Conceived and designed the experiments: JT PR RE PS HS. Performed the experiments: SC RE. Analyzed the data: JT MRB HS. Wrote the paper: JT PR MRB RE PS SC HS.
The authors have declared that no competing interests exist.
An automated technique for the identification, tracking and analysis of biological cells is presented. It is based on the use of nanoparticles, enclosed within intra-cellular vesicles, to produce clusters of discrete, point-like fluorescent, light sources within the cells. Computational analysis of these light ensembles in successive time frames of a movie sequence, using k-means clustering and particle tracking algorithms, provides robust and automated discrimination of live cells and their motion and a quantitative measure of their proliferation. This approach is a cytometric version of the
Computerized identification, discrimination and tracking of biological cells, in microscopy images, is vital to modern high throughput platforms that deliver automated scanning and capture of millions of images per day
These computerized approaches mimic human visual perception of form and motion where dense and complex image information is processed to obtain much simpler, abstract representations of objects and their position. However, through early studies by Wertheimer and others on the relationship between perception and simplified abstractions, such as points or lines, it is now known that human perception can operate directly at the level of the abstract object and so does not require detailed information – the human form of a ‘stick-person’ is recognizable despite consisting only of straight lines and a circle. This is the
In this paper we report on the implementation of the moving-light-display technique in cells by the use of lysosome-encapsulated quantum dots (QDs) to create clustered points of light, through which cells can be identified, tracked and analyzed. The creation of the optical sources through endocytosis of fluorescent markers provides a generic and innate encoding mechanism, applicable across multiple cell types. The nanoparticles provide robust, bio-stable and photo-stable fluorescence that can be tracked over multiple generations, and at nanomolar concentrations do not perturb cellular function
(A) Bright field image with binary element overlay; (B) fluorescence image of two neighboring A549 cells; (C) the light point cluster map derived from the fluorescent signal. The binary elements within the light point map are displayed as blue points and represent the locations of the centres of the quantum dot labeled vesicles; the centroids of the binary element clusters are displayed as red points. (see Movie S1 for mitosis animation).
(details presented in section on
In the following sections a demonstration of the moving-light-display concept within cells is presented; starting with details of the microscopy and image analysis techniques used and then expanded through reference to the biologist’s requirement for i. identification and discrimination of cells, ii. spatio-temporal tracking of their motion and division, iii. analysis of division events and iv. visualization of time dependent relationships in cell lineage maps. We conclude with a discussion on the applicability of the technique and a general summary of the nanoparticle MLD approach.
A549 (ATCC CCL-185) cells were maintained under G418 selection in McCoy’s 5a medium supplemented with 10% fetal calf serum (FCS), 1 mM glutamine, and antibiotics and incubated at 37°C in an atmosphere of 5% CO2 in air. For imaging experiments, cells were grown at a density of 1×106 cells ml−1 as a monolayer in either coverglass bottomed chambers (Nunc, 2 Well Lab-Tek II, Fisher Scientific) or glass bottomed (24 multi-well Sensoplate, Greiner Bio-one for 24 h prior to imaging. All cell concentrations were determined using a Coulter Particle Counter (Beckman Coulter, High Wycombe, UK).
Cells were loaded with commercially available targeted nanocrystals using the Qtracker® 705 (QTracker705) Cell Labeling Kit (Invitrogen (Q25061MP) at 4 nM concentration. The reagents in the Qtracker® 705 Cell Labeling Kit use a custom targeting peptide (9-arginine peptide) to deliver near-infrared-fluorescent nanocrystals into the cytoplasm of live cells via the endosomal pathway. Briefly, Qtracker reagent A and B were premixed and then incubated for 5 minutes at room temperature. 1 ml of fresh full growth media was added to the tube and vortexed for 30 seconds. This labeling solution was then added to each well of the cells and incubated for 1 hour at 37°C after which they were washed twice with fresh media. Subsequently 24 hours later, labeled cells were then analyzed by time-lapse, confocal microscopy.
(A) Bright field image of A549 cells overlaid with binary elements (blue) and assigned centroids (red); (B) full-field binary element representation of a 336×256 µm fluorescence image taken at the initial time point of experiment and (C) representation of the same field at the t = 40 hour time point. (see Movie S2 for evolution animation).
Confocal laser scanning microscopy (Radiance CLSM, BioRad Ltd) was used to track quantum dot labeled A549 cells over a 48 hour period. The Qtracker705 fluorescence was collected using 488 nm excitation and 680–20 nm emission filters; x,y,z,t optical sections (using ×40, 0.75 NA air lens) were collected every 5 minutes.
All image processing was done within the MATLAB programming environment. The fluorescence image data was provided in the format of multi-layer tiff stacks each containing 8 focal plane images for every 5 minute interval across the 48 hour experimental range. The capture of multiple focal planes allowed creation of a composite image using the highest contrast regions of all available images. This composite was created by segmenting the image space into smaller regions and applying the standard absolute gradient algorithm of the form:
Conversion to a binary representation is a simple two step process. A step function of the form A simple peak finder algorithm locating the maximum pixel intensity within a localized area by cross-referencing the linear profiles of the pixel rows and columns was applied to locate the maximum intensity points of the fluorescent signals corresponding to nanoparticle loaded vesicles and their x and y coordinates stored.
(main figure) An example of the characteristic curve of mean binary element separation as an individual cell undergoes a mitotic event and divides into two daughters cells. (sub-panels) The corresponding stills below show the binary element distribution, in a fixed frame position, at four time points spanning key stages of the process: (A) Cell prior to extracellular signs of mitotic committal; (B) localized contraction of the light-point markers as the cell prepares to divide; (C) marker distribution indicating telophase stage; (D) two daughter cells identified as independent clusters (frame scale is 40×30 µm). (see Movie S3 for animation).
A standard k-means clustering analysis algorithm was used to identify groups of pixels in the binary image (binary elements) corresponding to fluorescent vesicles within the same cell. For the binary elements coordinates
In successive images the centroid locations from the previous timeframe were used as a seeding set. To deal with the sporadic occurrence of rogue binary elements (noise) and binary elements representing cells entering the field of view, new centroids are assigned to regions containing binary elements that are more than 120 pixels (30 µm or ∼3 cell diameters) from their nearest centroid. New seeds are assigned to deal with both these occurrences as they are indistinguishable without time consuming comparisons with previous frames and interpretation of boundary events. It is temporally and computationally more efficient to assign binary elements to new clusters generated according to basic rules and interpret the nature of the groupings during later processes. A proximity validation was applied by identifying current seeds with no binary elements within a distance of 50 pixels (12.5 µm) and removing them. This primarily deals with the event of a cluster moving out of the field of view therefore leaving a centroid seed from the previous frame with no binary elements to define it and secondarily with centroids previously defined to account for rogue elements whose intensities have dropped back below the noise filter. The k-means algorithm was then run using the modified centroid set and a measure of the cluster fit taken by implementing a native Matlab silhouetting algorithm to acquire a cluster fit parameter. The Silhouette process provides a validation of the clustering by determining how well each binary element fits within its assigned cluster. For each binary element
Refinement of the spatial tracking was accomplished by temporal tracking of the centroids and by identification of cell division events through further analysis of the separation of binary elements within clusters.
The centroids in each time frame were linked to centroids in surrounding time frames through application of a nearest neighbor map. For two sequential frames with centroid sets
(see
Graphical representation of a single cell’s lineage evolution through space alongside conventional lineage portrait as defined by the automated centroid tracking. Diamonds mark the spatial and temporal location of mitosis events, red being the progenitor cell, green the second generation cells and magenta the location of the cell centroids in the final frame of the time-lapse sequence. The cross marks the loss of a cell behind an unidentified object in the image (non-cell). The arrows in the main diagram may be viewed as motility vectors reporting the mean velocity of a cell between mitotic events.
A typical example of the embedding of point light sources within cells, through the use of QDs, is shown in
Once an initial field of cells has been identified and cluster centroids assigned, a full spatio-temporal track can be obtained through linkage of the centroids through successive time frames. Centroid seeds for each successive frame are defined to optimize the speed of the k-means process, account for the possibility of losing and gaining clusters off the edge of the plane of view and deal with rogue binary element points occurring sporadically through time in otherwise empty regions. The general approach taken can be summarized as follows: 1. generate a foundation set using centroid locations from the previous time frame, 2. validate proximity of all centroids to binary elements to identify centroids whose binary element cluster has moved out of view and 3. validate proximity of binary elements to centroids to account for any noise elements that passed through the filter. Images were taken with a 5 minute time interval; this minimizes the probability of large changes in cell position and thus ensures accurate correlation of centroid locations from frame to frame.
A movie sequence of this centroid motion and the ‘birth’ of new centroids in daughter cells, taken over a 48 hr period, is shown in Movie S2; single images, taken at the 0 and 40 hr time points, are shown in
(A) Calibration curve of initial image showing the relation between the intensity threshold filter cut-off and the number of binary elements identified. (see Movie S4 for animation) (B) Plot showing the operational range of the system through the binary element dependence of the number of centroids (cells) identified (black curve), and the centroid displacement of a selection of cells (colored curves). The vertical, red dashed line indicates the minimum point number requirement for maximization of the number of cells automatically identified. (C) Bright field images at t = 0 and t = 24 hours with centroids overlaid. Red spots indicate automatically identified centroids, blue spots correspond to incorrectly assigned centroids and yellow spots represent manually identified cells which are unidentifiable by the automated analysis. (see
As cells go through mitosis they undergo distinct morphological changes that can be tracked via the binary elements. The mean distance of the binary elements from each cell centroid provides a characteristic parameter, from which a ‘mitotic curve’ may be constructed (
The mitotic curve provides not only a clear digital marker of cell division events but also a quantitative, analogue track through the mitotic phases which informs on the kinetics of the cell division process (see supporting media for further examples). Here the moving point display technique provides a tool with which the shape and motion of cells can be analyzed. The linkage of cell division events through spatio-temporal tracking provides a functional lineage analysis capable of describing the clonal relationships of a wide range of morphological and motility measures. As an example, a motility lineage map is shown in
In order to numerically quantify the operational range of the nanoparticle-encoded MLD we take a representative sample image and investigate the effects on cluster identification and assignment of centroid position due to variation in the noise filter threshold (
If the noise threshold filter is reduced too far the binary element number begins to relate to background noise rather than valid QD encoding pixels. Whilst the presence of these noisy pixels does not immediately invalidate the identification of cells (a random noise source adds pixels to each cell cluster with equal probability) it does affect the centroid co-ordinates and so leads to inaccurate determination of cell position. As the threshold filter approaches the noise floor (intensity ∼ 400 units) and the total number of binary elements increase beyond 1500 the centroid positions show marked deviations in excess of average cell diameters (>10 µm). However SNR values as low as 1.1 can be tolerated before this noise weighting produces a cluster position offset of 10 µm.
To assess the accuracy of centroid seeding (cell identification) and spatio-temporal tracking two image frames corresponding to the 0 and 24 hour time points were chosen. An initial centroid set was chosen at the intial timepoint by visual inspection of the bright field image; the number of centroids in the timeframe captured 24 hours later was then automatically assessed using the MLD algorithms and compared to direct visual identification. The 24 hour imageframe is shown in
Quantum dots are widely used for cellular labeling as they provide both a photo and bio stable fluorescence marker that can be spectrally tuned. Here we show that the processing of these nanoparticles by the cell is as important as their innate photonic properties; they are naturally taken up and concentrated into vesicles that are dispersed throughout the cytoplasm and as such provide bright, point like light sources across the majority of the cell cytoplasm. This bio-processing thus enables image analysis by these discrete points which are the fundamental elements of a moving light display. Concentration on representative sub-sampling of the image rather than full visualisation has two major benefits: i. major data reduction and ii. access to rapid and efficient image examination based on cluster analysis techniques. The processing of a bright field image to produce a binary map of QD light-points (
To summarise, we have used quantum dot nano-particles as point like markers within a cell to demonstrate the concept of moving light display to resolve cellular mitosis events and lineages without the need for complex interpretation of a complete visual picture of the cellular field. In the study of cell movement and proliferation the knowledge we seek can be resolved at the whole cell level and so the reduction of the data set to a minimal set of Cartesian co-ordinates provides a much greater efficiency of information processing, distilling the experimental measurements down to match just that needed to understand the biology. In future applications this approach will allow real-time data processing during image acquisition and the direct storage of biological knowledge, i.e. cell position and familial relationships.
(TIF)
(WMV)
(WMV)
(WMV)
(WMV)