Automated electron microscope tomography using robust prediction of specimen movements

Abstract

A new method was developed to acquire images automatically at a series of specimen tilts, as required for tomographic reconstruction. The method uses changes in specimen position at previous tilt angles to predict the position at the current tilt angle. Actual measurement of the position or focus is skipped if the statistical error of the prediction is low enough. This method allows a tilt series to be acquired rapidly when conditions are good but falls back toward the traditional approach of taking focusing and tracking images when necessary. The method has been implemented in a program, SerialEM, that provides an efficient environment for data acquisition. This program includes control of an energy filter as well as a low-dose imaging mode, in which tracking and focusing occur away from the area of interest. The program can automatically acquire a montage of overlapping frames, allowing tomography of areas larger than the field of the CCD camera. It also includes tools for navigating between specimen positions and finding regions of interest.

Introduction

Electron tomography has become increasingly important as a technique for examining cellular structures and macromolecules in three dimensions (Baumeister, 2002, Leapman, 2004, McIntosh et al., 2005, Steven and Aebi, 2003). The method involves taking a series of electron micrographs while tilting the specimen over a range of angles, typically at 1–2° intervals up to ±60–70°. These projections of the specimen are aligned and a three-dimensional reconstruction is computed by methods such as weighted backprojection (Frank, 1992). The recent growth of tomography has been fostered by continual improvements in the instrumentation, computers, and software needed for these tasks. In particular, much effort has been expended to automate the acquisition of the series of tilted projections, a task that has been facilitated by improved interfaces for microscope control. Acquisition requires a sequence of operations because of imperfections in the goniometers and specimen holders now available: after tilting to a new angle, the features of interest must be recentered in the field of view, and any change in vertical (Z) height must be compensated by refocusing, before the final image is acquired. When done manually, these operations are slow and tedious at best, and because the sample is exposed to the beam during these steps, they are particularly difficult for beam-sensitive samples. To automate this process, an initial tracking image is taken on a CCD camera after tilting to a new angle; the image is cross-correlated with a comparable image from the previous tilt, and electronic image shift is used to recenter the specimen. Defocus is determined by taking two images with the beam tilted in opposite directions; the displacement between two such images is proportional to the defocus, so the objective lens can then be adjusted to bring the specimen back to the desired focus (Koster and de Ruijter, 1992, Koster et al., 1987).

The first generation of automated tomography software took one or two tracking images at each tilt, and also measured defocus once or twice on each tilt. Some software followed a strict low-dose approach by taking tracking and focusing pictures on areas displaced from the region of interest, which is essential for working with frozen-hydrated material (Dierksen et al., 1992, Dierksen et al., 1993, Rath et al., 1997). Other software was more suited to work with plastic sections, but still provided a huge savings in dose to the specimen and an increase in throughput, quality, and convenience (Fung et al., 1996, Koster et al., 1992, Koster et al., 1993).

With the availability of better goniometers, new approaches have been devised with the goal of dispensing with some of the tracking and focusing steps by taking advantage of the predictable properties of the stage. The method of precalibration involves measuring stage movements in X, Y, and Z through a coarse tilt series (e.g., at 5° intervals), possibly at a lower magnification, and then applying these movements to acquire one or more full tilt series at a finer interval (Ziese et al., 2002). Precalibration is the basis for the tomography package supplied by the FEI Company (Eindhoven, Netherlands). Most recently, a prediction method has been developed (Zheng et al., 2004) which assumes that each point on the specimen moves in a circle around the tilt axis as the stage is tilted and predicts changes in Z-height from recent changes in the lateral position of the specimen. This method, which will be referred to here as Z-prediction, requires essentially no tracking or focusing images during the tilt series and represents an elegant solution to tilt series acquisition in situations where the goniometer and specimen holder satisfy its assumptions.

This paper describes a different approach to skipping steps and speeding up data acquisition. Specimen position is predicted in X, Y, and Z based only on the movements during recent tilts. The prediction is relied upon, i.e., tracking or focusing is skipped, only when the statistical error of the prediction is low enough. Rather than a prediction based upon a previous trajectory, as in precalibration, or a prediction of one coordinate based upon geometrical considerations, this method uses a conservative and adaptive prediction from movements in the current tilt series. The program described here, SerialEM, was developed after experimenting with the precalibration approach and becoming concerned about its vulnerability to less than ideal circumstances: when there is specimen drift or nonreproducible holder movements, when the eucentric height is not set equivalently between a precalibration run and the actual tilt series, or when it is necessary to readjust specimen position during the series. It will be shown here that the program succeeds in achieving the speed of precalibration where conditions are favorable, while providing the reliability of the traditional approach in less ideal cases.

This paper is organized as follows. Various components and features of the program needed for successful automation will be described first. The prediction algorithm and its implementation by a module called the “tilt series controller” are then explained. Examples are given of the behavior of the program in typical and extreme circumstances. The performance of this method relative to the precalibration and Z-prediction methods is evaluated, and data are presented indicating that the Z-prediction method is particularly susceptible to nonideal stage performance. Finally, additional features of the program are described, such as its ability to work with an energy filter and to take montaged images from overlapping CCD images. One motivation for describing some aspects of the program in detail is to highlight the various nonideal properties of current microscopes and cameras that keep automated tomography from being straightforward, and to present ways in which these problems can be overcome.

This work was reported previously in abstract form (Mastronarde, 2003).