Commentary: Tracking Biological Cells in Time-Lapse Microscopy
Sijie Wen
Z5185708
Introduction
In this paper, the author proposed a new method to solve
the difficulty in detection and tracking cells in time-lapse
microscopy.
In the field of cellular structure researching, tracking cells
to see how they change with time is an important method in
researching, since they could clearly see how the cellular
structure changes under certain conditions and discover new
findings. This leads to the development of time-lapse
microscopy; this is the tool to record the behaviour of cells
over time. However, even with such tool, it is still
challenging for scientists to track down cells, because this
means that they have to record the whole process and track
cells in real-time and also have to stay around the
microscopy to make sure that they did not lose track of cells,
this is a huge workload for scientists and could make
mistakes easily. This leads to the research of automated
cellular structure tracking, by using this technology,
scientists could be relieved from the tedious work of
detection and tracking.
Nonetheless, this technology is not perfect for now, it
may wrongly detect cells because of overlapping or mitosis
and many other unexpected reasons, thus the algorithm
would lose track of cells and lead to cumulative error.
Previous methods did not solve these problems perfectly, in
this paper, the author proposed a pipeline model that uses an
adaptive technique and topological features to improve the
performance of detection and tracking, they also tested it on
the datasets to shows their improvement compare to the
previous method.
Methods
In this part, we will give a brief discussion about the
proposed model which includes his basic theory.
Additionally, we also simply discuss some possible
improvements that can be done in this paper.
The detection model used in this paper is based on the
morphological top-hat filter [1] to correct illumination error
and h-maxima transformation [2] to split cells. Because in a
real scenario, a cell is a dark region surrounded by a bright
halo. This leads to low contrast between the cell and its
background thus making it hard to detect. In order to solve
this problem, the paper used a top hat filter to correct the
illumination error around the cell and remove the shading
artifact so that the cell is clearly represented. After removing
the illumination of cells, we need to segment these cells,
because different cells may have overlap or under the process
of mitosis. To detect, the author discovered that the cell
center is slightly brighter than its boundary, thus can
transform this problem into a simple form, which we just
need to suppress the local maximum which corresponds to
the boundary. To do this, the paper first used a Gaussian filter
to generate a unique intensity maximum inside each cell and
also remove some noise, because noisy data might influence
the detection result, these preprocessing steps could benefit
the finding of the cell center. After the filter, the author used
an h-maxima transformation to suppress local maxima and
find regional maximum, where h is a parameter computed
based on several feature parameters of the cell. In order to
reduce the effect of irregular boundaries of cells, they also fit
an ellipse to the result of every detection, where the ellipse is
found by using the Least Square method. H-maxima
transformation used in here is a very nice choice because it
originates from the feature that found by the author and is a
simple and clear method to solve this problem efficiently,
additionally, in the original edition, the h is a hyper
parameter that is set by trial, however, in this paper, the
author proposed a new method the adaptively calculate the h
value.
Another important aspect would be tracking, in order to
track, you need to identify whether a cell in a consecutive
frame is the same cell. In this paper, they introduced several
topology features for cell, like displacement, skewness area,
and so on. They use these features to describe certain
properties of a cell thus could help in tracking cells since
these properties may not change drastically over time. Then
they use these features to match cells, they did it in a graph
way, they compute the match cost of different cells thus
having graph-like data, where the data of the table represents
the cost of a cell matching another cell. They take this as a
bipartite graph, then they use the Hungarian algorithm [3] to
find the best match.
After the successful match, we also need to recover the
trajectory of cells, because cells might leave, enter, or having
mitosis, thus the trajectory might be broken from this frame
to the next frame for a single cell. To recover the trajectory,
the author firstly assigns IDs to different cells and categorise
them into 3 types based on the appearance of ID on the
previous or consecutive frame. Then they generate an
elliptical matching template which is used to convolve with
frames, then it could classify the broken trajectory into 3
types that they described above, then they could recover the
trajectory based on the different broken trajectory types. For
the algorithm used here, I do not think this is a good option.
Because it just uses a template to match and find results
based on it, which is very rigid. Unlike many other things,
the structure of cells is evolving with time, therefore the
trajectory of cells is also related with time, which means
previous location influences future location, however, this
important factor is not considered in this part, since a
template matching method has no consideration between the
relationship of different frames. In this part, I think that the
Bayesian model could perform better because it can relate
future location and previous location through conditional
probability rather than a simple template matching.
Results
The result of this paper can be concluded into 2 parts,
first is the detection experiment, second is the trajectory
experiment and the performance-related with the number of
cells.
In the first part, the author tests the proposed model on
the detection. From the result, the proposed model could
obviously detect cells better than other algorithms. We could
see that it can generate more obvious boundaries than other
methods and thus having a clearer segmentation of cellular
structure. However, even the result seems nice, this
experiment has some problems, one of the problems in the
experiment is that the baseline method is too weak. The
baseline used in this detection experiment is only a simple
threshold method, which did not use any advanced
algorithms in edge detection like Soble or Laplacian operator
[4]. Outperform such easy baseline is not a surprising result,
in order to convince the reader, it needs to be compared with
other complicated methods to show the improvement, at least
comparing to such a simple baseline would not persuade me
to use this model in cell detection.
Another problem is that it did not give any quantitative
result in the cell detection, it only shows the qualitative test
example in this experiment, however, a quantitative result
could validate the author’s argument even more, since this is
a detection task, which gives an accuracy table is not hard.
However, such a result is not given in this part, which makes
this experiment results in less convincing.
The second experiment is tracking, it tracks 4 cells in this
experiment for consecutive 700 frames, we could see the
performance of tracking 4 cells is promising which reached
83% accuracy and cells after mitosis is also detected and
tracked, this is mainly due to the template matching method.
It uses both qualitative and quantitative measurement, for
quantitative, it gives the tracking accuracy that is calculated
based on the observed segments and valid segments.
However, one fatal flaw here is that it did not even compare
to any other tracking method in this experiment, not even a
baseline model was given in this part. It means we cannot tell
whether this model performs well or not since we do not
have any information about how other tracking models
would perform in this experimental setting.
The author also tests the model’s performance under a
different number of cells, however, one fatal flaw is also
about the choice of baseline, it mentions about the 2 baseline
model, but it does not give even any description about
different features of these baseline tracking models, thus we
do not know whether these modes are an advanced model or
old models. Another problem is that since it compares with
these 2 baselines, but why it did not compare it in the
tracking accuracy. This is a confusing problem that did not
have a proper explanation in this paper.
Additionally, another point is that in the accuracy table he
gives, there is no column of tracking accuracy when a cell
enters or leaves the view, however, in the paper, this is
described as one of the major reason that a broken trajectory
situation might happen, which makes me question whether
this model could handle this problem.
From the reasons that I concluded above, I will not use
his model because of a lack of reasonable comparison
between models and also the lack of proper experiments
about another situation.
Conclusions
In this article, we first introduced the background of
automated cell tracking and its importance to biology
research, then we have discussed some major difficulties in
this research area.
After the introduction, we give a brief description of the
paper’s proposed model, this model is in a pipeline manner
that combines different technologies to address detection and
tracking problems. In detection, it uses a top-hat filter and h-
maxima transformation to segment the cells then an elliptical
matching is applied to match cells. After detection, they use
different topology features to identity cells in different
frames in which the best match is optimised using the
Hungarian algorithm. Then this paper turns to address the
broken trajectory problem that might happen in tracking,
which is due to mitosis and entering or leaving of cells, this
paper used an elliptical template matching method to recover
the trajectory from frames.
We also had a simple analysis of the resulting experiment
of this model, in the detection task, the qualitative result
seems very promising which could generate more obvious
cell boundaries, however, this paper did not give a
quantitative result and the baseline is too weak. The second
experiment is in tracking, in this part, the paper shows it can
track cells with 83% accuracy, however, in this part, it did
not test on the entering or leaving situation and also did not
have a baseline model to compare with. Additionally, it tests
accuracy on the different numbers of cells with 2 baseline
models, but we know little about these 2 baselines thus we
could not conclude that the proposed model is whether state-
of-the-art or not.
The strength of this model is that it captures the topology
features of cells, which discovered many useful features to
detect or track, the algorithm is also simple and clear.
However, pipeline model could be a weakness, since a
pipeline model is complicated than an end-to-end model if
this model could be an end-to-end model, it could be much
better, another point that could be improved is in tracking, a
template matching is really not a good idea since it lacks
relatedness between different frames, this can be addressed
using conditional probability model like Bayesian[5] which
the current step can be established based on previous steps
using conditional probability.
For automated tracking, there are still several areas that
could be further researched. For example, in a frame, there
might be different types of cells, it could also benefit the
scientist if the algorithm could automatically classify cells in
the tracking process. If we could add cell classification into
automated tracking, it could also benefit scientists and help
them to do better research.
References
(1).An Introduction to Morphological Image Processing by
Edward R. Dougherty, ISBN 0-8194-0845-X (1992)
(2)Soille, P., "Morphological Image Analysis: Principles and
Applications" (Chapter 6), 2nd edition
(2003), ISBN 3540429883.
(3) Harold W. Kuhn, "The Hungarian Method for the
assignment problem", Naval Research Logistics Quarterly, 2:
83–97, 1955. Kuhn's original publication.
(4) Gilbarg, D and Trudinger, N. Elliptic partial differential
equations of second order. Springer.
2001. ISBN 978-3540411604.
(5) Gelman, Andrew; Carlin, John B.; Stern, Hal S.; Dunson,
David B.; Vehtari, Aki; Rubin, Donald B. (2013). Bayesian
Data Analysis, Third Edition. Chapman and Hall/
CRC. ISBN 978-1-4398-4095-5.