Data Analysis for Three-dimensional Volume Scanning Electron Microscopy

Published January 25, 2017

In recent years, three-dimensional (3D) scanning electron microscopy techniques have gained recognition in the biological sciences. In particular, array tomography, serial block face scanning electron microscopy (SBFSEM) and focused ion beam scanning electron microscopy (FIBSEM) (described in Three-Dimensional Scanning Electron Microscopy for Biology) have shown an increase in biological applications, elucidating ultrastructural details of cells and tissues across a variety of different sample types.

In standard electron microscopy, such as brain cancer electron microscopy, sample preparation and data analysis takes more time than sample imaging. The time required for data analysis increases exponentially when you use 3D EM techniques. Thus, the two main challenges in 3D analysis are having enough research time and computer power to process the datasets.

Both necessary time and computer power vary depending on the type of information required from the data and the software used.^1–2To accommodate the different needs, there are a variety of techniques and software packages available at each stage of the 3D SEM data processing and analysis process.^2–3 The stages, with a brief outline of the available options, are below:

Alignment and Stacking

Array tomography, SBFSEM and FIBSEM produce data in the format of a series of sequential images. Each image is captured from the surface of a sample. For array tomography, the sample is a physical slice of the specimen while the sample for SBFSEM and FIBSEM is the sample itself after a slice is removed (see Three-dimensional Scanning Electron Microscopy for Biology for details). You need to stack and calibrate the resultant images (in the header information of the data stack) to correspond with slice thickness and generate correct z-axis and voxel (3D pixel) dimensions (Fig. 1).

Data Analysis for Three-dimensional Volume Scanning Electron Microscopy

Figure 1 shows the process for alignment and stacking of SBFSEM images of liver tissue. First, adjacent images from the data are aligned to one another using either an overlay (top left) or red/green tints to the images (Middle top). Arrows indicate obvious regions in the image where structures are misaligned. Once aligned, the stack of images requires correction to the z-axis dimension (bottom middle and right).

In addition to z-axis stacking, images often require alignment to each other (registration) because of drift between slices or, in the case of array tomography, because sections on the substrate are in slightly different orientations. In other instances, you may have to remove individual from the data stack to avoid skewing the rest of the data (for example, if debris has fallen across one of the slices, obscuring the sample and creating an extreme brightness and contrast range).

Filtering and Binning

3D SEM techniques commonly experience noise and reduced contrast. Therefore, filtering is important to optimize the data output. There are several types of filtering and processing steps you can apply to the data stack. The processing steps you use depend on the original image data and the type of information you seek. Common data processing steps include:

normalizing the contrast and brightness across all images
noise reduction
inverting data and
data binning

In addition to reducing noise, data binning also decreases the size of the data, making subsequent processing steps faster (Fig. 2).

Figure 2 shows the effects of filtering on SBFSEM data of liver tissue. The Gaussian filter reduces noise. Changes in contrast and brightness, which can alter during data collection, are eliminated by normalizing the contrast and brightness across all slices. Inversion of data makes segmentation and data display easier.

3D Data

Most data processing software display a 3D stack of images. You can use original image files, if you have the slice thickness. However, more often you will use images derived from an assembled 3D data file, which has been filtered and processed as outlined above. This is particularly true for very large datasets; they often cannot be displayed until some level of binning is applied.

At this stage, you can conduct a basic analysis of ultrastructure and quantification of repeating features. Often, however, the complex spatial arrangement of structure will require modelling. You can also trim the data once the region of interest has been identified, to make it easier to handle and process.

Segmentation

Segmentation aids in interpretation of structure and enables quantification. Consequently, researchers frequently apply this technique to 3D SEM data is (Fig. 3). In this process, data is subdivided based on ultrastructural characteristics and reconstructed as a 3D model. Segmentation is either fully automated, manual or a combination of the two.³ The segmentation process takes the greatest amount of time.¹ Detailed reconstructions, even with some automated segmentation, can take months and many people to achieve significant results.^4–5

Figure 3 shows different types of data display using the same liver sample shown above. An automated display, called maximum intensity projection, is based on threshold values and, highlights the brightest parts of the data (inverted data is required for this) as a 3D model (left). This is extremely useful for rapid analysis of data. Alternatively, regions or objects of interest can be selected manually on each slice of data based on ultrastructural details (middle). Once selected, this can be viewed as a 3D model (right, nucleus is in blue, mitochondria in green and a red blood cell in red), which can be analyzed using quantitative analysis tools.

Quantification

One of the most exciting aspects of 3D SEM is the potential to obtain significant replicates for quantification. The software and extent of data processing determines the type of quantification you can achieve. The greatest range is obtained with segmented data. For example, you can obtain volume and surface areas of segmented volumes as well as make manual measurements for length and distance. More complex measurements include feret diameters, sphericity, eigenvalues and eigenvectors, distribution of structure and distances of objects from specified points. You can do this across multiple cells, organelles or tissues to obtain statistically relevant information.

Display and Data Output

3D data also lends itself to more dynamic forms of presentation, along with more traditional images of modeled data and quantification results.

Video 1 shows sequential SBFSEM images of liver tissue.

For example, you can make simple videos that scan through original images (Video 1), rotate segmented or reconstructed models (Video 2) or create stereoscopic images on posters, virtual reality visualization and 3D printed models.⁵

Video 2 shows rotation of the model in figure 3.

Specific file formats vary, but you can usually find a range of options that are engaging and informative for presenting results to peers and the general public.