Reconstruction of Microtubule Centerlines from Electron Tomograms

Abstract

The organization of the mitotic spindle, a structure that separates the chromosomes during cell division, is an active research topic in molecular cell biology. It is composed of microtubules, elongated tubular macromolecules with a diameter of 25 nm. The only volumetric imaging technique that is available to a wide community and provides the required resolution to capture details about microtubules is electron tomography. However, the automatic detection of microtubules in electron tomograms is a difficult task due to the low contrast of the data. Furthermore, thick samples have to be cut into 300 nm thin sections before electron tomography can be applied. Software for automatically segmentation and stitching of the microtubules are not available and therefore these tasks have to be performed manually. Unfortunately, manual segmentation is time consuming for large samples and manual stitching of the tomograms is often infeasible due to the lack of prominent features for registration. Conclusions drawn from electron tomographic data is currently mostly based on either small samples containing few microtubules or single sections of complex structures. Consequently, simple properties, such as the length of microtubules in the spindle or their number, are still unknown for most model organisms. In this thesis, we present methods for 1) an automatic segmentation of microtubule centerlines in electron tomograms, and 2) an automatic stitching of the lines extracted from serial sections. For the centerline segmentation, we use 3D template matching and exploit knowledge about shape of microtubules and microscopy artifacts to design the templates. For the registration of the lines, we present a way to model the orientation of lines as a mixture of Fisher-Mises distributions where we estimate transformation parameters with the expectation maximization algorithm. The final line matching problem is formulated in terms of a probabilistic graphical model. To find the correct correspondences of line ends, we use belief propagation. We handle the poor convergence properties of this algorithm by detecting ambiguous and conflicting assignments of lines automatically. An expert can then influence the final output of the algorithm by solving conflicts manually. A detailed error analysis on true biological data and assessment of the reliability of the results is the prerequisite for analyzing the resulting line representations of the microtubules. To this end, the developed workflow for segmenting and stitching of microtubule centerlines is evaluated on plasticembedded samples of C. elegans early embryos and of spindles from X. laevis egg extracts. Our results suggest that the output of the presented algorithms together with little manual correction is of sufficient quality to allow a detailed analysis of dense microtubule networks. Finally, we exemplarily show results for the centrosome of a C. elegans mitotic spindle.