Cell migration is a fundamental process that is essential to life and it is linked to many important physiological and pathological events such as immune response, wound healing, tissue differentiation, embryogenesis, and tumour invasion [1
]. Modern microscopy with image capturing devices is at the heart of techniques to study live-cell migration, such as those driven by chemotactic signalling during directed migration [8
]. The advent of high-throughput, high-resolution microscopy, and imaging techniques means that cell outlines and migration trajectories or pathways may be visually accessible and manually traceable by a human operator (through manual tracking) to result in what is known as the gold-standard ground truth.
With the ever-increasing use of microscopy, a humongous amount of experimental data is routinely generated in many laboratories. Most of these data are collected in the absence of a monitoring segmentation and tracking algorithm. Due to the sheer magnitude of the data involved, manual tracking is often cumbersome and the development of computer algorithms for automated cell tracking is, thus, highly desirable [12
]. In contrast, a noticeable downside of fully-automated tracking is that the recognition capability of computers is very limited (comparing to a trained human operator) despite the sizeable efforts made from the wider academic community [16
Most of the research focuses on building fully-automated algorithms that post-process experimental outcomes without investigating the procedures underpinning data generation. The algorithms are thus blind to the underlying properties associated with quality data acquisition and a key parameter is the time step interval within the microscope. We have not explored any other parameters, since we feel these do not correlate directly with the cell tracking algorithm. Studies of this nature could be undertaken in a similar fashion. It is here that the novelty of our contribution lies, we strongly believe (and will demonstrate) that the time step interval embedded within the microscope camera setting at which experimental images are recorded is a crucial parameter that should influence the accuracy of the computer algorithms more drastically. The common choice of the time step interval between consecutive and successive camera shoot images currently published in the literature are in an ad-hoc trial-and-error approach that is generally based on the intuition and experience of microscopist or experimentalist. Typically, 5 and 10 min time step intervals are among the most common choices, for example, [21
By using a novel approach for fully-automated algorithms for cell tracking, we investigate viable options during experimental microscope data acquisition to allow for accurate and high-quality acquisition of images, thereby contributing to fully-automatic big data analytics in cellular biology. Here, we focus on different choices of the time step interval within the microscope. Taking into account the drawbacks that arise from using computer algorithms (i.e., lack of superior pattern recognition and error-handling capability compared to its human counterpart), one may propose to take constant and frequent sequences of images with an exceptionally short time step interval (e.g., a few seconds apart or even multiple frames within a second). However, when an image is taken, light is shed on the cells for clear visualisation. Yet, large doses of light have the potential to damage or even kill the cells under experiment [23
]. Recent studies show that yeast cells that appeared healthy directly after irradiation with a very low light-dose from a fluorescent microscope failed to divide when left overnight, whereas their non-imaged neighbours divided normally [27
]. In fact, it has been pointed out that fixation, plasma membrane permeabilisation, and cytoskeleton destruction are among the typical light-induced cell damages in live-cell microscopy [24
Evidently, a trade-off emerges on the choice of the time step interval and the quality of the images collected. Images acquired with a short time step interval would be beneficial for cell tracking algorithms, since cells will be much closer to their previous locations, but may be too harmful to the cells, as discussed above. On the other hand, a sequence of images obtained with a larger time step interval might result in less accuracy in cell tracking, since the cells could have moved further away and even out of focus or might have gone through large morphological changes.
Hence, the aim of this study is to investigate the relationship between the choice of the time step interval and the accuracy of the cell tracking algorithm, within the context of a fully-automated positive feedback loop between the quality of tracking cells and the quality of data acquisition. To the best of the authors’ knowledge, such a study has not been carried out. The majority of cell tracking algorithms published have only considered experimental data as static and given information. Here, we conduct specific experiments with known outcomes (i.e., the direction of cell migration) by either using an effective chemoattractant or employing no-chemoattractant. By acquiring images at 30 s time step intervals, we generate relatively large datasets on which we employ our cell tracking algorithms. To avoid generating extra data at say 1-, 2-min time step intervals, we use the same dataset generated with the 30-s time step interval and simply read out results at other larger time step intervals, to form different subset datasets with different time step intervals. We then evaluate the accuracy of our existing cell tracking algorithms to illustrate the effects of these different time step intervals. We establish that there exists certain fundamental relationships between the tracking accuracy and parameters that are associated with experimental microscope data acquisition.
Our fully-automated approach differs from the current-state-of-the-art cell tracking methodologies in that we investigate the dynamic interplay between segmentation and tracking (which are the hallmarks of the cell tracking algorithms) and the variable adaptive time step interval for generating quality image datasets from the microscopes. We recommend an adaptive approach, where, if the images being generated by the microscope for a particular time step interval are detected by the algorithm to be of less quality (below a certain threshold, hence poor cell tracking), the time step should then be reduced by some factor, automatically (on the fly). Conversely, if the images that are generated by the microscope at a particular time step interval are of high quality (above a certain threshold, hence too high quality cell tracking), then the time step interval can be increased by some factor, automatically. This positive interplay between the quality of the segmentation and cell tracking and the time step interval naturally leads to an adaptive optimal time step selection, such that the quality of the images is appropriate for quality segmentation and tracking and vice-versa.
This research provides a new perspective and benchmark results for the cell tracking community. Multiple experiments involving directed and random cell migration are used to capture live-cell migrations with two different cell lines in order to carry out these investigations. We include both induced (i.e., chemotactically driven) and non-induced experiments that result in directed and random cell migration, respectively. Phase-contrast microscopy is chosen to record the experimental data. This type of microscopy is generally considered to be less harmful to cells, since the intensity of lighting is an order of magnitude lower than other types, such as fluorescence microscopy. Besides the non-destructive light intensity, phase contrast is also less harmful to cells than fluorescence microscopy, because it is a label-free method [29
]. We select a number of time step intervals and evaluate each result against expected cell migration pathways and the corresponding manual tracking indices, respectively. We first select two representative algorithms [30
] to test automatic cell tracking and then consider a new algorithm that is based on an improved segmentation method.
We analyse the four datasets described in Section 2.1
and summarise our findings in this section. The only effective chemoattractant is in Dataset 2, where the chemoattractant is placed at the south side of the view plane. Thus directed cell migration between north and south is expected from this dataset. On the other hand, we expect cells to adopt a random-walk from the other three datasets between north and south directions. Furthermore, all four datasets should show cells adopting a random-walk between east and west directions.
We allow cells to move freely in and out of the boundaries of the view plane; thus, the number of cells in the experiments varies. In addition, a noticeable increase to the cellular area is observed and it is likely due to developed filopodia towards the end of the experiments. We use the confluences [39
], i.e., the ratio between areas of cells and total view plane, to evaluate cell density. We summarise the details of each individual experiment below.
- Dataset 1:
Total number of frames: 2880 (24 h), confluences vary between 7.30–15.02%, random-walk pattern is expected.
- Dataset 2:
Total number of frames: 2880 (24 h), confluences vary between 17.26–48.37%, directional migration towards south, random-walk pattern between east and west is expected.
- Dataset 3:
Total number of frames: 2761 (23.01 h), confluences vary between 4.59–12.44%, random-walk pattern is expected.
- Dataset 4:
Total number of frames: 2566 (21.38 h), confluences vary between 6.42–14.94%, random-walk pattern is expected.
For the first round of tests, we employ all frames (i.e., using a 30 s time step interval) from each individual dataset to track the cells. The second round of tests includes a 1-min interval, meaning we take every second frame in each dataset in our tracking algorithm and completely ignore the intermediates. Thus we only use half of the total number of frames. Tests are further repeated with 2, 3, 5, 7.5, 10, and 15 min intervals. For completeness, the choice of D within our tracking algorithm described in the previous section is 10 pixels (roughly around m) for the 30-s interval and gradually increases to around 50 pixels (approximately 83 m) for the most infrequent time step interval (i.e., 15 min). The minimum threshold on the duration of a cell trajectory is set to be of the experimental duration. In all cases, this duration is roughly h, which means a trajectory continually lasting at least h is counted as acceptable otherwise it is rejected. There is no upper bound threshold, so a trajectory may theoretically last the full length of the experiment, but this is neither practically common nor possible.
For illustrative purposes, we use the results from edge-detection-active-contour segmentation. We showcase some of the results graphically in Figure 3
. We plot all the tracked trajectories of Dataset 2 from four different time step intervals: 30 s, 5, 10, and 15 min, respectively. These trajectories are shown on both white backgrounds (Figure 3
(a1–c1)), as well as overlaid on the first image (Figure 3
(a2–c2)) of Dataset 2 to give the reader an alternative perspective on the interplay between segmentation and cell tracking. We observe that, as the interval enlarges, less trajectories are found by the cell tracking algorithm. The length of trajectories as well as the accuracy reduce substantially. When considering the standard 10 min setting for manual tracking, we see that this choice of time step interval plays a much more vital role in determining the quality of computerised cell tracking that is different to the manual cell tracking gold-standard approach.
In Table 1
, we show all results from the four datasets obtained when using the edge-detection-active-contour segmentation approach. The first column from the left indicates the different choices of the time step intervals. The second column shows the total number of trajectories found, which is equal to or greater than the required duration (i.e.,
of the total duration of the experiment). The third column shows the number of trajectories whose ending points are on the southern side of the starting reference point. The percentage between south moving and the total number is presented in the fourth column. In the fifth column, we show the number of east moving trajectories and their corresponding percentage in the sixth column. To fully evaluate the performance of our algorithms, we compare these tracked trajectories from our automatic cell tracked results to their corresponding counterpart results obtained manually from a trained human operator. We employ a strict approach, where, if a mistake is made (e.g., tracked to a different cell), we consider such track as a failure to provide accurate information. The details of this comparison are illustrated in the seventh column.
It can be seen from Table 1
that the total number of trajectories decreases as the time step interval gets large. This is demonstrated by all four datasets that confirm our observation, as illustrated in Figure 3
. We further observe that the majority of percentages from the random-walk pattern vary around
with an exception of
, which comes from a 15 min time step interval with a low identification count. In addition,
of figures lay between 40.0–60.0%. Whereas, for directed cell migration, all of the figures show a tracking accuracy of more than
. This result shows that fully-automated algorithms are capable of determining the patterns from experiments of chemotaxis. In terms of high accuracy, only within the 3 min time step interval does the algorithm achieve an accuracy of
We further analyse two selected datasets (Dataset 2 and 4) using an alternative segmentation technique that is based on background reconstruction and substitution, as described earlier, in order to demonstrate that patterns are independent of the choice of the technique. It is worth noting that, because of the extra fragments, the tracking accuracy from this segmentation technique is known to be worse. We show the results in Table 2
. We can see the accuracy within the 3 min time step interval is only
or above. On the other hand, we see that the method is still good enough to distinguish directed cell migration from a random walk. It can be observed that the former stays above
in all figures and the latter varies between 40.0–61.5%, with only one exception of
To graphically illustrate the difference between directed migration and random walk, we use the results from two datasets (Dataset 1 and 2). A star-plot format is employed here, which overlays all of the starting points of each trajectory to
and only links this reference point to the ending point of each trajectory. The results from eight different intervals from these two datasets are shown in Figure 4
and Figure 5
In terms of efficiency, any test case presented here can be fully-automatically analysed within 40 s on a moderate PC with GHZ Intel processors. Whereas, for the human operator, it could take between one to several hours to complete the manual tracking on one data test set.