- Single pixel errors in mouse click position can add up to cause significant errors in track length
- A lack of standardisation causes substantial variability between users and poor repeatability
- Livecyte's automated tracking removes these limitations by standardising the tracking process and consistently measuring cell centration
Cell motility is a vital metric when characterising many physiological and pathophysiological processes. During wound healing studies it is essential in identifying the collective migration of cells . Similarly tracking immune cells can give vital information on their migratory activity during an immune response . Therefore, it is critical to be able to evaluate cell migration in vitro objectively and precisely .
The most common method for directly observing and characterising migration is time-lapse microscopy. An important task in many time-lapse experiments is the tracking of large numbers of cells and the analysis of their dynamic behaviour. The plot in Figure 1 shows the increasing interest in cell tracking in biological science (and related) literature .
Manual tracking using a point and click system is the most common method for measuring motility of multiple cells over time. Whilst manual tracking is commonly deployed throughout the time lapse, it is time and labour intensive, suffers from inter-operator variability, ill-defined cell centroid positioning, and an intrinsic lack of morphological data . In many cases, the vast number of cell images collected during a time-lapse means only a subset of cells are tracked within a population leading to a poor approximation of migration rates. Multiple available tracking tools offer a certain level of image pre-processing and background filers which may also perturb tracking measurements from one user to the next, depending on the method of tracking used .
Automated tracking allows for analysis of large time-lapse data sets to truly understand and analyse cell behaviour in an efficient, reproducible and statistically robust way . Most simple automated tracking approaches, however, are dependent on high contrast images (as seen in fluorescence) where cells may be segmented by thresholding, i.e pixels above an intensity threshold are seen as cell and the rest is background.
Studies comparing automated and manual tracking approaches have shown some preliminary success with highly customised automated techniques. Yet a previous study comparing manual and automated tracked populations of pancreatic cells showed the manual approach to be highly variable and lead to a miscalculation of migration rates by up to 410% . However, in a separate paper the trajectories of manually tracked clusters of epithelial cells showed close agreement with an automated approach; the differences being within the limits of user variability. Further to this, the automated algorithm was estimated to have saved an approximately 320 person hours of labour .
Although published automated approaches may overcome some of the challenges of manual tracking, many are specifically designed algorithms which have been optimised for a unique purpose. This kind of tracking approach can suffer in the case of inconsistent fluorescence intensity or imaging artefacts and therefore a more a sophisticated solution is needed .
In this study we sought to illustrate and understand the root causes of manual tracking inaccuracies on a fundamental level and how using computer algorithms can eliminate them. Such automation has the added benefits of substantial time savings and removing the subjective nature of human interaction.
High-contrast quantitative phase images were automatically captured using the Livecyte Kinetic Cytometer. Cells were imaged with an Olympus PLN 10X (0.25NA) objective and 1 mm x 1 mm field of view (FOV) per well, for 24 hours, at 12-minute intervals. Cells were maintained inside an environmental chamber at 37°C with 5% CO2/95% humidity.
11 researchers from 4 different academic institutions volunteered to participate in our cell tracking study. They were provided with quantitative phase images (tif. stacks) generated by Livecyte and asked to track the same 5 cells manually using ImageJ's MTrackJ plugin, repeating each track 3 times. The 5 cells were chosen purposely for having significantly different motile behaviours from each other.
Afterwards, participants completed a questionnaire detailing the amount of experience they have with manual cell tracking, the software packages they use, the point on the cell they try to click on when tracking, the time spent tracking each cell in this assay and whether they used a mouse or touch pad.
Livecyte's Analyse software has easy-to-use algorithms that automatically segment and track all cells in a field of view. Whilst these algorithms can be customised to allow for tailoring to a user's specific requirements, the automatic tracking data we present here was generated using the default parameters.
Analysis & Results
The 11 volunteers tracked the paths of 5 cells across 121 frames using ImageJ (Figure 2). Most volunteers aimed to click on the brightest part of the cell, with the intention of tracking the cell nucleus. On average, it took the volunteers between 53 seconds (cell 5) and 98 seconds (cell 1) to track one cell through all 121 frames. This study highlighted some of the inconsistencies often associated with manually tracking cells over a period of time. Figure 2 shows two of the five cells as an example. In figures 2a) and 2d), example frames from the time lapse show the clicked position from one user (repeated three times giving the red, yellow and green data points). The inconsistency in the exact pixel on which a participant clicked can be clearly seen, hence introducing an error in determining the centre of the cell. We found that between repeats every volunteer showed variability in the click position. Even small, one-pixel errors in position per frame can accumulate over the entire track, leading to a difference of over 150µm in track length in 24 hours. This can have profound implications for the derived cell track length and velocity measurements.
We also sought to quantify the effect of this click position error on motility metrics that are commonly derived from manual tracking data. We determined, for each cell at each frame, the median click position from all participants. This enabled us to quantify a click position error for each click (such as those shown in figure 2a & d) as the deviation from this median position. A compilation of click position errors highlights the potential disparity between participants tracking the same cells (figures 3a & b). Errors in tracking inevitably lead to inaccuracies in track length and hence the cell motility measurements derived from this. This was demonstrated when analysing the variability in track lengths between the two participants (figure 3c).
For both Participant 1 and 3 there was little variability between repeats, with a percentage variability of 3% and 5% respectively. However, the difference in determined mean track length determined by each participant was much higher; between 19% and 31% for different cells. This could cause incorrect data interpretation which in turn could limit the sensitivity of treatment effects or give rise to inconsistent and unrepeatable published data. The full extent of this variability can be seen when expanding this to consider all participant data, as shown in figure 4. We calculate a mean variability by considering the standard deviation of the manual track lengths and expressing this as a percentage of the median manual track length, per cell. Averaging these standard deviations across the five example cells gave a mean variability of 20%.
Furthermore, differences in user functionality of manual tracking software such as ImageJ mean there is no standardised procedure for manually tracking cells. For instance, some software tools snap the clicked position to the brightest nearby pixel within a user defined distance. However, this gives no guarantee of accuracy as the brightest pixel may be a different physical feature in different frames. In addition, the inconsistencies of whether these added tools are utilised or not can increase user-to-user variability. Finally, we also observed differences in how users handled cell division events, many but not all just picking one of the daughter cells to follow after the event.
Why Automated Tracking?
Tracking cells using computer vision algorithms represents a different approach, designed to remove the variability arising from a human doing the task. Such algorithms seek to identify and follow individual cells and determine their position in a traceable and repeatable way.
There are generally two hurdles to overcome when performing automated tracking; the first is the recognition of relevant objects and their separation from the background in every frame (the segmentation step), and the second is the association of segmented objects from frame to frame and making connections (the linking or tracking step) . Livecyte can automatically perform both these steps by utilising its high contrast quantitative phase imaging (QPI) to reliably segment single-cells, with varying size and morphology, and track them from frame to frame using Phasefocus’s uniquely designed and patent-pending tracking algorithm. QPI also has the advantage of having contrast proportional to cell dry mass. The position of each cell is based on calculated centre of mass, leading to a highly consistent and traceable centroid determination.
Manual Vs Automatic Tracking
>Due to the self-consistent and invariant nature of Livecyte's automated tracking algorithm, it will always track the same cell in the same way, every time, for a given parameter set. This removes variances between repeats and is independent of the user conducting the experiment. As well as being consistent, we demonstrate in this study that it is also accurate.
In order to quantify track length variability, we defined a mean manual track length and a median-click track length. The mean manual track length was calculated by taking the mean of all 33 track lengths for each cell (three repeats from eleven participants). The median-click track length was determined by first calculating the median position of all clicks from all participants for each cell on each frame. This has the effect of determining a cell's position by the median of all eleven participants, three times each, per frame. The median-click track length is then calculated based on a track formed by these median-click positions.
Figure 5 shows the mean manual (blue) and median-click (orange) track lengths, as compared with automated (yellow) track lengths. We found that there was considerable variability in the mean manual track length when considering all participant data (a standard deviation of 20% of mean manual tracking length). Interestingly, when the click positions were averaged across repeats and participants, the resulting median-click track length was consistently lower for all cells than just the mean of the manual track lengths. We attribute this to the additive nature of errors and the subsequent "smoothing out" of centroid position in the median-click approach. The average percentage difference between the median-click track length and the mean manual track length is 18.9%.
In comparison, there was much better agreement between the median-click track length and the automated track length. The automated tracking algorithm employed by Livecyte uses a calculation of cell centre of mass for centroid determination. This is inherently a very stable measure of cell location as it is obviously independent of any human-introduced subjectivity and is robust to the variability of brightest pixel determination. Additionally, the average percentage difference between the median-click track length and the mean automated track length drops to just 4.6%.
Figure 1 Number of publications in PubMed database as a function of publication year with the words "cell tracking" in the title and/or abstract.
Figure 2: Illustrative phase images of two of the five cells used in the study. In figures a) and d), example frames from the time lapse show the clicked position from one user (repeated three times giving the red, yellow and green data points). This variation in click position is repeated over multiple frames subsequently altering the track length seen in b) and e). The click positions of all participants are shown in figures c) and f). The degree of error is escalated further when multiple participants track the same cell.
Figure 3: 3a and b show the spread of clicks for each repeat of five cells tracked across all frames; the origin is calculated as the average position from the three repeats for each cell in each frame. The click spread then shows the deviation (error) per click from this averaged position. Person 1 shows a large spread in where they click between repeats whereas person 3 click spread is less variable. 3c) shows the resulting calculated track lengths for both participant 1 (blue) and 3 (orange) for all five cells (repeated three times each) highlighting the resulting differences between users, but also repeats of the same cell.
Figure 4: The graph shows all the manual track lengths from all participants, measured for each cell. The three repeats for each participant are indicated by similar colours. Across all participant data, for the five example cells, the mean standard deviation of track lengths was 31% (expressed as a percentage of mean track length).
Figure 5: Graph showing the mean manual track length, mean-click track length and the automated generated track length. Mean manual track length, an average of all track lengths for each cell showed an increased margin of error. The mean-click track length, however used the mean click position of all participants for each frame to produce a track length. The automated track length was calculated using Livecyte's Cell Analysis Toolbox. (CAT). The mean-click track length and CAT track length showed good correlation for all of the tracked cells.
Time course cell imaging enables the dynamic phenotypic behaviour of individual cells to be investigated. The ability to individually track a single cell throughout an entire time lapse enables several morphological and dynamic phenotypes to be extracted. This information gives a far more accurate and in-depth analysis of the cell behaviour compared to single time point and/or a population averaged approach. However, it is practically impossible to manually follow many hundreds of cells through many tens to hundreds of image frames accurately, and sophisticated computerised methods are needed to perform these tasks.
In this study we showed how automated tracking can outperform manual tracking through removing the subjective nature of the manual approach, whilst also allowing the possibility of more cells being tracked in a far quicker manner. Specifically, we demonstrate that 11 participants tracking the same five cells, three times each, produces an average variability of 20% of the mean track length. Averaging all the participant clicks for each cell at each time point in order to reduce the position click errors, we compute a "median-click" manual track. Comparing this to both the mean manual track and Livecyte's automated tracking algorithm shows an average difference in track length of 18.9% and 4.6% respectively highlighting the reduced variability in Livecyte's automated approach.
Livecyte presents clear advantages to a researcher in the accuracy of derived motility measurements, as well as saving a considerable amount of time and energy.