Topic Introduction

Computational Image Analysis of Cellular Dynamics: A Case Study Based on Particle Tracking

Adapted from Live Cell Imaging, 2nd edition (ed. Goldman et al.). CSHL Press, Cold Spring Harbor, NY, USA, 2010 (in press).

INTRODUCTION

Obtaining quantitative data from live cell images is the key to testing mechanistic hypotheses of molecular and cellular processes. The importance of using computer vision-based methods to accomplish this task is well recognized. However, in practice, investigators often encounter obstacles that render the application of computational image processing in cell biology far from routine. First, it is not always clear which measurements are necessary to characterize a molecular system and whether these measurements are sufficient to characterize the cellular process under investigation. Second, even when the requirements for measurements are well defined, it often is difficult to find a software tool to extract these data. It can be even more challenging to find software tools to answer specific questions raised by the hypotheses underlying the experiments. One solution is for investigators to develop their own software tools. This is feasible for some applications with the assistance of commercial and open-source software packages that support the assembly and integration of custom-designed algorithms, even for users with limited computational expertise. Another solution is for investigators to develop interdisciplinary collaborations with computer scientists. Such collaborations require close interaction between the computer scientists and experimental biologists to optimize the data acquisition and analytical procedures, which must be tightly coupled in any project applying computational analysis to biological image data. This article introduces the basic concepts that make the application of computational image processing to live cell image data successful.

RELATED INFORMATION

Fuller discussions of the importance of computer vision-based methods in the acquisition of quantitative data from live cell images are presented in Eils and Athale (2003), Swedlow et al. (2003), and Dorn et al. (2008). Although the concepts discussed in this article are applicable generally, specific examples are presented from a case study of particle tracking (PT), one of the most frequently encountered problems in cell biology.

WHY USE COMPUTATIONAL IMAGE ANALYSIS?

Efficiency

Efficient extraction of quantitative measurements is a major motivation for the use of computational image analysis, especially in the context of screens. With the development of microscopes for live cell genome-wide screens (Smith and Eisenstein 2005; Bakal et al. 2007), it is possible to acquire vast amounts of data in ever-shorter times. For example, even at low spatiotemporal sampling, a live cell small interfering RNA (siRNA) screen of 49 mitotic genes can generate >100 GB of image data (Neumann et al. 2006). Such quantities make data management in general challenging and manual data analysis in particular unrealistic. Instead, such experiments require computational image analysis to extract image features for the classification of cell behavior in response to perturbations. For screens in particular, robustness is vital. Thus, for the most part, simple algorithms can be applied to produce meaningful features without the need for manual validation of image analysis results (Abraham et al. 2004). Alternatively, robustness can be achieved by manually training the computer to recognize a small number of phenotypes (Conrad et al. 2004; Chen and Murphy 2006).

Consistency

Computational image analyses yield consistent data, i.e., different experiments are processed using the same parameter settings and criteria to validate measurements. This eliminates uncertainty associated with subjective interpretations of image contents between investigators, or even by a single investigator analyzing different sets of data. Furthermore, computational image analysis permits the quantification of measurement uncertainty originating from noise in the raw imagery. High consistency and known uncertainty are particularly useful when the study of a certain cell function demands distinction between weak yet significant phenotypes (Dorn et al. 2005).

Completeness

Computational image analysis yields complete data: Every image event that fulfills an objective set of criteria is considered. Humans have a tendency, by nature or necessity, to concentrate on the apparently interesting events. This can bias analyses and increase the risk of overlooking rare events associated with weaker phenotypes. In contrast, complete image measurements permit the statistical selection of obvious and less obvious events, including highly transient events. Image transients are particularly relevant to establish functional linkages between the dominant image events.

CASE STUDY: PARTICLE TRACKING

Live cell images often consist of large numbers of punctate features (i.e., particles) representing single fluorophores tagging single molecules (Sako et al. 2000; Fujiwara et al. 2002; Groc et al. 2004), fluorophore clusters associated with subresolution molecular assemblies (Zenisek et al. 2000; Ewers et al. 2005; Danuser and Waterman-Storer 2006), or fluorophore blobs associated with vesicles or more extended organelles (Ehrlich et al. 2004; Tirnauer et al. 2004). To capture the full spatiotemporal complexity of subcellular particle dynamics and to link them to the underlying molecular processes, data must be extracted from live cell images using automated PT techniques.

PT consists of two major steps: (1) particle detection in each frame of a time-lapse sequence and (2) particle trajectory construction across the time-lapse sequence (Fig. 1 ). Although in some models, particle detection and trajectory construction are coupled and feed back into each other (Ponti et al. 2005; Racine et al. 2007), in most computational analysis frameworks, the information flows in one direction, from detection to trajectory construction. In either case, trajectory construction can be assisted by using particle-motion models that predict the particle positions in a frame based on the positions in the past, thus reducing the ambiguity of establishing particle correspondences between frames. Furthermore, PT generally must include a trajectory-diagnosis module to assess the quality of the tracking results and through which the tracking and motion modeling parameters can be optimized. The remainder of this section discusses the detection, trajectory construction, and motion-modeling modules; the following sections describe the design of experiments to yield image data optimized for automated image analysis and the diagnosis of the results of image analysis required to assess tracking quality and to adjust analysis parameters.

View larger version (26K): [in this window] [in a new window]

Figure 1. Particle tracking builds on essential steps (detection and trajectory construction) and optional but recommended steps (motion modeling and trajectory diagnosis). Image acquisition and analysis are tightly coupled.

Detection

The goal of particle detection is to obtain numerical representations of the location and properties of image features (Starck et al. 2000; Nixon and Aguado 2002). Image features are local intensity maxima whose intensity level is significantly different from their neighborhood. Consequently, particle-detection techniques must establish definitions for "neighborhood" (i.e., for the computation of representative background intensity distribution) and "significantly different." The most rigorous way to accomplish this is to cast the comparison of foreground-to-background intensity as a statistical test.

Subresolution features above a dark background, as encountered in single-molecule imaging, can be detected by comparing the intensity of local maxima with the local background intensity distribution (Jaqaman et al. 2008). For low signal-to-noise ratio (SNR) time-lapse sequences (e.g., SNR < 3, where SNR is the ratio of signal-above-mean-local-background to local-background variation), image time averaging can be used to enhance detection efficiency (Jaqaman et al. 2008). If features are subresolution but lie above a sea of fluorescence, for example in the case of speckles marking dense macromolecular assemblies, more sophisticated algorithms must be used to compare local intensity maxima with their neighboring local intensity minima (given a precalibrated model-of-camera noise; Ponti et al. 2003).

After the detection of significant local maxima, the subpixel positions and peak intensities of particles can be estimated via point-spread-function (PSF) fitting (Thomann et al. 2002; Yildiz and Selvin 2005; Jaqaman et al. 2008). For particles in isolation, the achieved positional precision depends only on the SNR; single-nanometer precision can be achieved if sufficient photons are collected (Yildiz and Selvin 2005). For particles not in isolation, iterative PSF fitting can be used to obtain unbiased position estimates and at the same time enhance resolution in detecting closely juxtaposed particles (Thomann et al. 2002; Dorn et al. 2005; Jaqaman et al. 2008). Based on simulations and indirect experimental evidence, iterative PSF fitting can overcome the diffraction-limited resolution of a microscope by a factor of 2-3 (Thomann et al. 2002). Thus, distances of 100 nm can be measured without the use of superresolution imaging (Bates et al. 2007; Shroff et al. 2007). The methods described here for the detection and localization of subresolution features are readily applicable in both two and three dimensions.

PSF fitting cannot be applied to detect particles representing objects larger than the diffraction limit, especially if their size varies. For particles with variable size but that are still relatively isotropic, wavelet-based algorithms can be applied (Olivo-Marin 2002). For particles with additional shape variations, an edge-detection-based algorithm to find particle contours can be used (Tvaruskó et al. 1999). Whereas the only properties of subresolution features are position and intensity, larger particles can also be described by their size and shape. These additional characteristics provide valuable information to support the construction of particle trajectories. Note, however, that the detection of anisotropic larger image features in three dimensions is a difficult problem currently without a general solution.

Trajectory Construction

Arguably, the key step of PT is the establishment of the correspondence between particle images in a sequence of frames to construct particle trajectories throughout the time-lapse sequence. Establishing correspondence is complicated by various factors, most notably high particle density, particle-motion heterogeneity, temporary particle disappearance (e.g., resulting from out-of-focus motion and detection failure), particle merging (i.e., two particles approaching each other within distances below the resolution limit), and particle splitting (i.e., two unresolved particles diverging to resolvable distances) (Meijering et al. 2006; Kalaidzidis 2007). Historically, many of these challenges have been overcome by diluting the fluorescent probes, resulting in a low particle density with almost unambiguous particle correspondence (Ghosh and Webb 1994; Crocker and Grier 1996). Under such conditions, PT is indeed reduced to a simple particle detection and localization problem. However, although low particle densities reveal motion characteristics, they do not allow probing of the interactions between particles. Also, the amount of data collected per experiment is low, limiting the observation of spatially and temporally heterogeneous particle behavior and hindering the capture of rare events. Furthermore, even with low particle density, low SNR and probe flicker complicate the search for particle correspondence. Therefore, for most cell biological studies, there is a great need for robust trajectory-construction methods that address the challenges mentioned above.

In cases of very low density where the ratio between particle displacement and the mean nearest-neighbor distance is significantly less than 0.5, particle frame-to-frame assignment can be achieved via a simple local-nearest-neighbor (LNN) algorithm (Fig. 2A ). Stepping through the list of particles in one frame, particles are linked to the closest particle in the next frame. The Crocker and Grier particle-tracking package, one of the most widespread trackers using LNN, can be downloaded from http://www.physics.emory.edu/~weeks/idl/.

View larger version (13K): [in this window] [in a new window]

Figure 2. Trajectory construction via LNN assignment. (A) LNN succeeds when the ratio () of the average frame-to-frame displacement/average nearest-neighbor distance << 0.5. (B) LNN fails when > ~0.2. Candidate assignments are represented as cands.

The LNN approach breaks down when particle density is high enough such that particles have more than one candidate assignment in the next frame. The outcome of a LNN algorithm in these situations depends on the order by which the assignments are made. In the example shown (Fig. 2B), if the triangle correspondence is assigned before the circle correspondence (sequence 1), then the triangle and circle will be assigned incorrectly. In contrast, if the circle correspondence is assigned before the triangle (sequence 2), then the assignments for all three interfering particles will be correct. In general, the best order of particle assignments is undefined. In some cases, simple heuristics can be sufficient to remedy the situation (Ponti et al. 2003). However, in general, a global solution is required to achieve satisfactory tracking results.

The most accurate and globally optimal solution to PT is provided by the method of multiple-hypothesis tracking (MHT; Reid 1979). In MHT, given the particle positions in every frame, all particle paths within the bounds of expected particle behavior are constructed throughout the whole movie. The largest nonconflicting ensemble of paths is then chosen as the solution (where "nonconflicting" means that no two paths share the same particle in any frame). This solution is globally optimal in both space and time, providing the best solution that can be found by simultaneously accounting for all particle positions at all time points. Clearly, MHT is computationally prohibitive even for problems with a few tens of particles tracked over a few tens of frames.

Heuristic algorithms with higher computational efficiency have been proposed to approximate the MHT solution. Most of these algorithms are "greedy"; that is, they seek to approach the globally optimal solution by taking a series of locally optimal solutions. Usually, this means that particle correspondence is determined step-by-step between consecutive frames, reducing computational complexity at the expense of temporal globality. Many tracking algorithms, referred to as global-nearest-neighbor (GNN) approaches, then solve the frame-to-frame correspondence problem in a spatially global manner. GNN approaches have been developed for radar tracking and computer vision, and many have been applied recently to cell biological studies (Vallotton et al. 2003; Bonneau et al. 2005; Sage et al. 2005; Sbalzarini and Koumoutsakos 2005; Shafique and Shah 2005; Genovesio et al. 2006). The Sbalzarini and Koumoutsakos particle-tracking package can be downloaded from http://www.mosaic.ethz.ch/Downloads/ParticleTracker. Other algorithms deal with the additional factors complicating PT such as temporary particle disappearance (Chetverikov and Verestóy 1999; Veenman et al. 2001; Bonneau et al. 2005; Sbalzarini and Koumoutsakos 2005; Shafique and Shah 2005; Genovesio et al. 2006), particle merging and splitting (Genovesio and Olivo-Marin 2004; Jiang et al. 2007), and particle-motion heterogeneity (Genovesio et al. 2006).

A recently introduced PT algorithm for cell biological applications is described by Jaqaman et al. (2008); the particle-tracking package can be downloaded from http://lccb.scripps.edu/. It uses a single, efficient mathematical framework, the linear assignment problem (LAP; Burkard and Çela 1999), to provide an accurate solution to all the PT challenges listed above. Given a set of detected particles throughout a time-lapse image sequence, the algorithm first links the detected particles between consecutive frames and then links the track segments generated in the first step to simultaneously close gaps and capture particle merge and split events. Thus, the initial particle assignment is spatially global but temporally greedy, whereas the subsequent track-segment assignment is accomplished via spatially and temporally global optimization, thereby overcoming the shortcomings of algorithms that rely solely on greedy assignment strategies. The algorithm is general and can be applied to both two- and three-dimensional problems. Overall, this approach defines an accurate yet computationally feasible approximation of MHT, permitting the robust tracking of particles under high-density conditions, as is typically found in live cell images.

Motion Modeling

The robustness of GNN assignment under high-density conditions can be increased by motion prediction. In this model, assignments are no longer made based on particle positions in the target frame t + 1 and the source frame t, but rather on particle positions in the target frame t + 1 and the predicted positions of the particles from the source frame t to the target frame t + 1. There are a number of possible approaches to particle-motion prediction between frames, such as estimating the global organization of particle motion iteratively from the available particle assignments (Ponti et al. 2005) or other tracking methods (Ji and Danuser 2005), or formulating explicit motion models for each particle, the parameters of which are inferred based on the already tracked particle paths (Genovesio et al. 2006; Jaqaman et al. 2008).

ACQUISITION OF OPTIMIZED FLUORESCENT IMAGES

PT algorithms will fail to capture live cell dynamics unless image acquisition is adjusted to the process of interest in terms of spatial and temporal sampling, SNR, and the movie length necessary to capture all possible process states. However, these imaging parameters are interdependent and conflict with each other. For example, high spatiotemporal sampling implies fast acquisition at high magnification, resulting in fewer photons reaching the imaging sensor and thus a low SNR. Conversely, although SNR can be improved by prolonging the exposures or by increasing the power of the illumination, this will increase photobleaching and phototoxicity, limiting the number of possible exposures and hence observation length.

Image-analysis algorithms impose an additional layer of conflicting requirements on data acquisition. For example, tracking quality decreases as the ratio of particle frame-to-frame displacements to interparticle distances increases. To improve tracking quality at the same particle density, images must be acquired faster. However, faster acquisition can reduce the image SNR, thus reducing detection quality and leading to more temporary particle disappearances. Similarly, the occurrence of temporary particle disappearances increases the risk of erroneous particle linking between frames under high-particle-density conditions. Image acquisition and tracking parameters must thus be adjusted iteratively to optimize tracking quality and minimize tracking errors (Fig. 1). As a general rule, image acquisition and analysis are tightly coupled in a quantitative live cell imaging project.

To design a quantitative imaging experiment, the minimum requirements of spatiotemporal sampling, SNR, and observation length need to be defined, and their compatibility with the available specimen and microscope hardware must be tested. For each of the three imaging parameters, the specimen and microscope hardware define a maximum performance point, which can be derived from the microscope specifications (e.g., fastest acquisition rate of the camera, highest magnification) or can be determined experimentally (acquisition time before bleaching or phototoxic damaging of the specimen; SNR obtained under very long exposures, e.g., in fixed specimens). Given the mutual interdependence between the parameters, the joint performance of an experimental setup can be conceptualized by the plane through the three maximum performance points (Fig. 3 ). If the minimum requirements of a specific experiment fall in a point beyond the performance plane of an experimental setup, it is impossible to acquire all the necessary image data using that setup.

View larger version (30K): [in this window] [in a new window]

Figure 3. Performance triangle of an experimental setup, as determined by the specimen, microscope hardware, and image-analysis software. Spatiotemporal sampling, SNR, and observation length are interdependent, and conflicting, imaging parameters. (Modified from Dorn et al. 2008 and reprinted with permission from Elsevier © 2008.)

There are two solutions to the problem of an insufficient experimental setup. First, the setup can be redesigned, for example, by investing in better microscopy hardware or by improving the stability and efficiency of the fluorescent probes. Second, one can compromise at the level of individual movies and instead combine data from different experiments at the analytical level under the assumption that cells imaged in different experiments are statistically equivalent. For example, data from short movies of fast temporal sampling were combined with data from long movies using slow temporal sampling to obtain comprehensive coverage of the wide distribution of clathrin-coated pit lifetimes (Loerke et al. 2009). The following sections provide some general guidelines on how to determine the minimal requirements for a specific experiment.

Sampling

To allow any computational image analysis of the spatiotemporal dynamics of a live cell, the specimen must be sampled at least three times finer than the highest spatial and temporal frequencies of interest (Stelzer 2000). Approximating the microscope three-dimensional PSF by an ellipsoid with short (and equal) semi-axes in the lateral direction and a long semi-axis in the axial direction, sufficient sampling means that the magnification of the microscope must be selected such that (1) the pixel side length is at least one-third the PSF short semi-axis in the lateral direction, and (2) the z-slice thickness is at least one-third the PSF long semi-axis in the axial direction (Inoué and Spring 1997).

For sampling in time, the characteristic timescale of the probed dynamics either is assumed a priori based on previous work or simulations of the molecular processes of interest, or is determined by analyzing the dynamic data themselves. For example, a common technique to estimate the diffusion coefficient of a particle undergoing confined diffusive motion is to plot the mean square displacement (MSD) of the particle over time lag. If the particle dynamics are well sampled, the MSD first grows linearly with time and then reaches a plateau reflecting the confinement radius (Fig. 4A , black dots). The initial linear part of the MSD plot can yield a good estimate of the diffusion coefficient, estimated by fitting a straight line through the first few points of the MSD curve (Fig. 4A, black line; Huet et al. 2006). On the other hand, for undersampled dynamics, the particle bounces from the boundaries many times within the sampling period. As a consequence, the linear phase of the MSD plot vanishes, precluding any accurate estimate of the diffusion coefficient (Fig. 4A, cyan and red symbols and lines). Similarly, if a particle undergoes periodic or quasiperiodic movements, the observed particle positions are dictated by the number of motion reversals within a sampling interval. If the dynamics are undersampled, the measured speed is inversely proportional to the sampling interval. The optimal sampling interval can thus be determined by first acquiring image data at the maximum sampling rate affordable, ignoring the limitations that too-fast sampling imposes on the observation length. Then, the image sequence can be downsampled artificially and the optimal sampling interval determined as the one just before the measured velocity versus time interval starts to deviate from a horizontal line (Fig. 4B).

View larger version (24K): [in this window] [in a new window]

Figure 4. Data analysis to ensure proper temporal sampling of the measured dynamics. (A) Effect of sampling rate on diffusion-coefficient estimation (line fits) for confined diffusive motion. Time lag is represented in seconds (s). (B) Effect of sampling rate on speed estimation for periodic and quasiperiodic movements.

SNR

SNR requirements are determined entirely by the image-analysis algorithm. Given a high SNR image of the specimen, e.g., one taken with long exposures of a fixed sample, the detection fidelity and the breakdown of an algorithm can be identified by the simulation of increasing noise levels on this image. Subsequently, the illumination conditions and the exposure times that produce an SNR above the breakdown limit can be determined experimentally. These imaging parameters must be defined such that the SNR conditions are satisfied at the end of the time-lapse image sequence, where the effect of photobleaching is strongest.

Observation Length

The observation length required to capture all possible states of a dynamic molecular system is the most difficult criterion to determine a priori. Data from multiple experiments must be pooled until there is statistical evidence that all system states have been sampled (e.g., parameter distributions converge to a fixed point), under the assumption that all cells behave equivalently. Because of the ability to pool data from multiple experiments, the observation length is often the least essential criterion to satisfy in a single experiment. In contrast, too-slow sampling or a too-low SNR will lead to the loss of primary information that cannot be recovered by data pooling.

ADJUSTMENT OF CONTROL PARAMETERS AND DIAGNOSTICS FOR TRACK EVALUATION

Maximum efficiency, consistency, and completeness in image measurements imply minimal user input for software control. Nevertheless, it is impossible to design image-analysis algorithms that are universally robust to achieve complete understanding of image contents in all applications. User input is always required to produce an algorithm with application-specific prior knowledge. The amount of information provided and the effect such information has on the outcome of image measurements must be analyzed carefully with every experiment. Also, a practical compromise between total automation (which might be impossible) and complete dependence on user-specified parameters (which can bias the data) can be achieved by designing self-adaptive algorithms that learn parameters during the process of image analysis with input from the user to define lower and upper boundaries to prevent drifts in self-adaptation. For example, instead of depending on a user-specified search radius for linking particles between frames, Jaqaman et al. (2008) used software to estimate the search radius from the constructed particle trajectories and relied on user input only to define the very fastest speed expected for a particle. Also, although the user might specify whether the algorithm considers only Brownian motion or both Brownian and linear motion, the decision as to whether an individual particle is undergoing linear or Brownian motion is determined by the software, as are the parameters characterizing each motion model.

Independent of the level of user input, software outputs always must be benchmarked carefully. Whereas visual inspection is a good first practice, in many cases of PT the visual impression of particle dynamics can be deceiving,and many significant particle behaviors are simply missed. Thus, before manual measurements are accepted as the ground truth for benchmarking, the inter- and intra-operator variability of the manual data set should be determined and documented. Several PT projects have reported performances as low as 50% and 70%, respectively (G. Danuser, unpubl.); such data are insufficient for evaluating the results of computational PT. The following sections discuss two approaches that allow a more objective benchmarking of tracking outputs.

Simulation-Based Benchmarking

Simulation experiments provide a means for determining absolute measures of false positives, false negatives, and other performance parameters of tracking under different conditions. For example, the selectivity of a point detector, which relies on a statistical test to distinguish particle signal from noise (Ponti et al. 2003; Jaqaman et al. 2008), is controlled via the confidence threshold required for accepting a local maximum as a particle. Lowering the threshold increases the number of false positives; raising the threshold increases the number of false negatives. The ratio between the two fractions is a nonlinear function of the threshold and the movie SNR. Thus, using simulated images of various SNRs, the performance of the point detector (i.e., the number of false positives and negatives it generates) can be evaluated. Such a performance graph can then be used to determine the optimal threshold for a movie with a particular SNR. Similarly, using ground truth tracks, the quality of the trajectory-construction algorithm can be evaluated as a function of particle density and movie SNR (Jaqaman et al. 2008). This identifies the breakdown point of the trajectory-construction algorithm and determines the expected quality of the constructed trajectories given the particle density in a movie and its SNR.

Data-Based Diagnostics

In addition to simulation-based software benchmarking, diagnostics that analyze and evaluate the particle trajectories obtained from the experimental images can be used to optimize the trajectory-construction parameters. For example, in determining the search radius for particle linking between frames, the distribution of particle displacements obtained from the tracking must be investigated for distortions. If the displacement histogram is cut off (Fig. 5A ), the search radius has been set to a value that is too small. In contrast, if the histogram decays gradually to zero (Fig. 5B), the search radius is large enough to capture all possible displacements. It is important to note that, when the trajectory-construction algorithm employs motion propagation, one should not analyze the particle frame-to-frame displacements but rather the distances between particle-propagated positions and the particles to which they are linked. Thus, in Figure 5A,B, the x-axes are labeled as frame-to-frame linking distance and not frame-to-frame displacement.

View larger version (31K): [in this window] [in a new window]

Figure 5. Trajectory diagnosis for evaluation of tracking. (A) A cut-off histogram of frame-to-frame linking distances is a sign that the search radius is too small. (B) A slowly decaying histogram shows the frame-to-frame linking distances, indicating a sufficiently large search radius. (C) Sampling rate-independent distribution of gap lengths, expressed in frames. (D) Histogram of gap lengths comparing a too-large gap closing time window (gray bars) and an appropriate time window (black bars).

Another parameter critical to optimizing trajectory construction is the time window for closing trajectory gaps resulting from temporary particle disappearance (Jaqaman et al. 2008). In live cell time-lapse sequences, particles temporarily disappear either because of detection false negatives or because of random particle motion in and out of focus. Two consequences follow from this: First, trajectory gap length distributions, when measured in frames, should be independent of movie sampling rate, provided the same exposure time is used (Fig. 5C). Second, for any sampling rate, longer gaps should be encountered less frequently than shorter gaps. Thus, a plateau in the tail of a histogram of gap lengths indicates that the time window used for gap closing is too large, resulting in falsely closed gaps (Fig. 5D, gray histogram). In such a case, the time window for gap closing should be reduced until there is no longer a plateau (Fig. 5D, black histogram).

CONCLUSION

Computational image analysis is a complex yet increasingly central component of live cell imaging experiments. Much has to be done to make these techniques useful for cell biological investigation. First, algorithms must be transparent, not necessarily at the level of the code but in terms of their sensitivity to changing image quality and the effect that control parameters have on the output. Second, the design of imaging experiments must be tightly coupled to the design of the analysis software. All too often, images are taken without careful consideration of the subsequent analysis and are forwarded to the computer scientist to retrieve information from the images. To avoid these problems, communication must be initiated early on, and experiments must be designed with the appreciation that data acquisition and analysis are equivalent components. Third, software development and application require careful controls, as is customary for molecular cell biology experiments. This article provides a brief introduction to the ideas useful for implementing such controls. Hopefully, the cell biological literature will include a more extensive discussion of the measures taken to substantiate the validity of results from image analysis. On the other hand, manual image analysis should no longer be an option. As discussed in this article, manual analyses fall short in consistency and completeness, two essential criteria underlying the validity of a scientific model derived from image data.

REFERENCES

Abraham VC, Taylor DL, Haskins JR. 2004. High content screening applied to large-scale cell biology. Trends Biotechnol 22: 15–22.[Medline]

Bakal C, Aach J, Church G, Perrimon N. 2007. Quantitative morphological signatures define local signaling networks regulating cell morphology. Science 316: 1753–1756.[Abstract/Free Full Text]

Bates M, Huang B, Dempsey GT, Zhuang X. 2007. Multicolor super-resolution imaging with photo-switchable fluorescent probes. Science 317: 1749–1753.[Abstract/Free Full Text]

Bonneau S, Dahan M, Cohen LD. 2005. Single quantum dot tracking based on perceptual grouping using minimal paths in a spatiotemporal volume. IEEE Trans Image Process 14: 1384–1395.[Medline]

Burkard RE, Çela E. 1999. Linear assignment problems and extensions. In Handbook of combinatorial optimization (eds. D-Z Du and PM Pardalos), Vol. A, pp. 75–149. Kluwer Academic, Dordrecht, Netherlands.

Chen X, Murphy RF. 2006. Automated interpretation of protein subcellular location patterns. Int Rev Cytol 249: 193–227.[Medline]

Chetverikov D, Verestóy J. 1999. Feature point tracking for incomplete trajectories. Computing 62: 321–338.

Conrad C, Erfle H, Warnat P, Daigle N, Lörch T, Ellenberg J, Pepperkok R, Eils R. 2004. Automatic identification of subcellular phenotypes on human cell arrays. Genome Res 14: 1130–1136.[Abstract/Free Full Text]

Crocker JC, Grier DG. 1996. Methods of digital video microscopy for colloidal studies. J Colloid Interface Sci 179: 298–310.

Danuser G, Waterman-Storer CM. 2006. Quantitative fluorescent speckle microscopy of cytoskeleton dynamics. Annu Rev Biophys Biomol Struct 35: 361–387.[Medline]

Dorn JF, Jaqaman K, Rines DR, Jelson GS, Sorger PK, Danuser G. 2005. Yeast kinetochore microtubule dynamics analyzed by high-resolution three-dimensional microscopy. Biophys J 89: 2835–2854.[Medline]

Dorn JF, Danuser G, Yang G. 2008. Computational processing and analysis of dynamic fluorescence image data. In Fluorescent proteins (ed. KF Sullivan), 2nd ed, pp. 497–538. Elsevier, San Diego, CA.

Ehrlich M, Boll W, van Oijen A, Hariharan R, Chandran K, Nibert ML, Kirchhausen T. 2004. Endocytosis by random initiation and stabilization of clathrin-coated pits. Cell 118: 591–605.[Medline]

Eils R, Athale C. 2003. Computational imaging in cell biology. J Cell Biol 161: 477–481.[Abstract/Free Full Text]

Ewers H, Smith AE, Sbalzarini IF, Lilie H, Koumoutsakos P, Helenius A. 2005. Single-particle tracking of murine polyoma virus-like particles on live cells and artificial membranes. Proc Natl Acad Sci 102: 15110–15115.[Abstract/Free Full Text]

Fujiwara T, Ritchie K, Murakoshi H, Jacobson K, Kusumi A. 2002. Phospholipids undergo hop diffusion in compartmentalized cell membrane. J Cell Biol 157: 1071–1081.[Abstract/Free Full Text]

Genovesio A, Olivo-Marin J-C. 2004. Split and merge data association filter for dense multi-target tracking. In Proceedings of the 17th international conference on pattern recognition (eds. J Kittler et al.), Vol. 4, pp. 677–680. IEEE, Piscataway, NJ.

Genovesio A, Liedl T, Emiliani V, Parak WJ, Coppey-Moisan M, Olivo-Marin JC. 2006. Multiple particle tracking in 3-D+t microscopy: Method and application to the tracking of endocytosed quantum dots. IEEE Trans Image Process 15: 1062–1070.[Medline]

Ghosh RN, Webb WW. 1994. Automated detection and tracking of individual and clustered cell surface low density lipoprotein receptor molecules. Biophys J 66: 1301–1318.[Medline]

Groc L, Heine M, Cognet L, Brickley K, Stephenson FA, Lounis B, Choquet D. 2004. Differential activity-dependent regulation of the lateral mobilities of AMPA and NMDA receptors. Nat Neurosci 7: 695–696.[Medline]

Huet S, Karatekin E, Tran VS, Fanget I, Cribier S, Henry J-P. 2006. Analysis of transient behavior in complex trajectories: Application to secretory vesicle dynamics. Biophys J 91: 3542–3559.[Medline]

Inoué S, Spring KR. 1997. Video microscopy: The fundamentals. Plenum, New York.

Jaqaman K, Loerke D, Mettlen M, Kuwata H, Grinstein S, Schmid SL, Danuser G. 2008. Robust single-particle tracking in live-cell time-lapse sequences. Nat Methods 5: 695–702.[Medline]

Ji L, Danuser G. 2005. Tracking quasi-stationary flow of weak fluorescent signals by adaptive multi-frame correlation. J Microsc 220: 150–167.[Medline]

Jiang S, Zhou X, Kirchhausen T, Wong STC. 2007. Tracking molecular particles in live cells using fuzzy rule-based system. Cytometry A 71A: 576–584.[Medline]

Kalaidzidis Y. 2007. Intracellular objects tracking. Eur J Cell Biol 86: 569–578.[Medline]

Loerke D, Mettlen M, Yarar D, Jaqaman K, Jaqaman H, Danuser G, Schmid SL. 2009. Cargo and dynamin regulate clathrin-coated pit maturation. PLoS Biol 7: e1000057. doi: 10.1371/journal.pbio.1000057.

Meijering E, Smal I, Danuser G. 2006. Tracking in molecular bioimaging. IEEE Signal Process Mag 23: 46–53.

Neumann B, Held M, Liebel U, Erfle H, Rogers P, Pepperkok R, Ellenberg J. 2006. High-throughput RNAi screening by time-lapse imaging of live human cells. Nat Methods 3: 385–390.[Medline]

Nixon M, Aguado A. 2002. Feature extraction in computer vision and image processing. Butterworth-Heinemann/Newnes, Oxford, UK.

Olivo-Marin J-C. 2002. Extraction of spots in biological images using multiscale products. Pattern Recognit 35: 1989–1996.

Ponti A, Vallotton P, Salmon WC, Waterman-Storer CM, Danuser G. 2003. Computational analysis of F-actin turnover in cortical actin meshworks using fluorescent speckle microscopy. Biophys J 84: 3336–3352.[Medline]

Ponti A, Matov A, Adams M, Gupton S, Waterman-Storer CM, Danuser G. 2005. Periodic patterns of actin turnover in lamellipodia and lamellae of migrating epithelial cells analyzed by quantitative fluorescent speckle microscopy. Biophys J 89: 3456–3469.[Medline]

Racine V, Sachse M, Salamero J, Fraisier V, Trubuil A, Sibarita J-B. 2007. Visualization and quantification of vesicle trafficking on a three-dimensional cytoskeleton network in living cells. J Microsc 225: 214–228.[Medline]

Reid DB. 1979. An algorithm for tracking multiple targets. IEEE Trans Automat Contr 24: 843–854.

Sage D, Neumann FR, Hediger F, Gasser SM, Unser M. 2005. Automatic tracking of individual fluorescence particles: Application to the study of chromosome dynamics. IEEE Trans Image Process 14: 1372–1383.[Medline]

Sako Y, Minoguchi S, Yanagida T. 2000. Single-molecule imaging of EGFR signalling on the surface of living cells. Nat Cell Biol 2: 168–172.[Medline]

Sbalzarini IF, Koumoutsakos P. 2005. Feature point tracking and trajectory analysis for video imaging in cell biology. J Struct Biol 151: 182–195.[Medline]

Shafique K, Shah M. 2005. A noniterative greedy algorithm for multiframe point correspondence. IEEE Trans Pattern Anal Mach Intell 27: 51–65.[Medline]

Shroff H, Galbraith CG, Galbraith JA, White H, Gillette J, Olenych S, Davidson MW, Betzig E. 2007. Dual-color superresolution imaging of genetically expressed probes within individual adhesion complexes. Proc Natl Acad Sci 104: 20308–20313.[Abstract/Free Full Text]

Smith C, Eisenstein M. 2005. Automated imaging: Data as far as the eye can see. Nat Methods 2: 547–555.

Starck JL, Murtagh F, Bijaoui A. 2000. Image processing and data analysis: The multiscale approach. Cambridge University Press, Cambridge, UK.

Stelzer EHK. 2000. Practical limits to resolution in fluorescence light microscopy. In Imaging neurons (eds. R Yuste et al.), pp. 12.11–12.19. Cold Spring Harbor Laboratory Press, Cold Spring Harbor, NY.

Swedlow JR, Goldberg I, Brauner E, Sorger PK. 2003. Informatics and quantitative analysis in biological imaging. Science 300: 100–102.[Abstract/Free Full Text]

Thomann D, Rines DR, Sorger PK, Danuser G. 2002. Automatic fluorescent tag detection in 3D with super-resolution: Application to the analysis of chromosome movement. J Microsc 208: 49–64.[Medline]

Tirnauer JS, Salmon ED, Mitchison TJ. 2004. Microtubule plus-end dynamics in Xenopus egg extract spindles. Mol Biol Cell 15: 1776–1784.[Abstract/Free Full Text]

Tvaruskó W, Bentele M, Misteli T, Rudolf R, Kaether C, Spector DL, Gerdes HH, Eils R. 1999. Time-resolved analysis and visualization of dynamic processes in living cells. Proc Natl Acad Sci 96: 7950–7955.[Abstract/Free Full Text]

Vallotton P, Ponti A, Waterman-Storer CM, Salmon ED, Danuser G. 2003. Recovery, visualization, and analysis of actin and tubulin polymer flow in live cells: A fluorescence speckle microscopy study. Biophys J 85: 1289–1306.[Medline]

Veenman CJ, Reinders MJT, Backer E. 2001. Resolving motion correspondence for densely moving points. IEEE Trans Pattern Anal Mach Intell 23: 54–72.

Yildiz A, Selvin PR. 2005. Fluorescence imaging with one nanometer accuracy: Application to molecular motors. Acc Chem Res 38: 574–582.[Medline]

Zenisek D, Steyer JA, Almers W. 2000. Transport, capture, and exocytosis of single synaptic vesicles at active zones. Nature 406: 849–854.[Medline]