Structure, Vol. 12, 1763–1774, October, 2004, 2004 Elsevier Ltd. All rights reserved. DOI 10.1016/j.str.2004.07.022 Ways & Means Methods for Generating High-Resolution Structural Models from Electron Microscope Tomography Data 1996; Li et al., 1997; Marko and Leith, 1996; Perkins et al., 1997). Although this method reveals the general shape and organization of logically distinct cellular com- ponents, the tedious and subjective nature of manual tracing makes it difficult to generate models that reliably David B. Ress,* Mark L. Harlow, Robert M. Marshall, and Uel J. McMahan Department of Neurobiology Stanford University School of Medicine Stanford, California 94305 represent the components at the full spatial resolution of the reconstructed volume. More sophisticated model-generation methods are Summary also available for EMT. One EMT software package, IMOD (Kremer et al., 1996), permits the creation of a Reconstructed volumes generated by tilt-image elec- model contour at a particular isodensity level on a series tron-microscope tomography offer the best spatial of virtual slices. General purpose commercial software resolution currently available for studying cell struc- applications such as Amira (ZIB, Indeed-Visual Con- tures in situ. Analysis is often accomplished by creat- cepts GmbH, Berlin, Germany) and AVS (Advanced Vi- ing surface models that delineate grayscale contrast sual Systems Inc., Waltham, MA) offer a plethora of boundaries. Here, we introduce a specialized and con- alternative segmentation, visualization, and analysis venient sequence of segmentation operations for methods. While these software packages each have making such models that greatly improves their relia- their advantages and limitations, there remains a sub- bility and spatial resolution as compared to current stantial need for additional approaches that offer im- approaches, providing a basis for making accurate provements in speed, convenience, and accuracy. measurements. To assess the reliability of the surface Here, we present a set of 3D image processing ways- models, we introduce a spatial uncertainty measure- and-means, specifically designed for EMT data vol- ment based on grayscale gradient scale length. The umes, which provide convenient and reliable generation model generation and measurement methods are vali- of full-resolution surface models. Independently control- dated by applying them to synthetic data, and their lable, 3D surface models of distinct structural compo- utility is demonstrated by using them to characterize nents are created using a two-step, dual-resolution iso- macromolecular architecture of active zone material density volume-of-interest (IVOI) approach. The first at the frog’s neuromuscular junction. step is a low-resolution slice-by-slice segmentation that produces small volumes-of-interest (VOIs) that enclose, Introduction but do not precisely delineate, individual components. The second step forms an isodensity surface as a full- Tilt-image electron-microscope tomography (EMT) can resolution model of the structural component enclosed be used to generate 3D reconstructions of sections from by each VOI. Because the initial IVOI segmentation step stained, plastic-embedded biological tissue samples is error tolerant, it can be performed manually, or by (Frank, 1992; Harlow et al., 2001; He et al., 2003; Horo- using interactive, semiautomatic methods. To assist witz et al., 1994; Ladinsky et al., 1999; Lenzi et al., 1999; both manual and automatic segmentation, we describe Martone et al., 1999; McEwen et al., 1986, 1993; Sedzik a parametric-spline path generation method. We also et al., 1992; Taylor et al., 1999; Woodcock et al., 1991). introduce an active-contour method designed to auto- The volume reconstructions make it possible to study matically segment the membranous structures that are the structure of cellular components within the depth of common in tissues imaged using EMT. Finally, we intro- the sections, acquiring information about sizes, shapes, duce a measurement, spatial uncertainty, which quanti- and relationships that cannot be obtained in any other fies the spatial reliability of a surface model based on way. However, analysis is often hindered by the abun- the local grayscale noise and gradient scale length. dance of structures, the complexity of their shapes and Our methods were validated using a synthetic volume relationships, staining inhomogeneity, and noise. To al- that simulated a typical EMT reconstructed tissue sec- leviate these problems, it is useful to form 3D models of tion. The simulated volume contained cylinders, spher- the components-of-interest within the reconstructions. oids, and a folded sheet which represent the most Surface models, which represent the boundary between common shapes of cellular components: for example components and the space surrounding them, are com- filaments, vesicles, and a plasma membrane, respec- monly used. tively. Surface texture was added to these simulated Because of the complexity and noise, most research structures to assess the accuracy of the model genera- in EMT has made use of manual surface-model genera- tion and measurement methods. Various amounts of tion methods. Typically, virtual slices are formed through noise were also added to the volume to assess the a reconstruction and the boundary of each component- performance of the segmentation methods for a range of-interest is traced in the slices in which it appears. of experimental data quality, and to quantify the accu- The traces are then interpolated to create a 3D surface racy of the spatial uncertainty measurement. We also (Frank et al., 1996; Hessler et al., 1992; Kremer et al., show how our methods were used to generate and ana- lyze surface models from EMT reconstructions of tissue sections. The sections were from fixed, stained, and *Correspondence: ress@stanford.edu; grantser@stanford.edu