The extracellular matrix (ECM) provides a biophysical and biochemical environment for cell mechanical behavior. In turn, cellular interactions with the ECM resulting from adhesive, proteolytic and migratory activity govern continuous matrix reorganization. One such example occurs in eutrophic inward remodeling of small arteries [34]. Here, the existing collagen matrix is rearranged around a smaller diameter. Such inward remodeling is a hallmark of many hypertensive disorders and has been shown to have predictive value for cardiovascular events [9].

In vitro setups of cell-seeded matrix scaffolds allow studying the specific components that are active in vascular remodeling [22]. Especially, matrix reorganization can be monitored in the presence of specific cells, matrix elements, and blockers or markers of remodeling enzymes such as the matrix metalloproteases and transglutaminases [1, 2]. Several approaches using multi-cellular preparations have been developed. Thus, free-floating collagen gels can be seeded with contractile cells, and the remodeling is monitored from the dimensions of these gels [8, 13, 16, 20]. In the so-called culture force monitor, collagen gels mixed with cells are allowed to polymerize between a fixed plate and a force transducer. Contractile force development is measured during subsequent culture of this preparation [12, 14, 22, 36]. However, interpretation of data derived from these multi-cellular approaches is troubled by variability in cellular mechanical activity, lack of information on cell density, and synergistic contractile effects [30]. Therefore, microscopic observations on matrix reorganization by single cells may provide more detailed and fundamental information on the mechanisms of remodeling.

Several groups have studied the effect of single cell activity on local deformation of matrices [5, 28]. The most frequently used substrates are silicone rubber [3, 6, 18, 21] and polyacrylamide sheets. Local deformation is then monitored by tracing the texture, sometimes facilitated by using fluorescent beads or imprinted micropatterns. Although these inert scaffolds provide a geometrical framework to study processes like cell locomotion, they do not resemble the physiological ECM. This has been solved in some cases by coating the polyacrylamide membrane with a thin layer of collagen [11, 17, 26, 27, 37]. However, remaining issues are the orientation of the collagen fibrils in the coating as compared to native collagen, and the inability of the cells to chemically modify (e.g. by cross-linking) the artificial sheets.

A possible way to overcome these issues is the use of collagen-based matrices. Cell-induced deformation patterns of these matrices and the degree of reversibility of such deformation provide information on physical remodeling and the underlying biochemical processes. Deformations have been assessed by manually tracing beads or landmarks in consecutive images [6, 28, 30]. However, so far, only a few algorithms were developed for automatic detection of substrate deformations. In particular, very few of such algorithms allow tracing in the absence of embedded beads [23, 33, 35], and more stable algorithms are required.

While single cell observations provide a more fundamental insight into matrix remodeling than macroscopic studies, a concern is efficiency. Time-lapsed video microscopy of single cells forms a labor-intensive and time-consuming experiment. Meaningful experiments require the comparison of large data sets, including different cell types, various matrix compositions, or the use of biochemical or molecular interventions. A fully automated, motorized microscopy setup is required that scans series of individual cells and their surrounding matrix. In addition, each individual cell should automatically be kept in focus during remodeling experiments, which take up 24 h or more. Finally, stable algorithms are needed for the automated analysis of geometrical reorganization with minimal user input.

The aim of this study is to develop an automated technique that can be employed to obtain and analyze local matrix remodeling by individual cells. The system that we present allows for monitoring of ∼75 cells in parallel, using time-lapse video microscopy and computer-controlled stage positioning. In addition, we present and evaluate a new algorithm for automated detection of collagen matrix deformation around these cells.

This study aimed at developing an automated technique for obtaining and analyzing matrix remodeling by individual cells. Emphasis was put on construction of an automatic, reliable algorithm for assessment of a detailed matrix displacement field. Especially, refinement of a cross-correlation based image analysis with a decomposition scheme was investigated. While “classic” direct CC sufficed for pairs of images with high correlation and low noise, this was no longer the case when substantial matrix remodeling occurred within the time frame between two consecutive images. This resulted in failing of CC at spots of high geometrical reorganization. However, when using decomposition CC, the gross displacements at these positions could be estimated by analysis of parent images with larger dimensions. This way, at a RD increase of 20.2% the average displacement error was lowered threefold in decomposition CC as compared to direct CC.

The method of refining displacement field accuracy with each decomposition step becomes progressively more important with larger displacements. Therefore we investigated the effect of a horizontal shift superimposed on a simulated compaction. Such shift occurred in our in vitro experiments when a group of neighboring cells pulled strongly on the matrix adjacent to a cell of interest. The shift interfered with direct CC because zero deformation was assumed when no proper correlation could be found. The result was an irregular displacement field (see Fig. 4). When using decomposition CC, on the other hand, the gross displacement was already detected in the first decomposition stage. Displacement values were then adjusted when local compaction was detected in subsequent decomposition stages. Using this strategy, a lack of local correlation at the finest resolution resulted in only a minor displacement error, with an average of 0.9 as compared to 14.6 pixels for direct CC.

Both cross-correlation methods differ not only in stability of displacement field estimation, but in efficiency of calculation time as well. Direct CC requires a large search area, i.e. the subwindows of I ₂ have to extend considerably beyond the boundaries of their corresponding I ₁ subwindow in order to perform a meaningful correlation. On the other hand, with decomposition CC n _cc can decrease with each stage (see Table 1), since each subwindow of I ₂ is repositioned according to the preliminary displacement as calculated for its parent window. Due to the increased number of calculations and data manipulations that have to be performed, analysis time was still about 50% larger for decomposition CC at the settings as stated in Table 1.

We chose a correlation threshold of 0.5 for both techniques, as well as cross-correlation windows as indicated in Table 1. These values were empirically determined as an optimum in the trade-off between calculation time, false positive displacements and overlooking local deformations. Clearly, these choices depend on the contrast and texture of the images and the nature of the deformation, and will need to be optimized for specific future experiments.

Cell traction is frequently assessed by quantification of deformations in a flexible substratum. Both 2D and 3D approaches have been used, resulting in different cellular morphologies [19, 20, 29]. While the latter seems more physiological, interpretation of the 3D experiments is more complex [35]. 2D substrates offer more straightforward tools for analyzing mechanical behavior of single cells. These materials can be enriched with fluorescent microbeads to increase image contrast. In order to achieve a high resolution in the vicinity of cells, which can change their morphology rapidly, manual tracking of specific landmarks was employed in several studies [6, 28, 30]. However, this labor-intensive method inherently limits the number of cells under study. In several studies particle tracking was performed on individual microbeads [11]. In this case, resolution depends on particle distribution, which is typically far from homogeneous in collagen substrates [23].

Several nested cross-correlation methods have been developed [35]. All of these algorithms are based on empirical determination of square size, distance for pattern search, and normalized CC threshold. In order to spatially limit the search region, a relative translational shift has been derived from image registration [23, 33]. Several schemes have been implemented to increase displacement field resolution in a step-by-step manner. In a single refinement cycle, Dembo and colleagues [26] decreased the subwindow size if significant displacement was observed. On the other hand, window size can be increased if correspondence failure occurs at high resolution [23]. The algorithm developed by Wang and coworkers [33] resembled our decomposition CC to a large extent. However, these authors aimed at construction of traction fields with a smooth nature by incorporation of filtering procedures at several stages.

Geometrical matrix reorganization provides a qualitative index of the traction forces present in the underlying material. The actual forces can be derived from a displacement field series using material properties of the ECM construct. However, for collagen scaffolds these are highly heterogenous. Frequently, local stiffness is estimated by either microneedles [5, 10, 21, 30] or optical tweezers [4, 15, 32]. Tractions can then be calculated by application of stress–strain relationships, including appropriate boundary conditions [7, 24, 25, 31, 37, 38]. However, this translation to quantitative traction forces lies beyond the scope of this article.

In conclusion, we presented integral methodology for the study of matrix remodeling. These techniques allow systematic screening of the role of matrix components such as collagen, elastin, fibronectin and laminin. Likewise, the function of stationary cells like smooth muscle cells, fibroblasts, and osteoblasts can be investigated. Our method can be applied under a wide variety of other experimental conditions. The single requirement for the image quality is a sufficiently high contrast in the material under investigation, without the neeed for laborious and potentially interfering micropatterning. Furthermore, our graphical user interface enables a flexible tuning of parameters such as number of decompositions, size of correlation window, search area and cross-correlation threshold. Using our motorized microscopy stage it is possible to patch individual images together in order to create one large field of view for the study of motile cells such as keratocytes. Finally, the method can be extended with fluorescence imaging of specific cell structures, cytokines, hormones and enzymes [15].