approach is that it enables easy imposition of real-space constraints such as
penalizing solutions that have large intensity variations along the centerline, which
can significantly improve performance [70].
12.5 Outlook and Conclusions
Many promising computational approaches for segmentation have become available
in recent years and it is likely that more such algorithms, taking advantage of the
recent hardware developments [14, 65], will emerge in the near future. However,
many of these algorithms are relatively complex to use, are only applicable to
subsets of electron tomographic reconstructions, often need tweaking of parameters
to yield optimal performance, and tend to require careful examination and adjustments by expert users. As a consequence, accurate segmentation still relies heavily
on human intervention.
One possibility to reduce workload and increase throughput is the use of
crowdsourcing analogous to what was done in the serial block face scanning
electron microscopy community. An example is the EyeWire project [98]. Anyone
can register on the EyeWire website and—after a brief training session—can segment neurons in three-dimensional serial block face scanning electron microscopy
reconstructions through the interface of an online game. Currently an impressive
number of over 200,000 individuals from 145 countries are part of the project [99].
While the EyeWire project relies on people volunteering their time, the use of
commercial services like the Amazon Turk can also be a viable solution [100].
However, block face scanning electron microscopy reconstructions have very high
contrast if compared to biological electron (cryo-)tomography data. Even in stained
biological electron tomograms, variations between expert-user segmentation results
are not negligible [43, 48]. To generate an interface that is amenable for crowdsourced non-expert segmentation with tolerable error rate is challenging.
Three-dimensional landscapes at the nanometer resolution of electron (cryo-)
tomograms are unfamiliar to the human eye and even in the absence of noise they
would likely appear quite chaotic owing to the dense crowding of macromolecules
in cells [19]. Another complication is that humans are susceptible to all sorts of
biases when evaluating visual cues, especially in the presence of high noise levels
[101–103]. A striking example is an experiment where participants were presented
with images containing random noise and were made to believe that 50% of those
images contain faces [104]. Not only did the participants believe to see faces in over
30% of the images containing nothing but random noise, fMRI measurements
showed activity in regions of the brain that are known to be associated with face
processing whenever a participant believed to see a face. While individual subjectivity could possibly be averaged out in segmentation tasks by averaging over
many individuals, it is doubtful that these species-dependent biases can be adequately addressed by averaging. Whether crowdsourcing is used or not, some
12 Segmentation of Features in Electron Tomographic Reconstructions
311
penalizing solutions that have large intensity variations along the centerline, which
can significantly improve performance [70].
12.5 Outlook and Conclusions
Many promising computational approaches for segmentation have become available
in recent years and it is likely that more such algorithms, taking advantage of the
recent hardware developments [14, 65], will emerge in the near future. However,
many of these algorithms are relatively complex to use, are only applicable to
subsets of electron tomographic reconstructions, often need tweaking of parameters
to yield optimal performance, and tend to require careful examination and adjustments by expert users. As a consequence, accurate segmentation still relies heavily
on human intervention.
One possibility to reduce workload and increase throughput is the use of
crowdsourcing analogous to what was done in the serial block face scanning
electron microscopy community. An example is the EyeWire project [98]. Anyone
can register on the EyeWire website and—after a brief training session—can segment neurons in three-dimensional serial block face scanning electron microscopy
reconstructions through the interface of an online game. Currently an impressive
number of over 200,000 individuals from 145 countries are part of the project [99].
While the EyeWire project relies on people volunteering their time, the use of
commercial services like the Amazon Turk can also be a viable solution [100].
However, block face scanning electron microscopy reconstructions have very high
contrast if compared to biological electron (cryo-)tomography data. Even in stained
biological electron tomograms, variations between expert-user segmentation results
are not negligible [43, 48]. To generate an interface that is amenable for crowdsourced non-expert segmentation with tolerable error rate is challenging.
Three-dimensional landscapes at the nanometer resolution of electron (cryo-)
tomograms are unfamiliar to the human eye and even in the absence of noise they
would likely appear quite chaotic owing to the dense crowding of macromolecules
in cells [19]. Another complication is that humans are susceptible to all sorts of
biases when evaluating visual cues, especially in the presence of high noise levels
[101–103]. A striking example is an experiment where participants were presented
with images containing random noise and were made to believe that 50% of those
images contain faces [104]. Not only did the participants believe to see faces in over
30% of the images containing nothing but random noise, fMRI measurements
showed activity in regions of the brain that are known to be associated with face
processing whenever a participant believed to see a face. While individual subjectivity could possibly be averaged out in segmentation tasks by averaging over
many individuals, it is doubtful that these species-dependent biases can be adequately addressed by averaging. Whether crowdsourcing is used or not, some
12 Segmentation of Features in Electron Tomographic Reconstructions
311