EllipTrack: A Global-Local Cell-Tracking Pipeline for 2D Fluorescence Time-Lapse Microscopy

Background

Biological processes are highly dynamic and display varying degrees of cell-to-cell heterogeneity. Time-lapse microscopy provides the opportunity of tracking and monitoring single-cell dynamics. However, rapid cell migration, high cell density, and treatments or modifications that alter cell morphology and/or behavior make it difficult to track single cells. Multiple tools for cell tracking have been generated. A classical tracker usually consists of three steps: segmentation, in which nuclei (or other structures) are identified; track linking, whereby nuclei (or other structures) are mapped throughout the time-lapse; and signal extraction, whereby fluorescent signals are extracted from regions of interest in each cell to obtain relevant biological quantitative data.

Over the last decade, various efforts have focused on improving the accuracy of each step, however tracking remains challenging due to the high accuracy needed for analysis. Often, manual verification and correction is still required, which limits the experimental throughput. An important breakthrough achieved in cell tracking was the creation of the global track-linking algorithm by Magnusson and colleagues (1). This algorithm uses machine learning tools to infer the probability of cell overlapping, cell migration, and biological events such as mitosis, apoptosis, etc, and then uses the Viterbi algorithm (a dynamic algorithm that aims to compute the most probable path) to iteratively search for, and assemble, the cell track that results in the greatest increase in the probability of existing cell lineages, until this probability can no longer be improved.

However, the use of Viterbi algorithm is a source of tracking mistakes, which Magnusson et al (1) fixed by introducing a swap operation whereby the algorithm examines whether swapping a cell track with another existing one will increase the overall probability of cell lineages. While the algorithm introduced by Magnusson et al increases the accuracy of cell tracking in a small-to-medium scale, it is only partially useful in very high-throughput experiments spanning multiple days, multiple rounds of replication and high cell densities. In the present work, Tian and colleagues (2) address this shortcoming and present EllipTrack, a global-local cell-tracking pipeline optimized for tracking cells in complex time-lapse-derived movies.

 

Key findings and developments

EllipTrack implements a global track-linking algorithm to construct tracks that maximize the probability of cell lineages, and then corrects tracking mistakes with a local track-correction algorithm (module) whereby tracks generated by the global algorithm are systematically examined and amended if a more probable alternative is found. The new algorithm optimizes the Magnusson swap operation, and offers a higher capacity of correcting tracking mistakes in multi-day, large-scale movies with densely populated cell cultures.

EllipTrack is based on the three-step procedure conventionally used in tracking, but introduces advanced features for each step. For track correction, the novel algorithm presented in EllipTrack consists of multiple steps, each addressing various aspects of tracking mistakes:

The core step optimizes the Magnusson swap operation to fix the mistakes in densely populated regions, using an iterative approach whereby tracking mistakes are progressively fixed.

For corrections, EllipTrack will consider the following alternatives:

1. swapping two cell tracks as done by the Magnusson swap operation.
2. assigning one cell track for mitosis while terminating the other one.
3. swapping two cell tracks and then breaking one of them into pieces.

In addition to the core step, the local track-correction module also fixes tracking mistakes related to under-segmentation, over-segmentation, and undetected mitosis events.

The authors tested the power of EllipTrack by comparing the tracking outcomes with the BaxterAlgorithm. They applied this algorithm to a 48h movie of mammary epithelial cells expressing a nuclear marker and a cell-cycle marker, and showed that EllipTrack could correct tracking mistakes of the BaxterAlgorithm. The authors then examined the ability of the EllipTrack pipeline to identify error-free cell lineages. For this purpose, they measured CDK2 activity as a marker of cell cycle. Using the algorithm, it is possible to evaluate the quality of identified cell lineages by examining whether they match the expected CDK2 activity. They determined that EllipTrack allowed correct identification of cells per lineage.

A systematic benchmarking of EllipTrack against various existing tracking algorithms was performed. These were applied to a variety of cells (HeLa cells, BJ5TA cells, RPE-hTERT cells, MCF10A cells and A375 melanoma cells), under various biological conditions, cell densities, treatments with various drugs, and time-frames, covering multiple generations of cells. The cell trackers were benchmarked on six different criteria:

1. SEG: segmentation quality
2. TRA: tracking accuracy
3. %CORR_S: percentage of cell nuclei correctly segmented before track linking
4. #COMP: number of complete tracks where cells were continuously tracked throughout the movie
5. #MIS_T: average number of mistakes among the complete tracks
6. %CORR_T: percentage of complete tracks correctly tracked

 The authors discuss the advantages and shortcomings of each tracking tool, and conclude that EllipTrack is a powerful tool capable of tracking nearly error-free cell lineages. EllipTrack also addresses a shortcoming of various other tools, which assume a constant migration speed of cells over an entire movie. This assumption can result in various tracking mistakes. To account for variable cell behavior, EllipTrack implements an option to perform time- and density-dependent inference of cell migration speeds from training datasets, which users can select for analysis. The authors demonstrate that the time- and density- dependent inference option, in addition to the local track-correction module, allows users to achieve a better tracking performance.

A further advantage of EllipTrack explored by the authors is that it requires minimal training, while still preserving high predictive power, a fact that represents significant time saving for users. Moreover, the authors identified that many tools used for tracking are not always user-friendly and this influences usage and learning. To improve practical usability, the authors developed two user friendly graphical user interfaces for parameter generation and training data.

 

What I like about this preprint

I like that the authors identified shortcomings associated with cell tracking, and developed their own tool to address those shortcomings. I think they systematically introduce the advantages and uses of EllipTrack in their manuscript. I like that they tested it in multiple biological conditions, and compared it with multiple existing tracking tools. I found the manuscript easy to read, easy to understand, and EllipTrack easy to use. I like also that the authors addressed a big issue affecting use of various image analysis tools: user friendliness. In their work they made sure that EllipTrack was easy to train, and was user friendly. They make their work fully available at github.com/tianchengzhe/EllipTrack.

Open questions

1. You tested EllipTracks’ ability to perform in different time- and density-dependent conditions. Is there a limit for either, upon which you find more errors even using EllipTrack?
2. You performed a thorough evaluation of EllipTrack in multiple cell lines. This allowed you to see the performance of EllipTrack in cells of different morphologies, migration characteristics, mitotic characteristics, etc. You discuss among the limitations of EllipTrack that cells with kidney shapes or non-elliptical nuclei are over-segmented. Are you planning an improvement on this on future versions, so that users can even choose the cell type being analysed? This would be comparable to the introduction you did of time- and density-dependence.
3. In all the cells you used, the nuclei of these cells are of considerable sizes. Is it possible to use EllipTrack to track much smaller cells? And what would the error rate be in such instances? for instance, since you discuss the potential of EllipTrack for drug discovery, I was thinking drug-testing studies in parasitology (including parasites such as Plasmodium or Toxoplasma), although this would involve tracking nuclei under 1µm diameter. Is this possible?
4. You mention in the discussion that EllipTrack does not consider unusual cell behaviours and you mention multipolar mitoses and cell-cell fusion as examples. Are there other specific instances where EllipTrack might have performance limitations?
5. A huge advantage of your work is that it is open access. Is it possible for users to contribute to the further development of EllipTrack?
6. How successful is EllipTrack for tracking fast-moving cells in 3D?

 

References

1. Magnusson, KEG, et , Global linking of cell tracks using the Viterbi algorithm, IEEE Trans Med Imaging, 34, 2015.
2. Tian C, et al., EllipTrack: a global-local cell-tracking pipeline for 2D fluorescence time-lapse microscopy, bioRxiv, 2020.

 

doi: https://doi.org/10.1242/prelights.21943

Read preprint (No Ratings Yet)