhood
cell
information
to
compute
numerical
likelihood
scores
for
cell
identity
between
each
pair
of
consecutive
time
steps
.
◦
Generation
of
frame

frame
correspondences
between
cells
and
lineage
map
.
4
SGN

9406 Signal Processing Graduate
Seminar IV
Introduction (Contd..)
Commercial
s/w
for
general
purpose
image
analysis
:
Imaris,Amira,Volocity,Metamorph,Matlab
Image
processing
toolbox
.
Public

domain
tools
contain
Image

J
for
general
purpose
image
analysis
whereas
CellProfiler,
Cell

ID,CellTracker
and
GoFigure
are
for
segmentation
and
tracking
in
more
specific
areas
.
5
SGN

9406 Signal Processing Graduate
Seminar IV
Image preprocessing and hybrid image
generation
Grey

scale
image
is
divided
into
background,border
and
undecided
.
◦
Modified non

linear range filter and dilation help identify background.
◦
Extended high

pass filter identifies border regions.
◦
Undecided regions need to be further classified.
Special
meaning
to
largest
values
in
grey

scale
image
creates
the
hybrid
grey

scale/black

white
image
.
Modification is required in normal filters to work for hybrid images.
◦
Masks are used to exclude background pixels in border identification.
6
SGN

9406 Signal Processing Graduate
Seminar IV
Filters for hybrid images
A class of mask

based filters can be designed for hybrid images
Suppose, a hybrid image is given by
H = (I,M)
with
I
being grey

scale image and
M
being the black

white mask. If a function
H’ = F(H) = F(I,M) = (I’,M’)
also gives a
hybrid image, then
F
is a hybrid filter if
F(I,M) = F(I.*M , M)
, where
.*
is element

wise product
Hybrid range filters(HRF) is a specific class of hybrid filters given by
where is the given image pixel, is the structuring element and
where is the set intersection opertator.
With extension of above definition, hybrid mean,median or rank
filters can be designed.
7
SGN

9406 Signal Processing Graduate
Seminar IV
Iterative mask

based cell segmentation
Hybrid image is processed with combination of hybrid filters to
erode the undecided regions, followed by hybrid dilation and
smoothing.
Iterative processing is performed to gradually change the
associated mask until a stopping rule is met. Final mask is the
resulting segmented image.
Cell object modeling and selection at the end of each iteration
enhances the above obtained result.
8
SGN

9406 Signal Processing Graduate
Seminar IV
Iterative mask

based cell segmentation
(Contd..)
Disconnected segmented regions
–
referred to as blobs
–
in the
binary mask are
labeled
, evaluated, scored.
All blobs are classified as cells or undecided,
based on the object
shape model,
where undecided ones are subject to further
segmentation.
Useful especially when the cells exhibit diversity in scale, e.g., cells
differing in diameter, in which case cells at different scales will be
picked up at different iterations.
9
SGN

9406 Signal Processing Graduate
Seminar IV
Cell tracking and lineage reconstruction
Two

step
algorithm
:
Construction
of
score
matrices
and
Optimization
algorithm
to
obtain
the
frame

to

frame
correspondence
matrices
.
Calculation
of
geometric
overlapping
scores(>
0
.
2
is
valid
here)
for
the
pair
to
construct
forward
and
backward
scores
.
A
cell
i
is
in
the
neighborhood
of
a
cell
j
if
they
overlap
at
least
once
if
we
shift
cell
i
within
given
pixels
p
.
Computation
of
a
pair
of
neighborhood
scores
over
the
neighborhood
of
each
cell
using
the
overlapping
scores
.
Repeatedly
shifitng
one
image
slightly
and
recalculating
overlapping
and
neighborhood
scores
for
cell
pairs
.
10
SGN

9406 Signal Processing Graduate
Seminar IV
Cell tracking and lineage reconstruction
(Contd..)
Combination
of
score
matrices
and
transformation
into
a
0

1
correspondence
matrix
.
◦
Dot
product
of
score
matrices
to
get
combined
score
matrix
.
◦
Calculation
of
row
and
column
maxima
in
combined
score
matrix
.
◦
Locating
values
that
are
both
row
and
column
maxima
and
assigning
matches
for
corresponding
cell
pairs
.
◦
Use
of
backward
score
matrix
for
selection
of
two
best
scores
for
assigning
the
matches
.
Reconstruction
of
cell
lineage
trees
by
linking
the
sequence
of
correspondence
matrices
.
11
SGN

9406 Signal Processing Graduate
Seminar IV
Results
12
SGN

9406 Signal Processing Graduate
Seminar IV
Results(Contd..)
13
SGN

9406 Signal Processing Graduate Seminar IV
Results(Contd..)
14
SGN

9406 Signal Processing Graduate
Seminar IV
Results(Contd..)
15
SGN

9406 Signal Processing Graduate
Seminar IV
Results(Contd..)
16
SGN

9406 Signal Processing Graduate
Seminar IV
It
is shown that the neighborhood score based tracking method
provides a much sharper score matrix with substantially improved
cell correspondences.
The
algorithm reconstructs almost 98% of actual tracks from the
last frame back to the first frame in the movie, as compared to 58%
by the one in
Sigal
et al.
The
analysis was a direct application of the automated method
applied to the data with no additional parameter tuning.
Robust approach because of using common features and avoiding
modeling
cell growth transformation.
Advantageous to use hybrid images and filters.
Comparison and discussion
17
SGN

9406 Signal Processing Graduate
Seminar IV
One particular implementation of overall strategy and resulting
algorithm is presented.
Improved segmentation results, or more efficient algorithms, may
be achieved by tailoring the workflow to more specific features of a
data set.
The
emphasis is on robustness and portability.
Conclusion
18
SGN

9406 Signal Processing Graduate
Seminar IV
QUESTIONS ?
Thank You!!
19
SGN

9406 Signal Processing Graduate
Seminar IV
Comments 0
Log in to post a comment