Automated segmentation of electron tomograms for a quantitative description of actin filament networks

Abstract

Cryo-electron tomography allows to visualize individual actin filaments and to describe the three-dimensional organization of actin networks in the context of unperturbed cellular environments. For a quantitative characterization of actin filament networks, the tomograms must be segmented in a reproducible manner. Here, we describe an automated procedure for the segmentation of actin filaments, which combines template matching with a new tracing algorithm. The result is a set of lines, each one representing the central line of a filament. As demonstrated with cryo-tomograms of cellular actin networks, these line sets can be used to characterize filament networks in terms of filament length, orientation, density, stiffness (persistence length), or the occurrence of branching points.

Introduction

Electron tomography provides, in a digital format, information about the three-dimensional density distribution within the object under study or, more precisely, its locally varying electron scattering power. For perception of these data by the human visual system, it is necessary to convert them into images either in the form of a series of two-dimensional (2D) images (‘slices’) or three-dimensional (3D) images generated by surface or volume rendering. For some purposes this form of visualization is sufficient. However, for a deeper analysis of structural features in the complex settings of cellular landscapes, it is necessary to characterize them in terms of parameters such as surface areas, volumes, distances or angles. A quantitative analysis of this kind makes it necessary to identify and localize the features of interest and to separate them from the background. In other words, the tomograms must be segmented.

Here we describe a novel approach for the automated segmentation of filamentous structures, in particular actin filaments. Currently, such segmentations are mostly performed manually under visual control by drawing contour lines, sometimes supported by drawing programs (Kremer et al., 1996, Stalling et al., 2005). Manual segmentation, however, is slow and tedious and not as reproducible as a thorough quantitative or statistical analysis might require. Although the expertise of the person executing this task may be a valuable asset, there is an element of subjectivity that might introduce some bias. Moreover, some features are easily overlooked since they must be segmented in 2D-slices; for example an actin filament running perpendicular to the plane of a tomographic slice is hard to recognize. Even though the results of an automated segmentation might not be perfect when compared with the ground truth – which is usually unknown – it is reproducible and allows a meaningful comparison of results from a multitude of experiments. Obviously, the results of an automated segmentation depend on user-defined parameter settings, but if the segmentation process is fast, many parameter settings can be explored and an optimal setting can be chosen, which is then applied throughout a set of experiments.

The poor signal-to-noise ratio (SNR) of electron tomograms recorded under low-dose conditions and the artifacts arising from incomplete data (‘missing wedge’) make segmentation a challenging task. Existing segmentation methods as used widely in biomedical imaging cannot cope with data of this kind (Al-Kofahi et al., 2002, Huang and Stockman, 1993, Kirbas and Quek, 2004). We have shown previously that template matching allows to identify known macromolecular structures in tomograms of frozen-hydrated phantom or real cells (Böhm et al., 2000, Frangakis et al., 2002, Ortiz et al., 2006). Subvolumes containing structures similar to the template can then be extracted from the tomograms, aligned and classified, and finally averaged. Template matching is usually preceded by a ‘denoising’ step. Here, we use denoising followed by template matching as a first step in the segmentation procedure. A generic cylinder, whose length and diameter has to be adapted to the imaging parameters, is used as a template. The cross-correlation involved in the template matching step provides not only a map of cross-correlation coefficients as a measure of the similarity between the template and the local density in the tomogram, but also a map representing the best local orientation of the cylindrical template. These two maps are then used to calculate a similarity function that allows to evaluate the likelihood that two neighboring voxels are connected by a filament. The second step of the segmentation process is a new tracing algorithm scanning all voxels for possible connection on the basis of the similarity function. This results in an array of filaments each represented by its central line.

The concept of using an orientation field for the segmentation of thin structures has been reported by Sandberg and Brega (2007). These authors applied a type of local Hough transform in order to derive the orientation field, in contrast to our approach, which is based on template matching. Mainly designed for the segmentation of surfaces, the former approach works only in 2D, making it necessary to process tomography data slice by slice. The resulting slices are still grey level images and need further processing for a geometric representation of surfaces.

We have used our automated segmentation approach to characterize actin networks in cellular tomograms. Actin polymerization and depolymerization underlie cytoplasmic organization and cell motility (see e.g. Pollard and Borisy, 2003, Ridley et al., 2003). Actin filament networks with different characteristics exist in distinct cellular territories as shown schematically in Fig. 1 and there is a need to describe them in quantitative terms. We have used the results of automated segmentations to describe properties of individual filaments (length, stiffness) and of actin networks (density, mutual orientations). We have also addressed the issue of actin branching, as mediated by the Arp2/3 complex (Pollard and Beltzner, 2002). The very existence of such actin branching has recently become a matter of some controversy (Higgs, 2011, Insall, 2011, Urban et al., 2010). Clearly, the detection of genuine branchings and their characterization cannot be based on visual inspection of tomograms or on manual segmentation. Large data sets must be analyzed and branching must be characterized by parameters based on statistics – a task calling for automation.