staining and other sample preparation artifacts that may complicate subsequent
analysis.
Because of the low signal-to-noise ratio and the other factors mentioned above,
reconstructions obtained by electron tomography are difficult to interpret. This
difficulty is further aggravated in highly complex, crowded cellular systems [19].
As a consequence of these factors, image and signal processing methods developed
for other imaging domains are not straightforward to apply to electron tomography
data. While there have been substantial efforts during the last few years which
specifically address segmentation of features from electron tomograms, progress has
been much slower than in related imaging fields. In fact, the relative lack of adequate tools for automatic and objective extraction of information has been recognized as a critical barrier to progress in the field of electron tomography [20–23]
and the task is often carried out manually, using programs that allow tracing within
slices to create iso-contour models of the features of interest [24, 25]. This type of
hand tracing tends to be time consuming, tedious, and subjective. The remainder of
this chapter describes computational approaches specifically targeted to segmentation of electron tomograms with special emphasis on electron cryo-tomography.
12.2 Membrane Segmentation
Membranes tend to be relatively easily identifiable by eye in tomogram sections
perpendicular to the electron beam direction. This includes cell membranes as well
as membrane compartments or vesicles in cells. Many computational segmentation
approaches specifically developed for electron tomography target improvement of
manual segmentation using various types of surface or curve fitting approaches.
These methods include simple spatial gradient optimization in two dimensions [26],
three-dimensional geodesic active contours [27], and full-fledged dynamic level-set
based approaches [28, 29]. All these methods are based on some form of
energy-minimization, thus having tendencies to get trapped in local optima and
being subject to scalability issues. These complications result in the requirement for
reasonably good starting models as well as careful fine-tuning of the algorithm
parameters to ensure correct convergence.
Other edged-based methods that do not rely on manual pre-segmentation include
a bilateral edge-detection algorithm [30], a method based on orientation fields and
line segment detection [31], a dual-contour fast marching method with automated
seed selection [32], and an approach based on differential geometry and use of the
Hessian tensor [33]. The latter was improved at a later stage by including the ability
to classify the detected membrane structures [34]. These membrane detectors do a
good job in enhancing the membrane signal but can suffer a number of drawbacks.
For example, gaps that can appear in membrane delineations due to experimental
imaging conditions may not be properly filled or structures that protrude from the
membrane may be segmented as part of the membrane. Also, because these detectors
are primarily sensitive to line-like features, they tend to be problematic in cases
12 Segmentation of Features in Electron Tomographic Reconstructions
303
analysis.
Because of the low signal-to-noise ratio and the other factors mentioned above,
reconstructions obtained by electron tomography are difficult to interpret. This
difficulty is further aggravated in highly complex, crowded cellular systems [19].
As a consequence of these factors, image and signal processing methods developed
for other imaging domains are not straightforward to apply to electron tomography
data. While there have been substantial efforts during the last few years which
specifically address segmentation of features from electron tomograms, progress has
been much slower than in related imaging fields. In fact, the relative lack of adequate tools for automatic and objective extraction of information has been recognized as a critical barrier to progress in the field of electron tomography [20–23]
and the task is often carried out manually, using programs that allow tracing within
slices to create iso-contour models of the features of interest [24, 25]. This type of
hand tracing tends to be time consuming, tedious, and subjective. The remainder of
this chapter describes computational approaches specifically targeted to segmentation of electron tomograms with special emphasis on electron cryo-tomography.
12.2 Membrane Segmentation
Membranes tend to be relatively easily identifiable by eye in tomogram sections
perpendicular to the electron beam direction. This includes cell membranes as well
as membrane compartments or vesicles in cells. Many computational segmentation
approaches specifically developed for electron tomography target improvement of
manual segmentation using various types of surface or curve fitting approaches.
These methods include simple spatial gradient optimization in two dimensions [26],
three-dimensional geodesic active contours [27], and full-fledged dynamic level-set
based approaches [28, 29]. All these methods are based on some form of
energy-minimization, thus having tendencies to get trapped in local optima and
being subject to scalability issues. These complications result in the requirement for
reasonably good starting models as well as careful fine-tuning of the algorithm
parameters to ensure correct convergence.
Other edged-based methods that do not rely on manual pre-segmentation include
a bilateral edge-detection algorithm [30], a method based on orientation fields and
line segment detection [31], a dual-contour fast marching method with automated
seed selection [32], and an approach based on differential geometry and use of the
Hessian tensor [33]. The latter was improved at a later stage by including the ability
to classify the detected membrane structures [34]. These membrane detectors do a
good job in enhancing the membrane signal but can suffer a number of drawbacks.
For example, gaps that can appear in membrane delineations due to experimental
imaging conditions may not be properly filled or structures that protrude from the
membrane may be segmented as part of the membrane. Also, because these detectors
are primarily sensitive to line-like features, they tend to be problematic in cases
12 Segmentation of Features in Electron Tomographic Reconstructions
303