15. As long as they are tracking well, select a high number of
fiducials to track in the tomogram (upwards of 20). If there
are fewer than 10 points available, the tracking will still be
adequate if they are distributed evenly across the field of view
or around the object of interest. If there are fewer than 3–5
well-distributed fiducials, using the “patch tracking” method
to generate a fiducial model may be the best option. In this
situation, there must be good contrast for the algorithm to
align to.
16. Not every fiducial must be tracked completely through the tiltseries. If some do not track completely or go outside of the field
of view, it is ok, as long as there are multiple fiducials that do
track across the entire tilt-range or multiple that overlap in
their tracking across the whole tilt-series.
17. Until you are sure you are done processing a tomogram
completely, do not delete intermediate files. It may be that
you will want to come back and reconstruct at 1 Â binning
instead of 2 Â binning, remove gold, or perform CTF correction. To do this, navigate to the directory the reconstruction
files are in using the command terminal and run the command:
etomo <filename.edf>. This will launch your reconstruction
workflow just as you left off, allowing alterations to be made
easily.
18. This segmentation feature of EMAN2 is likely to upgrade
quickly, so it is recommended that these instructions be
checked against the newest documentation on the EMAN2
website if they are not working (https://blake.bcm.edu/
emanwiki/EMAN2/Programs/tomoseg). If your AI segmentations are very noisy, it may be worth exploring different
frequency bands for training. Also, unlike the workflow in the
original publication of this approach [7], our workflow treats
all membranes as the same, and uses Amira to separate them
into different materials later. We find that the fidelity of the
segmentation is higher if the neural networks only have to
distinguish between membranes and non-membranes, versus
distinguishing between the membranes of different organelles.
19. In this workflow, pen size 2 was used for membranes and pen
size 3–4 for microtubules and their lumen, but this number
will vary depending on the pixel size in the tomogram.
20. The option “from_trained” will be used if you are refining a
training set against a previously trained network. While not
included in this protocol, it can improve the accuracy of segmentations in some cases.
Cryotomography of Neurons
47
fiducials to track in the tomogram (upwards of 20). If there
are fewer than 10 points available, the tracking will still be
adequate if they are distributed evenly across the field of view
or around the object of interest. If there are fewer than 3–5
well-distributed fiducials, using the “patch tracking” method
to generate a fiducial model may be the best option. In this
situation, there must be good contrast for the algorithm to
align to.
16. Not every fiducial must be tracked completely through the tiltseries. If some do not track completely or go outside of the field
of view, it is ok, as long as there are multiple fiducials that do
track across the entire tilt-range or multiple that overlap in
their tracking across the whole tilt-series.
17. Until you are sure you are done processing a tomogram
completely, do not delete intermediate files. It may be that
you will want to come back and reconstruct at 1 Â binning
instead of 2 Â binning, remove gold, or perform CTF correction. To do this, navigate to the directory the reconstruction
files are in using the command terminal and run the command:
etomo <filename.edf>. This will launch your reconstruction
workflow just as you left off, allowing alterations to be made
easily.
18. This segmentation feature of EMAN2 is likely to upgrade
quickly, so it is recommended that these instructions be
checked against the newest documentation on the EMAN2
website if they are not working (https://blake.bcm.edu/
emanwiki/EMAN2/Programs/tomoseg). If your AI segmentations are very noisy, it may be worth exploring different
frequency bands for training. Also, unlike the workflow in the
original publication of this approach [7], our workflow treats
all membranes as the same, and uses Amira to separate them
into different materials later. We find that the fidelity of the
segmentation is higher if the neural networks only have to
distinguish between membranes and non-membranes, versus
distinguishing between the membranes of different organelles.
19. In this workflow, pen size 2 was used for membranes and pen
size 3–4 for microtubules and their lumen, but this number
will vary depending on the pixel size in the tomogram.
20. The option “from_trained” will be used if you are refining a
training set against a previously trained network. While not
included in this protocol, it can improve the accuracy of segmentations in some cases.
Cryotomography of Neurons
47