Fig 4. Using advanced tools for 3D cell structure analysis to help visualize and verify connections between Golgi cisternae in glucose-stimulated beta cells. A potential "disadvantage" of our preparative methods (i.e., high-pressure freezing/freeze substitution) comes from the fact that the cytoplasm appears very dense in comparison with cells prepared by classical methods, in which the Golgi and other subcellular membranes are more readily distinguished. This increased cytosolic density makes the process of segmentation (i.e., identifying and modeling biological features of interest within the 3D data) a time-consuming task that necessitates an experienced eye. Typically, connections of interest between nonequivalent cisternae do not lie exactly within the x–y plane. (A) Model contours of two medial-cisternae (green and red) in the Golgi stack are shown in a normal x–y view of a pixel-thick slice extracted from the 3D reconstruction of the Golgi region presented in Fig. 3. (B) The normal x–y view presented in A zoomed by a factor of 3, with model contours omitted to allow the membranes to be more clearly visualized. By using the Slicer tool that is part of 3DMOD within the IMOD software package (1), we rotated the data in the x, y, and z axes to unambiguously visualize the connections both without (C) and with (D) model contours drawn (x,y,z rotations of 11.0°, 5.0° and –15°, respectively). The image shown in D represents the sum of three individual tomographic slices. Model contours from five successive tomographic slices are drawn, because model contours are now oblique to the x–y plane. To visualize Golgi membranes and follow connections between cisternae at different levels more easily and to take preliminary steps toward the use of fast and accurate automated methods for detecting and defining the boundaries of Golgi membranes in 3D, the tomographic data presented in A–D were preprocessed with an iterative 3D median filtering algorithm (E). This type of filter efficiently reduces noise while preserving the location of edges (2). In the example shown here, six iterations of median filtering were used. We then applied a 3D variant of the watershed transform to segment the continuum between the cisternae at different levels shown in E in a semiautomated manner (3). (F) The resulting segmentation mask was then applied to the original, unfiltered tomographic data.

1. Kremer, J. R., Mastronarde, D. N. & McIntosh, J. R. (1996) J. Struct. Biol. 116, 71–76.

2. Yang, G. J. & Huang, T. S. (1981) Comput. Graphics Image Proc. 15, 224-245.

3. Volkmann, N. (2002) J. Struct. Biol. 138, 123-129.