BactImAS: a platform for processing and analysis of bacterial time-lapse microscopy movies

Preparation of images

Individual images from up to three recorded channels must be stored as image sequences. Alongside the set of original images, the user has to provide the 8-bit PNG versions (selected because of their small file size) that the software uses internally for displaying and tracking tasks. Both these sets are easily prepared using, for instance, the attached ImageJ program. This does not limit the format of the original images, as measurements can be taken on any image format supported by ImageJ.

Due to the imprecision of the motorized stage, there is often a translational shift between two successive frames in an image sequence. This must be corrected before cell segmentation. To this end, BactImAS has a simple and robust built-in registration algorithm (see Additional file 1) to identify and correct this. If this algorithm fails, images can be registered using external software.

Apart from these mandatory steps, the user may opt to define parameters such as pixel-to-micron ratio, specify frames that should be ignored (e.g. out-of-focus frames), and associate experimental information with the corresponding frame (e.g. change of growth medium).

Cell segmentation and tracking

Once the images are loaded, the user must delineate contours of the cell(s) of interest in the first frame. For easier tracking of cell poles, generations, and lineages in general, initial cells are named in an alphabetical manner and their progenies in binomial fashion (e.g. cell A divides to AA and AB, cell AB then divides into ABA and ABB etc). The user then selects an algorithm that segments and tracks each cell for the assigned number of frames on a selected channel.The user can easily inspect the segmentation results throughout the movie processing step. Upon selecting a frame of interest, the program produces a new image stack with ROI selections superimposed on the images from each channel (e.g. blue, green, and red) and of merged combinations of channels (B+G, B+R, G+R, B+G+R) (Figure 1B). In addition, BactImAS can generate an image sequence of all processed frames composed of ROI selections superimposed on phase-contrast images. If errors are observed, the segmentation can be repeated with different algorithm parameters or by using images from a different channel (if available). Alternatively, any erroneously segmented cells can be manually corrected. This not only corrects the ROI selection of the selected frame, but also facilitates the correct segmentation for the subsequent frames.

In addition to manual selection of cells, it is possible to set a specific cell to be ignored (e.g. if a tracked cell partially exits the image borders), to be defined as “dead”, or to be assigned any other property from a certain frame onwards.In many cell types, a division event is easily defined by a membrane constriction between two adjacent cells, visible on phase-contrast images. In contrast, in mycobacteria a division event often precedes any identifiable membrane constrictions. This mycobacterial feature makes it challenging for an algorithm to define the exact frame and intracellular position of the division event. BactImAS therefore relies on the user to define the division events throughout movie processing. The user is assisted in this decision-making process in two ways: image sequences of each channel are separately displayed in the GUI and the aforementioned image stack, displaying a combination of channels for a selected frame (Figure 1B). After recognizing a division event, the user has to define which specific cell divided (this automatically generates a name for each daughter cell and allows the program to keep track of cell lineages) and then to delineate the contours of the two daughter cells.

Mycobacterial segmentation algorithm

Previously developed algorithms cannot deal with typical M. smegmatis time-lapse movie due to challenges illustrated in Figure 2. To segment mycobacteria-like cells we developed an algorithm (for detailed description see Additional file 1) based on the following assumptions: cells change only slightly from frame to frame thus making tracking straightforward; cells are sometimes in close juxtaposition to neighboring cells but most juxtaposed cells have at least some segments of visible edge; cells are worm-shaped with a relatively fixed width but varying length; and almost all growth happens along the length of the cell.

The algorithm requires three parameters to be defined, corresponding to cell width, maximum pole elongation, and the maximal area increase (see Additional file 1for the precise definition of these parameters and details on how to use them). For any given frame, the algorithm consist of the following steps:

•  Original image (Figure 3A) is processed by applying a 3 x 3 Sobel edge detection filter [24] (Figure 3B).

Selections are calculated in a round robin fashion, which is a drawback of this algorithm, as it makes it order dependent. However, this feature comes into play only in situations where there are no detectable borders and it is not clear to which cell to initially assign the joint area.

Extraction and visualization of quantitative data

The user can extract quantitative data for the annotated cells in a configurable and fully-automated way. Measurements can be taken from images in any file format, as long as they are of the same resolution and alignment as those used in segmentation and tracking (such that ROIs would delineate the same image areas). Additionally, it is possible to define whether to morphologically erode the shape before taking measurement (e.g. for the membrane region to be excluded from the cytoplasmic fluorescence measurements). Furthermore, BactImAS can acquire measurements from a user-defined background area on all channels for each frame. This information is utilized to provide an additional set of background-corrected values for each cell.

Many additional pre-defined variables are provided in the SQLite database to simplify subsequent data analysis, such as interdivision time and growth rate (see Additional file 1: Table S1–S3). The user can export all or a subset of data as a CSV file and then analyze them externally. We provided a built-in SQL editor to query the database and perform basic time-lapse analysis but the user can also use other specialized software such as SQuirreL SQL [27].

BactImAS enables visualization of the cell lineage according to user-defined graphic parameters. In such a lineage tree, the branch length corresponds to the time between divisions and branching points represent division events. If there is more than one cell at the beginning of the experiment, a separate lineage tree is plotted for each cell. Optionally, the designated name of each progeny cell is plotted. The user can select any two variables and visualize them on the lineage tree, one as the branch width and the other as branch color.

BactImAS as a platform for bacterial time-lapse movie analysis

An important novelty, with respect to all other bacterial time-lapse analysis software, is data storage in the form of a relational database. The advantages of this approach are manifold: easy data backup and exchange, elegant retrieval of results via the standard SQL language, and simple data maintenance/handling in general. We are using the SQLite database as a data repository for all the information extracted from the movies (see database diagram in the Additional file 1). The SQLite is portable, has low memory requirements, and is freely available. However, it is easy to replace SQLite with any other database management system (e.g. MySQL, PostgreSQL) since the Java database connectivity standard is used to connect the program with the database. This replacement would be advisable should the program be configured to use one centralized database. By default, the platform is single-user oriented, with the ability to exchange data with other users by simply transferring a single SQLite file (and associated images).

The BactImAS platform is built in a modular fashion (Figure 1A) to facilitate addition of new functionalities, primarily additional cell segmentation and tracking algorithms. Similar to Schnitzcells [6] and MicrobeTracker [7], it is possible to build an algorithm collection so that users can perform the segmentation and tracking with the best fitting algorithm. To add an algorithm, one has to write a Java class implementing the given programming interface (for detailed description see Additional file 1). BactImAS adopts ImageJ data structures and functionality for various image processing tasks (e.g. edge detection) and adds a number of its own functionalities, which facilitates the development of new algorithms.

Segmentation and tracking of microfluidics-grown mycobacteria

We validated BactImAS on time-lapse movies of microfluidics-grown M. smegmatis cells (Additional file 2). Details on the experimental setup are given in the Additional file 1. The BactImAS algorithm tracked the growth of every individual cell in a microcolony using phase-contrast images with parameters for cell width, maximum pole elongation, and the maximal area increase set to 3, 10, 35, respectively (Figure 4 and Additional file 3). In the presented time-lapse movie, 6.3% (201 out of 3184) of the ROI selections had to be manually corrected. Manual interventions were also required to define division events. However, division events represent only a small percentage (3.7% in the presented movie) of total cell selections in a typical mycobacterial movie. Similar results were obtained with other processed movies, resulting in, at least, a 10-fold decrease of processing time when compared to manual analysis. As these results were obtained using suboptimal images (pixel size was 0.129 μm), we expect a further reduction in the percentage of manual selections when using higher resolution images.

Currently available software solutions were developed for bacterial species where the typical frame-to-frame difference between cells is high, but the programs could rely on the clear cell contours to obtain good segmentation. In our experimental system, the situation is the opposite: cell delineation is difficult, but the tracking is simple because cells grow slowly and minimally change position between two successive frames (Additional file 2). To examine if the developed algorithm is more widely applicable, we processed published time-lapse movies from three different bacterial genera: fast-growing rod-like Bacillus [28], crescent-shaped Caulobacter [7], and filamentous Streptomyces [29]. The most promising results were obtained by processing the Streptomyces time-lapse movie. The main limitation of the algorithm lies in its sensitivity to significant frame-to-frame changes in cell shape and position. Should the skeleton from the previous frame fail to overlap with the contours of the cell body in the following frame, an erroneous selection would likely be generated. This presents a problem when detecting fast-growing organisms (e.g. E.coli or B. subtilis) and/or when the frequency of imaging is such that significant changes occur between subsequent frames. Hence, selection of an appropriate image acquisition frequency, one that enables good automation of the tracking step, should be considered when designing the time-lapse experiments.

We intend to build a multi-channel version of the algorithm to simultaneously use the information from all available channels. This would potentially allow automated detection of cell division events in mycobacterial strains that express a fluorescently-tagged Wag31protein, a reporter that accumulates at the position of cell division septum and poles [16], clearly marking division events.

Visualisation and analysis of bacterial gene expression

We illustrate the usage of BactImAS by analyzing real-time expression of icl, a gene previously shown to be indispensable for chronic infection of M. tuberculosis in mice [30, 31]. A change of carbon source in the growth medium from glucose to acetate affects the quantity and activity of ICL in M. tuberculosis[32]. To investigate the role of icl induction in this process, we followed the ICL reporter strain of M. smegmatis (described in detail in the Additional file 1) in which a green fluorescence protein (GFP) signal serves as a proxy for ICL production within the cell (Figure 2, middle row and Additional file 4). Beside reflecting changes in icl expression, the GFP fluorescence is also influenced by fluctuations in the global gene expression machinery (e.g. altered number of ribosomes) within each cell. To account for differences in icl expression that stem from these other cellular factors, we monitored, in parallel, the signal of DsRed2, a red fluorescent protein constitutively-expressed from an independent chromosomal locus (Figure 2, bottom row and Additional file 5). Dual-fluorescent ICL reporter cells were cultured in a custom-made microfluidic device [4] and imaged every 10 minutes. For the initial 29 hours cells grew in a glucose-based flow medium and then the glucose was replaced with acetate. We processed three time-lapse movies and extracted quantitative data for every cell of the microcolony. This data is the first such demonstration of single-cell gene expression dynamics in mycobacteria.

Although time-lapse data is commonly represented as a lineage tree, researchers predominantly use unavailable in-house software. BactImAS includes a novel and advanced visualization tool that can display lineage tree(s) superimposed with any two variables from the database. A representative tree, with cell size displayed as proportional to the branch width and mean green fluorescence mapped to a color scale, demonstrates the advantages of such visualization for representing complex information (Figure 5A). Upon switching to an acetate-based growth medium, there is a clear increase in icl expression, with concomitant decrease in typical cell size. Phenotypic heterogeneity in icl induction and cell growth, observed even between sibling cells, is clearly evident from such visualization.

The built-in SQL editor allows data analysis to be performed directly with BactImAS. With a single query to the database (see Additional file 1), we analyzed the dynamics of icl expression in all processed movies. We calculated the green/red fluorescence ratio for every cell (7306 ROIs) and then plotted the average ratio as a function of time (Figure 5B). The icl expression was found to be stable in the 9 hour interval prior to the glucose-to-acetate switch. Upon the switch, icl expression continuously increased over the next 5 hours before reaching a 6-fold higher plateau. This is in very good agreement with the 5.9-fold difference in the icl expression level determined after 24 hours by quantitative real-time PCR in M. tuberculosis[32].