1.10.1 Segmentation and denoising
Segmentation can be defined as the process by which electron-dense features in a
tomogram are ascribed identities and highlighted relative to other features (Fig. 1.6;
Chap. 12; [18]). This decomposition into structural components, e.g. membranes or
filaments is essential to understanding function [35]. Early examples of segmentation were actual models made with, for example, ping pong balls, corks [102],
balsa wood [103], and even clay [104]. Given that segmentation still tends to be
time-consuming, even for a specialist, the reasons for doing so are to give clarity to
the features of interest (via the use of colour and shading but also by purposefully
omitting distracting features), and to present the data in a way that does justice to
3D spatial representation. Here, the benefits of tomography are finally realised.
Judicious use of colour, lighting and perspective provides an intuitive representation of features in real space. There is no convention for colour-coding of features,
and colours can be chosen to emphasise distinct structures and even their functional
relationships. Segmentation of a 2D micrograph is certainly possible but it makes
assumptions about whether features are interconnected or merely in close proximity. 2D snapshots from any plane of the reconstruction can reveal details that
were not apparent from the original 2D perspective.
Computerised segmentation can be an entirely objective approach [105, 106],
not to mention the least tedious. Semiautomated (semi-objective) approaches can
also be highly efficient [84]. Nonetheless, tomograms are often segmented manually
because the available segmentation algorithms are often inferior to human anticipation, which can infer incomplete and/or anisotropic data. Examples of features
that can be delineated without a priori information (such as an electron-dense label)
include membranes, components of the cytoskeleton, and distinctive molecular
complexes such as the 26S proteasome [38]. In this case, structural signature alone
is sufficient for reliable identification [107].
Stained plastic sections are dominated by amplitude contrast and a correspondingly high signal-to-noise ratio. In the simplest case, manual thresholding of
intensities and removal of spurious (outlying) features may yield an acceptable
result for less complex tomograms. The ‘Watershed’ algorithm and other useful
segmentation methods are discussed in Chap. 12. Projections of unstained cryo
specimens are dominated by (weaker) phase contrast and noise. Bilateral filtering
and nonlinear anisotropic diffusion (NAD) [108] are examples of real-space, nonlinear denoising algorithms that are relatively stable with respect to noise; however,
they remove some signal component selectively, resulting in data with higher SNR
but suboptimal preservation of information (see Chap. 8). Therefore, denoising is
suitable as an aid to segmentation but not for further quantitative analysis or
subtomogram averaging.
1 Electron Tomography: A Primer
19
Segmentation can be defined as the process by which electron-dense features in a
tomogram are ascribed identities and highlighted relative to other features (Fig. 1.6;
Chap. 12; [18]). This decomposition into structural components, e.g. membranes or
filaments is essential to understanding function [35]. Early examples of segmentation were actual models made with, for example, ping pong balls, corks [102],
balsa wood [103], and even clay [104]. Given that segmentation still tends to be
time-consuming, even for a specialist, the reasons for doing so are to give clarity to
the features of interest (via the use of colour and shading but also by purposefully
omitting distracting features), and to present the data in a way that does justice to
3D spatial representation. Here, the benefits of tomography are finally realised.
Judicious use of colour, lighting and perspective provides an intuitive representation of features in real space. There is no convention for colour-coding of features,
and colours can be chosen to emphasise distinct structures and even their functional
relationships. Segmentation of a 2D micrograph is certainly possible but it makes
assumptions about whether features are interconnected or merely in close proximity. 2D snapshots from any plane of the reconstruction can reveal details that
were not apparent from the original 2D perspective.
Computerised segmentation can be an entirely objective approach [105, 106],
not to mention the least tedious. Semiautomated (semi-objective) approaches can
also be highly efficient [84]. Nonetheless, tomograms are often segmented manually
because the available segmentation algorithms are often inferior to human anticipation, which can infer incomplete and/or anisotropic data. Examples of features
that can be delineated without a priori information (such as an electron-dense label)
include membranes, components of the cytoskeleton, and distinctive molecular
complexes such as the 26S proteasome [38]. In this case, structural signature alone
is sufficient for reliable identification [107].
Stained plastic sections are dominated by amplitude contrast and a correspondingly high signal-to-noise ratio. In the simplest case, manual thresholding of
intensities and removal of spurious (outlying) features may yield an acceptable
result for less complex tomograms. The ‘Watershed’ algorithm and other useful
segmentation methods are discussed in Chap. 12. Projections of unstained cryo
specimens are dominated by (weaker) phase contrast and noise. Bilateral filtering
and nonlinear anisotropic diffusion (NAD) [108] are examples of real-space, nonlinear denoising algorithms that are relatively stable with respect to noise; however,
they remove some signal component selectively, resulting in data with higher SNR
but suboptimal preservation of information (see Chap. 8). Therefore, denoising is
suitable as an aid to segmentation but not for further quantitative analysis or
subtomogram averaging.
1 Electron Tomography: A Primer
19