Cells and cell culture
MG-63 cells were obtained from ATCC and were grown according to recommended protocol (ATCC). Specifically, MG-63 cells were cultured in Minimum Essential Medium (Gibco/Invitrogen) with 2 mM L-glutamine and Earle's BSS adjusted to contain 1.5 g/L sodium bicarbonate, 0.1 mM non-essential amino acids, and 1.0 mM sodium pyruvate with 10% heat-inactivated fetal bovine serum (FBS). KM101 cells, a human bone marrow stromal cell line  were grown in 10% FBS in Iscove's Modified Dulbecco's Medium (Gibco/Invitrogen). Cells were cultured at 37°C in humidified atmosphere with 5% CO2. Prior to plating, cells were trypsinized with 0.25% trypsin (Gibco) and resuspended at a density of 2000 cell/ml. Aliquots of 60 ul were seeded into 384 well plates (Costar, black wall) yielding approximately 5 to 15 individual cells in the camera viewfield at the outset of imaging.
The plate layout included triplicate wells for each compound at each coating concentration for both cell lines. All operations were performed using multi-tip pipettors so that inter-well variations within-treatments were minimized to the greatest possible extent. Although the ideal plating pattern would be fully random, for practical purposes, advantage was taken of the interspacing of wells that occurs using 96 well multi-tip pipettors for applying solutions from 96 well "seed" plates into every-other well of the 384 well plate, producing "checker-board" patterns that intermixed treatments, doses, and uniform mock-coated control wells across the plate. Cell suspensions were pipetted from solution basins uniformly across each row.
Extracellular matrix surface coating
Mouse laminin I (Cultrex), mouse collagen type IV (Cultrex) and collagen type I (Sigma) were stored and reconstituted according to the manufacturer's instructions. Dilutions for coating were performed either in sterile water for collagen types I and IV or in Iscove's Modified Dulbecco's Medium (IMDM; Gibco/Invitrogen) without added serum for laminin I. Control wells were mock-coated using IMDM without added serum. Aliquots of 10 ul of solution were added to each well of a 384-well plate (Costar), and the plate was incubated for 1 hour at 37C. The collagen solutions were aspirated, and collagen-coated wells were allowed to air dry for about 20 minutes in a laminar flow hood. The laminin and mock-coated wells were then aspirated and all wells were rinsed with IMDM and aspirated prior to addition of cell suspensions.
Cell culture imaging cystem
The cell culture imaging system consists of a custom made environmentally controlled biochamber on an electronically controlled x-y motorized stage driven by stepper motor drive systems (Ludl Electronics, Ltd.) mounted upon an inverted microscope (Nikon TE 300) with electronically controlled motorized focus. The stage moves precisely to each well in the multi-well plate or to any number of locations within each well with a positioning repeatability of ± 1.5 um over the longest distance traveled by the stage. The biochamber temperature is controlled via heating cartridges with temperature feedback loops, and the humidity and CO2 content are controlled via commercial sensors to control feedback loops to a water reservoir heater and low pressure CO2 solenoid valve, respectively. The glass windows for illumination and microscope imaging are specifically heated through an electrically conductive indium tin oxide (ITO) coating using feedback temperature regulation so that condensation does not occur on these surfaces. A custom instrument control program, written in Visual Basic and C++, integrates control of the microscope stage, focus, optical filters, shutters, camera, fluidics, image storing functions, thermal zones, and subsystems through a specialized serial interface board with eight RS-232 connections.
Video time-lapse imaging and analysis
Images were acquired at 30 minute intervals with a 10× objective on a Nikon TE300 inverted microscope with a Photometrics SenSys high resolution (7 × 9 mm, 1036 × 1318 pixel chip) CCD camera (Roper Scientific). Image sequences were processed using a custom software program that identifies and records the location and morphological characteristics of cell-like objects (see Cell Segmentation and Outline Determination). The centroids of cell-like objects are then linked through sequential images to construct "tracks" that trace the route of individual cells according to proprietary algorithms [see Cell tracking algorithm]. The information accumulated during the processing is represented in probabilistic form so that the decision-making process does not have to be "black-and-white" (e.g. is the given object a cell or not, or is the given track the right track or not?), but is postponed until the end of the decision-making chain allowing for corrections for "lost" cells.
In each experiment for this study, 96 or more wells were imaged for at least 3 days, yielding more than a million records, each containing multiple measurements acquired for each object at each imaging time-point, i.e. every 30 minutes. These measurements fall into the categories of motility (scalar and vector forms of velocity, linearity, measures of deviation and frequency of oscillation of cell paths, track size and boundary), morphology and texture (area, perimeter, elongation, eccentricity, roughness, intensity and variation of intensity), summary statistics (object counts, cell counts, apoptotic frequency when fluorescence vital staining is applied) and complex parameters such as cell motility persistence , proximity analysis (cell-cell interaction, frequency and duration), division detection, growth rate, and viability (not all are applicable).
Cell segmentation and outline determination
The optical characteristics for inverted light microscopy of 384-well plates present challenges for robust segmentation of live cells. The meniscus obviates phase contrast; brightfield images are low in contrast and require significant processing. So a succession of filters is used to increase the difference (signal to noise ratio) between the background and foreground (cell-like objects), followed by an efficient region-growing operation that segments cell-like objects from the background. Heterogeneity of illumination across the image is reduced using local histogram equalization. Variations in illumination between images (across time) are handled using histogram matching. Background variations are smoothed using anisotropic filtering and adaptive median filtering, preserving cell detail and texture. Finally with brightfield microscopy, the cell boundary produced by the cytoplasmic membrane easily blends into the surrounding background, so a unique set of gradient variation and texture filters is applied to enhance the cell outline. Following filter-based enhancement, a region-growing operation identifies contiguous areas of cell-like or background-like pixels to segment cell-like objects from the background. A still more involved cell boundary determination can be achieved via active contour techniques (snakes) at the operator's discretion.
Cell tracking algorithm
The time-lapse interval in multi-well experiments is dependent upon practical considerations including the cell-type specific rate of motion, the total number of wells, the rates of stage movement, camera operation, and so on. For very fast cells such as T cells, imaging is performed on subsets of wells on a rotational basis in order to achieve intervals short enough for reliable track construction. At longer intervals, and particularly for view-fields containing many similar cells, cell track linking across the interval becomes increasingly unreliable because cells change shape and direction frequently, and as their paths converge, incorrect links may be assigned probabilities equal to or greater than correct ones.
Tracking is achieved by linking the matched cell-like objects between consecutive images in a probabilistic manner using a succession of increasingly stringent criteria. First, for each cell-like object, a set of candidate matches is chosen from within maximum speed and acceleration limits. Within this set, cell-like objects (blobs) are assigned to tracks based on match probabilities. Blobs are compared using multiple features such as location displacement, size difference, eccentricity changes, grayscale intensity IQR changes and normalized cross-correlation of the respective image portions. Converting distances into probabilities is done using a Gaussian probability density function. Because we assume that the features are independent, we can also assume that the resulting probabilities are independent. Therefore we can combine the different probabilities,
into the matching probability P(match
| i, j), where match
is a candidate match between track i and blob j and d
is the distance for feature n.
Cell objects in one image may compete simultaneously for multiple matches to different cells in the next image. A rule-based algorithm develops tracks based upon the values of the matching probability and tracking scenario. "Tracking scenario" includes the recent history and circumstances such as cell merging and splitting. For example, when two tracked cells that are similar collide, the merged object cannot be assigned to either track so both tracks are terminated. When two merged cells subsequently split apart, two new tracks are initiated because the matching of cells before and after the collision is ambiguous [see Additional file 3]. The cell tracking algorithm is very efficient in comparison to the cell segmentation; both are completed at the rate of 32 images/min, for 658 × 517 images on an Intel P4 2.8 GHz PC with 512MB RAM.
Both magnitude and directional velocity information are output from the linked positions of objects in sequential images. In this study, and for general investigation, we use a scalar average across several time points to smooth variation due to many factors. This average velocity represents the actual distance travelled, as determined by the movement of the centroid of the cell, divided by the elapsed time. Track lengths with fewer than 3 segments were not considered, and a maximum of 10 segments were included such that, for tracks longer than 10 segments, the velocity represents a running average. An exception to this method was used for "instantaneous velocity", as shown in Figure 8, where calculation was based upon displacement divided by time for the single track segment between the previous and current image. In order to exclude objects containing multiple cells in clusters or colonies, only velocity measurements for objects with areas less than three standard deviations above the mean area for each cell type were considered. The mean cell area and standard deviation was 1800 +/- 400 and 3100 +/- 900 square microns for KM101 and MG-63 cells, respectively. The lower limit for object area was 1200 square microns in both cases, based upon optimal visual exclusion of non-cell objects during set up of image processing variables. Velocity measurements were not included when the total area occupied by cells was greater than 30% of the viewfield area, i.e. when the cells were greater than 30% confluent. Finally, in order to filter out plate surface imperfections and adherent particles that gave rise to cell-like objects, data were excluded from tracks with fence sizes of less than 20 microns (See Results).
The technical accuracy of the imaging processing and data conversion steps were verified manually by constructing idealized image sequences with objects "seeded" at known pixel distances in order to generate known velocities based upon typical magnification and binning settings. Microscope optical magnification levels have been verified and calibrated with image pixel dimensions using a standard reticle.
Technical precision and quality control
Technical precision is here defined as the square root of the average variance for sets of triplicate velocity measurements (each measurement representing the mean velocity for all cells in the image) at single time points. Since absolute precision tended to increase in value for wells with cells of higher average velocity, and since there were sufficient data for analysis across the full range of velocities, we calculated summary precision values for triplicate sets ranked according to their measurement means (Figure 9). These data may be interpreted to indicate that the uncertainty in the mean velocity for a single well at a single time point ranged from approximately ± 0.08 um/min at the 0.2 um/min mean velocity level (CV = 40%) to approximately ± 0.12 at the 0.8 um/min level (CV = 15%) for KM101 cells, and it was slightly better for MG-63 cells. Sources of this measurement variability include the irreducible variation expected from random sampling from populations shown in Figure 6, taking into account the number, n, of cells in each sample, i.e. the number of cells imaged in each view-field as indicated in Figure 7.
Incidentally, the bin-based population variance, in which these precision estimates are rooted, should be employed when using the t-Test, rather than the variance of individual sets of replicates, to evaluate the significance of differences observed within experiments between triplicate means, i.e. for testing whether cell velocity was affected by an experimental treatment compared to a control treatment. The reason for this is that the sample mean and sample variance are independent when sampling from a normally distributed population. In other words, counter-intuitively, the mean of a set of three replicates that are widely spaced is likely to be as close to the "true value" as the mean of a set of replicates that are very closely spaced. The abundance of measurements made available by automation supports the validity of this claim, and would allow for slight adjustment of this principle when the data depart from normality, as it does here.
At the most fundamental level, because the stage mechanism re-centers each well into view after each time interval, images and data derived from them are subject to random errors associated with slight misalignment of the culture plate at each scan. The limits of this misalignment were determined by expanding images and manually tracking the motion of highlighted features of small imperfections on the culture surface throughout example image sequences. Such features were confined within a boundary of 3 square pixels over greater than 100 sequential images in all three experiments. Assuming Gaussian statistics, a limit of less than 1 pixel was estimated for the standard deviation of alignment of a single-pixel object, i.e. "stage noise". Under the magnification (10×) and binning conditions (2 × 2) used in these experiments, one pixel corresponds to 2 microns; this displacement for an object within a 30 minute time interval yields an upper limit for velocity of approximately 0.07 um/min. This calculation represents an upper limit to the contribution of stage noise, because in practice, stage-positioning error is expected to be a relatively small component of the total error that includes image-processing variability in determining cell outlines and object centroids.
Comparison of manual and automated tracking
Manual tracking was performed with a custom viewing program that enabled the user to store x and y coordinates by clicking on cells in sequential images with a computer-mouse. The 517 × 658 pixel images were displayed at approximately 7 × 8.5 inches on the monitor (approximately 75 pixels/inch) with a zooming option. A total of 27 cells were manually tracked from each of the four selected treatments. (For this exercise, a 27-cell limit was imposed by the nature of data output into a Microsoft Excel worksheet). The "rules" for manually placing the "centroid" and for terminating or initiating tracks were somewhat discretionary, e.g. the author continued to manually track cells through periods of contact with other cells, even though such scenarios were expected to involve track termination and re-initiation by the automated tracking algorithm [see Additional file 3].
For comparison of tracks on an object-by-object, point-by-point basis, we developed an algorithm to match objects from automated image analysis with manually tracked cells from each image. This algorithm identified and tabulated data from objects with automatically located centroids that fell within 10 pixels (20 um) of the x and y coordinates of the manually located cell positions that were determined by clicking with the mouse pointer. Following the initial comparison, both the manual tracking and the automated analysis were repeated on the four identical image sets, and the algorithm was applied to identify matched objects between the replicate manual and automated operations as well. Regression analysis and t-Test calculations were performed using R Project for Statistics [V. 2.0, see http://www.r-project.org].
Cell numbers and growth rates
Individual cells within colonies and cell clusters are not reliably recognized by current automated software using brightfield imaging. Instead, colonies and clusters are segmented as individual objects, and the areas of these objects provide a basis for estimating cell numbers. Doubling rates are calculated using linear regression of log transformed area-based cell numbers over time. Briefly, the slope of Log2(Cell Number) versus Time equals the doubling rate. As shown in this study, however, cell numbers do not correlate linearly will total cell area, and so doubling rates based upon an exponential growth model for area were called "area-doubling" rates.
Experimental design, normalization, and statistical analysis
A randomized complete block design  was used so that all informative factors (cell-line, compound, and dose) could be individually evaluated and separated from nuisance factors contributed by between-experiment variation and technical variability as evaluated between replicate wells within each experiment.
The behaviour of the cell velocities over time was analysed using 4-hour intervals and linear regression. First, velocity measurements were averaged for wells with the same cell line and treatment. Each data-point represents the average for 8 images (acquired every 30 minutes over 4-hour time intervals) from triplicate wells in each of 3 experiments. Second, a linear trend line was fitted to each profile. Finally the intercept (initial average velocity) and the slope (behaviour of average velocity over time) were compared across cell-line and treatment. The model fitting was done using SAS Proc Mixed. The significance of coefficients in these models is tested with student's t test. Overall tests for the equivalence of starting velocities are performed using the Chi-square test. Tests for equivalence of slopes are also performed. Pair-wise comparison with Bonferroni adjustment is also employed to see differences within each pair of treatments.
The dose-response of the cell-velocities was analysed within a 24-hour period. There was considerable variation between experiments as measured by the intra-class correlation (ICC), or ratio of variance from individual factors to the total variance . Therefore, prior to combining across experiments, all data for velocity and growth rates were normalized using an additive model based upon the difference between mock-coated plastic wells for each experiment. That is, the normalized measurements, Mn, were calculated from the original measurements, Mi, as follows: Mn = (Mi - Pi) + Pe, where Pe equalled the overall mean measurement for mock-coated wells for all experiments, and Pi equalled the individual means for mock-coated wells within experiments. Means were estimated using Tukey's biweight single-step M-estimator . The ICC for informative factors increased from 0.41 to 0.89, while that for the major nuisance factor, between-experiment variability, decreased from 0.58 to 0.09. The proportional contribution from technical variability (replication error) increased slightly from 0.005 to 0.011 due the decrease in overall variance after normalization.