Processing of 2D Electrophoresis Gels?

Andrés Almansa2 , Mijail Gerschuni1 , Alvaro Pardo1?? , and Javier Preciozzi1

1
DIE, Facultad de Ingenierı́a y Tecnologı́as, Universidad Católica del Uruguay
2
INCO, Facultad de Ingenierı́a, Universidad de la República.

Abstract. Here we present results on 2D electrophoresis gel image process-

ing using methods that provide a measure of meaningfulness. This work
is part of an ongoing project on biomedical image analysis. Biomedical
sciences have a long tradition in our country and therefore there is ex-
tensive experience in this area. Unfortunately, economical restrictions
sometimes do not allow our researchers to have the latest technology.
As we will see later, 2D electrophoresis gels are an extremely useful and
affordable technology, which can be used in our countries.
From the image processing point of view we believe that there is room
for improvements when thinking in the final software application. With
this in mind we present an algorithm that covers all the steps for gel
image registration. First we present a method for robust and meaningful
detection of spots. Then we study two improvements on the computation
of the distance between spots using shape contexts. Finally, we present
an iterative random sampling process which deals with spot differences
between images to give a gel registration. Across all the steps we address
the easy interaction with the user based on a measure of meaningfulness
of the results.

1 Introduction

This work is part of an ongoing project on biomedical image analysis. One of

the goals of this project is to study methods and algorithms for image analysis
that provide, together with their results, measures of confidence. That is, we are
not only interested in obtaining automatic results with probably some errors;
we look for a measure of the confidence on them. Even in the case when the
algorithms turn to be semiautomatic, this measure of confidence will be decisive
to guide the user towards the issues that need human intervention.
In this paper we address the processing of 2D electrophoresis gel images.
Two-dimensional electrophoresis is a well known method for protein separation
which is extremely useful in the field of proteomics. The basic idea is to separate
proteins contained in a sample using two independent properties such isoelectric
point and mass. In Fig. 1 we show an example of the kind of images obtained.
Each spot in the image represents a protein accumulation and its size depends on
?
Supported by Proyecto PDT-S/C/OP/46/18. A. Pardo is on leave from Facultad de
Ingenierı́a, Universidad de la República.
??
Corresponding author [email protected].
the amount of protein present in the sample. The grayscale on top of each image
is placed to allow grayscale calibration. Although it may seem a simple task,
the manual processing of this kind of images is very cumbersome. Furthermore,
since gel electrophoresis is generally used to compare samples, we need to process
several images in a single experiment. For this kind of differential analysis we
need to register two images finding the spot correspondence. One of the reasons
of the popularity of 2D gel electrophoresis is its simplicity. As a counterpart,
the experimental setting and the materials used do not allow a highly controlled
experiment. This means that strong variations between corresponding sets of
spots are expected. All these elements show that, although 2D gel images may
seem simple, the complete task of individual spot matching and gel registration
is a complex and time consuming process.
Most of the existing methods for gel matching start extracting point features
which represent the spots. These point features are then used for point and/or
gel matching. In some cases this point features can be used to establish point
correspondence before obtaining the complete gel matching [9]. This can be done
for example using shape contexts [3]. In [9] the authors review several feature
distances and matching procedures, and devise an iterative algorithm to perform
the gel registration (based on iterative closest point and graph matching). Some
works also include image correlation at some stage to refine the results.
Some of the methods reviewed propose a simple spot detection algorithm
which gives a large number of false positives[8]. This is then solved during the
matching step with the application of robust methods that discard outliers. We,
on the other hand, reject this approach as it produced, in our experiments, too
many outliers. We propose a more conservative detection of the spots to increase
robustness and reduce computation complexity.
With respect to the use of image correlation we observed that it works for
synthetic images, when the same image is compared after deformation, but is
does not work in real cases. Real pairs of images not only have strong differences
between corresponding spots but also several spots may share the same elliptical
appearance.
Finally, although in [9] the authors criticize the use of semiautomatic meth-
ods, no matter how precise the methods could be, for this kind of applications
some user validation will always be performed. For this reason we present meth-
ods which provide information that can be easily used by the operators.
The contributions of this work are the following. First, we present a ro-
bust method for spot detection based on level lines that provides a measure of
meaningfulness. Second, we include two modifications to the use of shape con-
text which clearly improve the results. Third, we present an iterative sampling
process which deals with outliers. After the last two steps we compute a couple
of matrices that give the similarity between each pair of spots. These matrices
can be also be used by the user as a measure of meaningfulness to validate the
results.
 (a) (b)

(c) (d)

(e) (f)

Fig. 1. (a,b) Low complexity pair of images. (c,d) High complexity pair of images. (e)
Maximal meaningful boundaries detected of image (d). (e) Resulting shapes obtained
applying the isoperimetric relation, over the lower level sets of the image and enters of
the spots on the obtained shapes.
2 Spot Detection using Level Lines
The process to establish correspondences between two images must be based
on invariant features present on them. Due to the nature of the generation of
the gel images explained before, features as the shape or intensity of the spots
may vary between experiments, turning them useless. One of the features that
remain invariant is the relative distribution of the spots within the image. This
is the reason why most of the existing methods are based on point-matching
approaches, where the points to be matched are, in general, the center of the
spots. Before applying such an algorithm, we must obtain for each image, the
set of points to be matched.
The proteins submitted to an electrophoresis process suffer deformations from
an initial circular or punctual concentration that turns them into ellipses. Given
a set of spots, the points can be obtained as the darkest point of each spot. The
problem then is how to obtain the set of valid spots of the image. On this work
we propose to detect them using the meaningful boundaries [5] together with
several criteria related to the shapes of the spots. We will start in next section
with a review of the meaningful boundaries approach.

2.1 Meaningful Boundaries

Meaningful boundaries are based on two concepts: the level lines (connected
components of topological boundaries of level sets) of an image [10] and the a
contrario models [4]. In fact, the meaningful boundaries of an image are obtained
by selecting from all of its level lines, those which are meaningful (that is much
unexpected) under the a contrario model. Lets recall both concepts:
Definition 1 (Level Sets). Given an image u : Ω(⊂ R2 ) → [a, b](⊂ R), we can
define for any λ ∈ R the (lower) level sets of u as χλ (u) = {x ∈ Ω, u(x) ≤ λ}.
In the same way we can define the upper level sets. Using both, lower and upper
level sets, the topographic map of an image is defined as the collection of all of
its level lines. Level sets constitute a hierarchical representation of the image
which can be encoded in a tree structure, tree of shapes [2], to represent the
topographic map. Since we are only interested on the detection of dark spots
from now on we concentrate only on lower level sets and their corresponding
level lines.
A contrario models. Computational Gestalt Theory was first presented
by Desolneux, Moisan and Morel [4] as a way to obtain a quantitative theory of
the Gestalt laws. Gestalt theory [6] states that visual perception is a grouping
process where geometric objects are grouped together by similar features or
gestalts (color, size, shape, etc). Although the Gestalt theory is consistent from
a qualitative point of view, it lacks of a quantitative way to determine when a
set of objects have the same gestalt. Computational Gestalt uses the Helmholtz
Principle to define a quantitative measure of a given gestalt [4].
Helmholtz Principle: The observation of a given configuration of objects in
an image is meaningful if the probability of its occurrence by chance is very small.
Therefore we ask the following question: is the observed configuration probable
or not in our model?. If not, this proves a contrario that this configuration is
meaningful. The Helmholtz principle can be formalized by the definition of the
Number of False Alarms and the definition of an ²-meaningful event:
Definition 2 (Number of false alarms - NFA and ²-meaningful event).
Given an event of the type “a given configuration of objects has a property”, the
number of false alarms (NFA) is the expectation of the number of occurrences
of this event under the uniform random assumption. An event E of the type “a
given configuration of objects has a property” is ²-meaningful if the NFA is less
than ²: NFA(E) < ².
Since the NFA defined in Def. 2 is in general very difficult to obtain in most
practical cases an upper bound definition is used instead. From now on we will
use the following re-defintion of NFA.
Definition 3 (Number of false alarms - NFA). The number of false alarms
(NFA) of an event E is defined as: NFA(E) = N · P[E ≥ E|H1 ] where N is the
number of possible configurations of the event E and H1 is the background or a
contrario model.
Meaningful Level lines. In [5], a definition of meaningful boundary is pre-
sented, taking into account the topographic map of the image and the framework
reviewed above. Using the norm of the gradient as a measure of the contrast, the
authors propose the following ²-meaningful boundary definition of a level line C:
N F A(C) = Nll P [min(|Du(x)|)]l/2 < ²
x∈C

where where Nll is the number of level lines in the image, l is the length of C
and P (x) is the probability of the contrast x under the a contrario model.3
A final observation is that on real images, boundaries width are bigger than
one pixel, which leads to the detection of several meaningful boundaries for each
real one. This problem is solved by applying a maximality criterion over the set
of all meaningful boundaries of a monotone section of the tree of shapes.

2.2 Meaningful Spots

Fig. 1(e) shows the result of the meaningful boundaries detection algorithm
applied to a real gel image. We can note from the results that most of the
spots are detected. Nevertheless, some problems arise. Shapes that correspond to
several spots: In many cases, several spots are grouped into a bigger one, given as
a result, instead of a boundary for each spot, a unique boundary for the whole
shape. Non detected spots: Some of the spots are not detected as meaningful
boundaries. These problems may impact on the point-matching process. In order
to solve this problem we must include characteristics of the spots we are looking
for. We will address this point in next section.
3
The power of l/2 comes from considering every second pixel’s contrsat (or gradient
magnitude) on the level line as independent from the others. The a contrario model
is obtained by using the empirical distribution of the gradient magnitude |Du| as
the probability distribution.
2.3 Adding spots characteristics
We already mentioned that we consider only lower level sets in order to detect
dark spots. Since most of the spots have an almost elliptical shape we propose
to filter the shapes obtained with the meaningful boundaries method using the
isoperimetric ratio p2 /a where p is the perimeter and a is the area of the shape.
In the case of a circle this relation has a value of 4π.
Although the inclusion of this feature improves the results, some problems
are still present. In fact, the problem of detecting the spots that are contained in
a bigger one remains unsolved. Let us analyze this case in detail. If several spots
are joined on a bigger one two things could happen. If the shape of the grouping
has also an elliptic shape it will be detected by the algorithm. On the other
hand, if this shape is not elliptical enough to be detected by the algorithm no
shape will be detected and some spots will be missed. The first case is shown at
Fig. 2, where several spots are grouped on a bigger one with elliptical boundary.
In this case, since the criterion to obtain the center of the spot is to keep the
darkest pixel, we obtain the center of one of the spots. Nevertheless, the other
spot is lost and no information could be obtained from it. In the second case, no
inner spot is taken into account, and all this information is lost. In next section
we show how to find spot correspondence.

3 Spot Matching
The best methods for spot matching are based on point-matching techniques.
In our case we use the metric Shape Context (SC) [3] following [9] where this
metric was applied to gel images. The idea behind SC is to describe each point
(spot) with the distribution of points on its neighborhood. Using a set of bins
in polar coordinates the number of points in each bin is computed to obtain a
two-dimensional histogram in polar coordinates. We will denote the normalized
histogram at point i as hi (k) where the index k identifies the bin. Given this
metric we can compute the distance between the SC of two points i and j using
the χ2 distance:

1 X [hi (k) − hj (k)]2

dχ2 (SCi , SCj ) = .
2 hi (k) + hj (k)
k

When comparing two SC we must recall that small discrepancies between

corresponding points may be encountered. These discrepancies have different
sources. First, there are genuine differences due to the appearance or disappear-
ance of spots. Second, the misdetection or errors in spot detection. Third, the
gel deformation. To deal with these discrepancies we propose two modifications:
a kernel estimation of the histograms and a different metric to compute the
distance between SC.
To add robustness and generalization capabilities to the SC we added a kernel
estimation of the histogram. This improves the results when small discrepancies
are encountered between both images.
Fig. 2. Extract of a gel image where we can see that two spots are detected on a bigger
shape that contains them. We can see also that the center of one of the spots is well
detected.

The χ2 distance it is sensitive to small variations in the histograms. In [7] the

authors show how to improve the SC matching using the Earth Movers Distance
(EMD). The EMD is a measure of similarity between probability distributions
(histograms) that copes with variations. Given two probability distributions S
(supply) and D(demand) the EMD distance is the transportation cost from the
supply S to the demand D given a cost of transportation between bins: CTij .
Hence the EMD distance between S and D is the minimum effort to transport
one probability distribution to the other (For details see [7]). The application of
EMD to our problem is relatively straightforward. The probability distributions
are the two dimensional SC normalized histograms. The cost of transportation is
computed as linear in the radial and circular axis. That is, the distance between
bins equals the Euclidean distance between their centers.
At the end of this step of spot matching we obtain a matrix C where each
entry Cij = d(SCi , SCj ). Therefore, for each spot in one image we obtain the
similarity with each spot in the other one.
To evaluate the performance of both modifications we performed a set of
simulations with landmark data obtained from [1]. We used 19 pairs of corre-
sponding images with 399 spots total. In Table 1 we present the results with
and without a kernel based estimation of the SC using χ2 and EMD distances.
We show the number of mismatched spots before and after gel registration. The
number enclosed in parenthesis corresponds to the results of spot mismatched
after gel registration using the method we will propose in section 4. As we can
see, the inclusion of the kernel estimation consistently improves the spot match-
ing results. When using χ2 distance the kernel estimation greatly improves the
results. This shows that certain amount of flexibility (generalization) must be
allowed in order to capture the intrinsic variations of spots between images. For
the same reason, since EMD already allows some discrepancies between SC, the
improvements of the inclusion of the kernel estimation while using EMD are not
so impressive. Based on these results we select the number of angle bins, nθ = 24,
the number of radial bins, nr = 5, and outer radius of the neighborhood.

nθ = 12 nθ = 16 nθ = 24 nθ = 12 nθ = 16 nθ = 24
Kernel EMD 6 (6) 7 (5) 5 (0) 4 (6) 5 (7)
EMD 9 (6) 6 (14) 5 (11) 4 (9) 5 (5)
Kernel χ2 8 (6) 8 (5) 10 (0) 10 (6) 11 (5) 11 (6)
χ2 24 (6) 21 (14) 25 (11) 23 (9) 19 (7) 29 (15)

Table 1. First three columns: results for outer radius two times the average spot
distance. Last three columns: results for outer radius four times the average spot dis-
tance. In each cell we show the number of mismatched spots before and after (between
parenthesis) gel registration. The last column for EMD is not computed due to its
computational cost.

4 Gel Registration

Once we have a list of potential correspondence candidates for each spot we

must solve the problem of gel registration. That is, if possible, finding a unique
correspondent for each spot. A first approach to obtain such matching could be
to find a global set of correspondences which minimize a global merit function.
For instance, we could use the method applied in [3]. Given a matrix of similarity
between spots, Cij , the idea is to find the optimal assignments to minimize the
total cost of matching: X
min Cij Pij (1)
Pij
ij

where Pij is a permutation matrix which encodes the matching. Since we may
have outlier spots we also include a set of virtual spots with cost ε for rejection
purposes. Obviously the problem with the above procedure is that no global
coherence is imposed.
As discussed before when pursuing gel registration we have to deal with
two types of errors. First we have differences in the spot sets (either genuine
or produced during spot). Second, the aforementioned spot differences produce
discrepancies in the SC used for spot matching and gel registration. To over-
come this problem we developed an iterative random sampling procedure. At
each iteration we randomly sample a subset of spots from both images and com-
pute the SC and the corresponding distances between pairs of spots. Along this
process, for each pair of corresponding spot we record the smallest distance in
a matrix, Cij , and the number of times each pair is matched in a matrix Nij .
We use these matrices to compute the gel matching using (1). For each spot in
one image we obtain a set of possible corresponding spots. Using Cij and Nij
we obtain a measure of confidence of each pairing. This can be used to assist the
user in the rejection of false matches and refinement of the registration solution.
No matter how robust the automatic methods could be, there will be always
potential mistakes that would need high level information to be resolved. That
is why we equipped every step of the method with a measure of confidence to
rapidly assist the user.

5 Results and Conclusions

To test the methodology here presented we divided image pairs in three groups
of: low, medium and high complexity. In Fig. 3 we show the results. For both
pairs we have the ground truth of corresponding points. In the first example
of low complexity, in Fig. 3(a), we detected 36 spots in each image and the
ground truth contains 29 pairs. The proposed registration process finds 26 correct
pairs, 1 erroneous pair and two points with no correspondent. In the second
example of medium complexity, in Fig. 3(b), we have 61 spots in one image, 56
in the other one and 45 corresponding pairs in the ground truth. The proposed
registration process finds 44 correct pairs and 1 erroneous pair. As we can see
the global results are extremely accurate for low and medium complexity pairs.
Unfortunately these results are not achieved in high complexity examples. In the
example showed in Fig. 3(c) only 25 out of 56 correct pairs are found from 39
spots without correspondent the method correctly finds 22. The lack of global
constraints and the complexity of these pairs are not resolved with the proposed
method. In this case the strong differences between gels conspire against the
results. Although, some of the erroneous correspondents are close to the valid
ones, the actual matching is incorrect. This is a clear difference with other works
which report the error in pixels. At the end of the day we may have small error
but a huge number of incorrect matching to be resolved by the user. Our method
intends to overcome this problem and the results for the first two examples seem
promising. Below we discuss possible improvements to cope with high complexity
pairs.
The spot detection using meaningful boundaries with the addition of the two
criteria explained on this article gives good results. Most of the spots present on
the images are correctly detected and a very low number of them are missed.
Furthermore, we can obtain the meaningfulness for each spot which allows the
user to supervise, if needed, the less confident ones. In the future we will ad-
dress the problem of grouped spots. We expect that the joint inclusion of spots
features, instead of in sequential order, will allow us to consider a level line to
be meaningful if it is contrasted enough but also has an isoperimetric relation
similar to the circle.
The use of a kernel based estimation of the SC histogram and the EMD
distance showed to improve the results.
The iterative random sampling process gives good results despite it does not
include global coherence. Furthermore, the distance matrix, Cij , and the matrix
with the number of matches, Nij , can be used to decide the confidence on the
correspondence of pair of spots. We are currently investigating how to include a
more detailed validation of the matching and registration transformation using
the theory of Computational Gesltalt. Also, we are exploring the inclusion of
constraint on the permutation matrix Pij to reject invalid matchings.

References
1. http://www.lecb.ncifcrf.gov/2dgeldatasets/.
2. C. Ballester, V. Caselles, and P. Monasse. The tree of shapes of an image. Technical
report, ENS, 2001.
3. S. Belongie, J. Malik, and J. Puzicha. Shape matching and object recognition
using shape contexts. IEEE Trans. on Pattern Analysis and Machine Intelligence,
24(4):509–522, April 2002.
4. A. Desolneux, L. Moisan, and J. Morel. From Gestalt Theory to Image Analysis. A
probabilistic Approach. Springer “Interdisciplinary Applied Mathematics”, 2007.
5. A. Desolneux, L. Moisan, and J.-M. Morel. Edge detection by Helmholtz principle.
Journal of Mathematical Imaging and Vision, 14(3):271–284, 2001.
6. G. Kanizsa. La Grammaire du Voir. arts et sciences. Editions Diderot, 1997.
7. Haibin Ling and Kazunori Okada. An efficient earth mover’s distance algorithm
for robust histogram comparison. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 29(5):840–853, May 2007.
8. M. Rogers, J. Graham, and R.P. Tong. 2 dimensional electrophoresis gel reg-
istration using point matching and local image-based refinement. In BMVC04,
volume 2, pages 567–576.
9. M. Rogers and M. Graham. Robust and accurate registration of 2-d electrophoresis
gels using point matching. IEEE Trans. on Image Processing, 16(3):624–635, March
2007.
10. J. Serra. Image analysis and mathematical morphology. Academic Press, 1982.
 Corresponding Spots

Warped image 1 over image 2

50
30 36
23 34
32 28 33 141315
191718 12 28 100
31 23 2019 9 22 29
1516 2122 30 36 18
17 31 35
20 29 35 16 27 150
2413 34 1121 33 8
1412 27 10 32 200
25 26 11
25
7 26 24 250
8
109 7 3 5 6
6 5 4 300
4
3 350
2
400

450
2
1 1
(a)
500
50 100 150 200 250 300 350 400 450 500

Warped image 1 over image 2

50
60 54
61 494847
14 56 50
39 38
29 46
100
20 44
19
13 25 22 10 32
23
12 24
261823 41 56 22 33
21
20
2816 2127 42 11 55
150
31 55 241912
1415 47
464 8
17 51 11 17 37 45
15 37 52 50 53 18
16 30 27 43
59 200
303835363334 57
58 53 2526 54 49
32 10 1331 29 52
45 4039 9 34 35 28
36 250
40 51
300
86 7 4443 5 42
7 8 41 9
350
4 3 6 5
2 4 400

3 450

2
(b) 1 500
1 50 100 150 200 250 300 350 400 450 500

Corresponding Spots

31 117 115
87 53
86 52
5436 3534 66 22
5133 59 85 95 32 78
48 7172 68 41
84 46
47
44
43
4542 69 70 84
22 83 94 40 42
70 1921
4140 60 82 96 67
7775 76
2016
79 17 383958
37 56
81 74 93 116
18
80 49 26 55 57 39 4948 73 88
90 105
78 6561
152725 38 554647 53 104
548789 106
103 4
14
13 30 7776 90 37 44 52
43 51
50 107108
75 2829 2467 23 353645 6264101 9899 102
100
93 92 63
6160 119
69 507172 113 11133
110
112
65 118
127468 73 32 59 58 86 5695 114
31
81 57 3097 10925
64 98 63
10 62 11
83
80 799185 2928
82
7 91
4 65 10 27 26 2321 24
92 20
89 34 3 2 1 7
3 88 11 15
14 18
16 17 8 65
94 1312
66 2 19
1 9
(c)

Fig. 3. Results for low (a), medium (b) and high (c) complexity pair of images. (a, b)
Left: corresponding pairs. Right registration of both images with thin plate splines. (c)
Corresponding pairs.