A simple microcomputer-based system for real-time analysis of cell behaviour

Many tissue cells, in vitro, move over their substratum, be it a collagen gel or a plastic Petri dish. The phenomenon of cell locomotion has been studied intensively for many years, both as a topic of purely academic interest, and because of the role of cell motility and guidance in embryonic development, in wound healing and inflammation, and in cancer metastasis (Bellairs et al. 1982).

Time-lapse recordings of moving cells contain a plethora of information, most of which must be rejected in order to focus on the topic of particular interest. For example, an experimenter would wish to quantify the extent of cell spreading on a particular surface, or the degree of alignment of cells on some oriented substratum, like a stressed collagen gel. Alternatively, it is frequently desirable to measure the speed and locomotory parameters of cells moving toward a source of chemoattractant, in which case it is necessary to measure the cells’ positions over a period of time.

From this, it can be seen that there are two classes of data commonly required in studies of cell behaviour: ‘static measures’, for example of spread areas, number or length of processes, or cell orientation; and ‘dynamic measures’, for example of speed and direction of cell locomotion, rate of adhesion, or of response to chemoattractants. Static and dynamic measurements are very laborious to make, and both are highly amenable to computer techniques (Noble & Levine, 1986). In this paper, we will concentrate on the problem of dynamic cell tracking in real time.

Traditionally, studies of cell locomotion were performed manually, using a ‘post-analysis’ strategy: rather than tracing the movement of a single cell in real-time, the movements of several cells were first recorded on cine film or video tape. The paths of several cells could then be traced by hand, and displacements, speeds and angles calculated manually (Fig. 1A). That this procedure is extremely time-intensive is witnessed by the fact that Weiss, one of the pioneers of cell behaviour studies, left 10 km of cine film unanalysed at the time of his retirement (A. S. G. Curtis, personal communication).

Clearly, any reduction of labour is desirable. We have defined two possible levels of computer intervention in this manual procedure; computer-assisted and computer-centred. In a computer-assisted system (Fig. IB), the subjective judgement of cell position and manual tracking of the cell still rests with the experimenter; however, by using a digitizing pad, all subsequent measurement and calculation can be performed by a microcomputer (Lackie & Burns, 1983). This greatly simplifies the task, though a time lapse cine film of a 30 min experiment requires several hours of post-analysis, even after the film has returned from the processing laboratory.

A computer system that can automatically locate objectively and track cells would free the experimenter for more rewarding tasks. If such a computer-centred system (Fig. 1C) could perform quickly enough, it would be possible to avoid even having to record pictures and post-analysing them; the record of the experiment need only be a series of x,y coordinates for each of a series of cells. Systems that could perform such a task are available now, based on commercial framestores, but at formidable cost (£5000-£50000). This paper describes a simple microcomputer-centred system, based on an inexpensive (£100) framegrabber, which can track a large number of moving cells in real time, and thus offers many experimental opportunities.

For a computer-centred system to be useful, it must perform at least as well as the system it is intended to replace. A popular subject of cell motility studies is the polymorphonuclear leucocyte (PMN), or neutrophil (Wilkinson, 1982; Lackie & Wilkinson, 1984). In our department, a standard film sequence for analysis would be of 300 frames, recorded at 6-s intervals, and thus corresponding to 30 min real-time. From each run, the tracks of up to 20 cells would be recorded at 1-min intervals. A real-time computer-centred system, therefore, needs to record the positions of not one, but at least 20 cells, at less than 1-min intervals, to afford a significant improvement.

To be able to track cells whilst they are moving, it must be possible to grab the frame, analyse it, store the coordinate pairs, and grab the next frame, before the cells have moved too far to be relocated. The speed, or ‘cycle time’, in which these tasks can be accomplished is thus critical. Neutrophils move unusually fast (at up to 20 μm min⁻¹, compared with their diameter of only 8 μm). A sample rate of at least 2 min⁻¹ was found necessary in pilot experiments. For most other cells, which display a spread morphology, far longer cycle times are acceptable. BHK cells, for example, move at less than 0 ·5 μm min⁻¹, and so an image capture rate of one frame in 5 min is adequate.

A post-analysis strategy was considered, in which digitized images would be stored on disc, and then analysed later; however, the storage requirements for a single run (10 frame min⁻¹ × 15 min × 20 kbyte frame⁻¹ = 3 Mbyte per experiment) were impracticable. If x, y coordinates were stored, the disc storage required per experiment would fall considerably (15 readings × 50 coordinate pairs × 10 bytes per coordinate pair = 7 ·5 kbyte per experiment), so that a single floppy disc of 200 kbyte capacity, could hold data from perhaps 30 experiments. This was the approach adopted.

Cells were viewed through a Leitz Diavert or Leitz Ortholux microscope, equipped with phase-contrast, bright-field, dark-field or Nomarski optics. A Panasonic Newvicon-tube video camera was attached to a beam-splitting eyepiece, so that the image could be simultaneously viewed directly, and on a Panasonic 12-inch monochrome monitor. Cells were maintained at 37 °C by an air curtain device (a modified domestic fan-heater) (Wilkinson et al. 1982). The TV camera was connected to a Video digitizer unit (£100 : Watford Electronics), which was connected in turn to the ‘User Port’ of a BBC Master microcomputer, equipped with twin floppy disc drives, a plotter, a printer and a monochrome monitor. The combined system was capable of digitizing a TV image to a binary, 640 ×256 pixel (or image point), image within 2s. Using a 10 × lens, as in most of these experiments, a digitized field of 720 μm × 460 μm was obtained, corresponding to a pixel size of around 1 ·5 μm.

The central problem in computer tracking, be it of cells, limbs or missiles, is to identify, and measure the position of, the same object in successive frames. For some objects, or for noisy images, this problem can demand the very highest level of computer power. However, applications in cell biology need not be so demanding; the simple ‘box-search’ algorithm described below can be implemented successfully on a small microcomputer, providing certain experimental conditions are met.

The algorithm for the box search is simple (Fig. 2A); given that the x, y coordinates of the cell in the previous frame are known, and that the cell stands out clearly from its background (i.e. its pixels are distinctive), all the distinctive pixels within a specified radius of the last known position are counted, and the x and y coordinates of each such pixel summed. (Mathematically, this corresponds to taking the zero- and first-order moments of the object.) If no distinctive pixels are found, then the cell is known to be lost; otherwise the (visual) centroid of the cell is found by dividing the sums of the x and y coordinates by the number of pixels found in the box. Such an approach, using moments, is extremely widespread in image analysis: it is, though, rather surprising that so simple a search algorithm can perform adequately in cell tracking studies. The implementation of the box search in BASIC is described in Appendix 1; note that, strictly, a circular search area should be used, but it is far faster, and just as effective, to search a square ‘box’. The cell tracks will not be biased, provided that the diameter of the box is great enough to encompass even the fastest-moving cells.

There are certain obvious limitations on a box search. The first (Fig. 2B) is that fast-moving targets may move partially or completely out of the box between successive frames, and their tracks may thus end abruptly. To correct this, it is necessary either to increase the size of the box, or to decrease the sampling interval. In either case, the ideal box ‘radius’ (in μm) is clearly equal to the maximum expected distance moved by a cell between frames (i.e. cell speed (in μm min⁻¹) multiplied by the sampling interval (in min)). There are drawbacks either to very large boxes (the search time rises as the number of pixels within the box, or the square of the radius of the box) ; or to very short sampling intervals (few targets could be tracked in the spare time between frames).

The second limitation of the box search concerns the possibility that a second cell might intrude into another cell’s box (Fig. 2C). In this case, there are four possibilities; (1) the computer’s impression of the cell’s centroid will be temporarily distorted; (2) the box algorithm will switch allegiance to the other cell; (3) both cells move out of the search area in opposite directions, resulting in termination of the track; or (4) if both cells were initially tracked, then one cell might become tracked twice, while the other stops being tracked altogether. These problems can be overcome by keeping the cell density low, so that the chances that two cells move within one box radius of each other are slight; or by reducing the size of the box (but see the arguments above). In any case, it is possible to check, on each cycle, that no two sets of x, y coordinates are identical; if they are, one track is terminated. From the above, it is clear that a working compromise must be obtained between cell density, sampling interval and box size. In our hands, it was possible to track neutrophils at the density previously used routinely, of 0 ·5 ×10⁶ml⁻¹, corresponding to 50—100 cells per field of 720 μm × 460 μm. Taking the maximum speed of neutrophils as 20 μm min⁻¹, it was possible to track 40 cells every 30 s, using a box radius of 10 μm. Using a faster, machine-code box routine, 50 cells could be scanned every 20 s, using a box radius of 8 μm.x

It can be seen that there are two classes of error in tracking: premature loss of a track; and mistracking, or switching allegiance between cells. Neither of these errors will lead to a serious error in estimating population speeds or persistences (persistence, a concept derived from the random-walk theory, is roughly equivalent to the mean time between changes of direction); provided that tracking errors are not systematic (for example, unusually fast cells must not be lost preferentially), and that errors are rare. For example, if a tracking allegiance switched between two cells once during a run of 30 readings, only one data point (that at the transition) would be an inaccurate reflection of the cell population.

There are other, more sophisticated algorithms, for example ‘flood fill’ searches, in which only contiguous pixels contribute to the calculation of cell centroids, which perform better at high cell densities; or those in which higher moments for cell shapes are calculated (Dunn & Brown, 1986). However, these are rather more complicated and time-intensive to implement, and are more suited to ‘static’, rather than ‘dynamic’, measures.

Rabbit neutrophils were prepared by peritoneal lavage (Vickeret al. 1986), and stored at 5 °C for up to 3 days before use. A significant decrease in mean speed, and in the fraction of moving cells, was noted in this interval. They were observed in filming chambers of acid-cleaned glass in 50% Hanks’-Hepes salt solution, 50% peritoneal exudate. Cells were allowed 15 min to warm up before the start of the experiment, which ran for 25 min.

BHK fibroblasts (clone C13) were harvested from routine cultures in the department (Edwards & Campbell, 1971), and plated at 20000ml⁻¹ onto acid-cleaned glass coverslips, in HECT medium (Hepes-buffered Glagow-modified Eagle’s medium, supplemented with calf serum and tryptose phosphate broth). They were kept at 37°C at all times, and allowed to settle for at least 6h before filming.

PC12 phaeochromocytoma cells (gift from Drs A. McCruden & P.Schutz) were grown at 37°C on tissue culture plastic, in bicarbonate-buffered Eagle’s Medium, supplemented with 10% foetal calf serum and 10% horse serum. Neuronal differentiation was triggered by adding nerve growth factor (mouse 7 S ; Sigma Chemicals) at 100 ng ml⁻¹, one day before study. Multiple neurites are observed under such conditions (O’Lague et al. 1985).

The static and dynamic properties of the box search algorithm were tested by computer simulation. Repeated measurements were made of stationary discs generated by computer, and the mean and standard errors of the centroids thus determined were calculated. In another set of experiments, box and disc radii were set to values simulating a neutrophil-tracking experiment, and tracks obtained for different (known) speeds. The accuracy of speed measurement by the computer could thus be assessed.

The accuracy of the computer-centred cell tracking system was tested by comparison with the computer-assisted technique previously in use in the department. Time-lapse films, which had previously been analysed by computer-assisted methods, were projected onto the wall, and viewed with a TV camera. The microcomputer acquired and analysed each cine frame, advancing the projector automatically. As in the computer-assisted system, data from every tenth frame were stored for later analysis. After tracking was complete, data were processed by the same program that had been used for the computer-assisted system, and the population displacements, speeds and persistences obtained by the two systems compared. Speed, persistence, and augmented diffusion coefficient are parameters based on a pseudo-random walk model for cell locomotion, proposed by Dunn (1983), and applied to neutrophils by Wilkinson et al. (1984).

Repeatability of the computer-centred system was assessed by analysing the same sequence of cine film four times, and comparing the locomotion parameters obtained on each run. Similarly, the stability of the location algorithm was obtained by tracking the cells on a single frame as for a normal experiment, but without advancing the film between tracking cycles. In this way, a population of ‘stationary’ cells was obtained.

The accuracy of hand tracking was also assessed ; in these experiments, 26 individual tracks from a typical neutrophil-tracking run were plotted by computer. The experimenter then selected 20 ‘representative’ tracks and entered them, via a digitizing tablet, into a computer file. The population locomotion parameters obtained by this method were compared, using Student’s t-test, with the ‘true’ values obtained by the computer. This experiment thus assessed directly the reliability of a human experimenter in reproducing the intricacies of cell tracks.

Photomicrographs and digitized images of BHK cells are shown in Fig. 3. Bright-field images were of low contrast, and unsuitable for digitization. In phase contrast, cells are visible as dark grey objects, with white halos, against a mid-grey background (Fig. 3A). These images digitize rather well, and if a binary filtering level is chosen carefully it is possible to obtain a good picture of the cell, suitable for tracking (Fig. 3B). Although cells are clearly visible in the dark-field image (Fig. 3C), this illumination also shows up particulate debris in the medium, normally invisible in the light microscope. The resulting image (Fig. 3D) is thus far too noisy, despite the large ‘signal’ size. Finally, Nomarski optics, while providing plenty of information for the human eye (Fig. 3E), do not provide an easy image for computer analysis (Fig. 3F). The cells are shown only by a dark strip down one side, and a light one down the other. This provides a powerful pseudo-relief effect to the human eye, but is difficult for a computer to interpret, as the centre of the cell is the same brightness as the background.

Neutrophils provide extremely good phase objects (Fig. 4A), which digitize readily (Fig. 4B). Similarly, PC12 cells are clearly visible under phase conditions (Fig. 4C), and the club-like growth cones are particularly prominent in digitized images (Fig. 4D). For all three cell types, therefore, conventional phase contrast seems the best optical system for computer analysis.

Neutrophil tracks from a typical experiment are shown in Fig. 5. As can be seen, neutrophils move in zigzags, frequently doubling back on themselves. These tracks were plotted within 5 min of the end of the experiment, a considerable saving in time compared with the conventional methodology.

It proved possible to track, not only neutrophils, but also BHK fibroblasts and PC12 growth cones. The locomotion parameters of these cell types are shown in Table 1. It can be seen that neutrophils move an order of magnitude faster, and with an order of magnitude lower persistence, than either BHK or PC12 cells (which show a strong similarity in their locomotion parameters).

Cell locomotion parameters are plotted as a function of time since the start of the experiment in Fig. 6. Whereas the speed of cells is relatively constant, both the persistence and diffusion coefficient rise gradually to a peak, then decline. It can be seen that the normal warm-up period allowed (15 min) ensures that the cells are performing at maximum persistence; however, it is also clear that the cells locomote usefully for longer than the 30 min for which they are normally filmed. Why these changes occur is not clear; previous analyses have not revealed these trends so clearly and the observation raises interesting questions.

It is important to establish the reliability of the tracking system, in not losing cells too fast. From 60 –80 cells on the screen at the start of the experiment, the computer normally systematically scans the central screen area, selecting the first 50 cells encountered. Assuming that the distribution of the cells on-screen is initially random, this corresponds to a random cell selection. (It is also possible to select individual cells deliberately for tracking, although considerations of objectivity preclude this for routine use.)

Some of the cells initially selected will be lost as they move off-screen; some by collisions with other cells (the computer automatically terminates tracks with the same coordinates); and a few for ill-defined reasons. After the experiment finishes, the experimenter is presented with each track in turn, providing the opportunity to select only moving cells with at least five coordinate pairs, for further analysis. The maximum number of cells accepted by the computer-assisted analysis program was 20, so a fairly steep rate of attrition, from the 50 cells originally selected, was acceptable. Fig. 7 illustrates the drop-out rates as a typical experiment proceeded. Clearly, then, this system is capable of producing an adequate number of acceptable tracks.

The determination of the centroids of stationary, computer-generated discs revealed no inaccuracy or random fluctutation in the computer search algorithm. When moving discs were tracked, the speeds determined by the computer were accurate until the distance moved by the disc between frames approached the box radius. Under such circumstances, speeds were underestimated, and cells were lost prematurely (Fig. 8). This is unlikely to prove a serious problem in practice; the box radius used (24 pixels) would correspond to a radius of 50 μm when using a 10 × objective. In a neutrophil-tracking experiment, at 3 frames min⁻¹, this corresponds to a maximum trackable cell speed of 150 μm min⁻¹, at least five times greater than the maximum speed observed.

The quality of image produced by digitizing cine films projected onto the wall was surprisingly good, and the computer had little difficulty in tracking the cells. The population locomotion parameters for a typical experiment, obtained by computer-assisted and computer-centred systems, are compared in Table 2. It is clear that the repeatability of the computer-centred system is very high; that is, when presented with the same cells, it produces effectively the same results. However, these data differ slightly from those previously produced by computer-assisted hand-tracking. In particular, hand-tracked cells have higher persistences and diffusion coefficients than computer-tracked cells, although the differences in speeds measured by the two systems are small.

The differences between computer-assisted and computer-centred determinations of cell locomotion might be due to shortcomings in either system. When computer generated tracks are transcribed by hand, cell speeds, persistences and diffusion coefficients rise artefactually (Table 3). This is clearly due to a smoothing effect on the track data. Fig. 9 compares the tracks derived for the same cells in computer analyses, with those plotted by hand. Again, the repeatability of the computer-centred system is excellent, and the tracks it generates are noticeably more jagged than those plotted by hand.

The system described here is extremely simple, and yet highly effective at tracking neutrophils in real time. It also has potential in tracking spread cells. As cells in culture round up to divide, it would also be possible to estimate the number of divisions occurring, or even to adapt the tracking program to detect a division in a cell being tracked, and to continue to follow each daughter cell individually. On a finer scale, it may be possible to track neuronal growth cones, allowing the system to be applied to many problems in developmental neurobiology.

Results obtained by the computer-centred system differ slightly (but significantly) from those previously obtained. Speeds, persistences and population diffusion constants are all lower when determined by the computer-centred system. It seems that this reflects the superior ability of the computer to follow higher-frequency fluctuations in cell position; human participation in the analysis process leads to a significant ‘smoothing’ of the data. In this sense, the computer-centred system offers improvements, not just in speed of operation, but also in objectivity. A possible problem, however, is that cells that ‘jiggle’ in one position without moving can produce anomalously high speeds and low persistences, if samples are taken at the high rates possible in a computer centred system. Non-locomotory noise can also be produced by rapid shape changes in otherwise slow-moving cells. Care must be taken therefore in applying such powerful computer techniques to cell locomotion; and computer data recording should be seen as an adjunct to, not a substitute for, manual experimentation on any given system.

It is important to note that the ‘secret’ of the system is that all the potentially difficult, or slow, computer tasks are obviated by careful experimental protocol. In particular, there is no need to introduce sophisticated collision-detection software if cells are plated at reasonably low densities. Nor is there any need for image enhancement or filtering if care is taken to produce a clean, evenly illuminated light-microscopic image.

The microcomputer used in these experiments (the BBC) is almost ubiquitous in British universities. However, the algorithms discussed are applicable to the wide range of microcomputers for which imagegrabbing hardware can be purchased. In particular, users of IBM PC and Apple computers should have little difficulty in emulating the system described here.

As to cost, if it is assumed that all interested parties already have access to a microscope, video camera, monitor and a suitable, inexpensive microcomputer, then the only additional costs are for the digitizer (£100 at the time of writing), and for software development. It is hoped that this paper will assist in the latter.

Thanks are due to Miss S. Kitson for technical assistance, and to Professor A. S. G. Curtis and Dr F. Lyall for their constructive comments. This work was supported in part by BP Venture Research Group, and by the General Funds of the University of Glasgow.

Rather than provide a machine-specific BASIC listing, the major tasks required of a tracking program are described in flowchart form, from which it should be easy to derive a specific application. Comments are in parenthesis.

Array x(max. number of cells)

Array y (max. number of cells)

c=0 (this counts the number of cells being tracked)

Open an output disc file

Loop: get an image: until keypress (this allows the user to adjust the picture quality)

For n = 100 to 1000

Form = 100 to 1000 (Scan the central area of the screen with two DO-loops)

If POINT(n,m)>0 then PROCbox (If pixel is white, call box routine to find the centroid)

Store x and y coordinates in x(c) and y(c)

Increment c

Blank off the pixels (so they aren’t ‘found’ twice) If c = max. number of cells then jump to main loop

Loop:

Get an image

For n = 1 to c (i.e. for each track)

If x(n)>0 call box (x(n)=0 implies the cell has been lost) If cell found

write new coordinates to disc

store new coordinates in x(n) and y(n)

remove any duplicated tracks

Or if no cell found

set x(n) and y(n) to 0

write 0,0 to disc

Finish loop if max. number of cycles reached, or user presses key

Close disc file