“ptcls_bad” set, and then using the left mouse button, box a set
of “negative particles”, which should consist of features in the
tomogram that represent things that are not your FOIs. The
recommended minimum is 100 but more is better, if possible.
Pick as many different aspects of the tomogram as possible
(other cellular components, carbon hole boundaries, gold fiducials, cytoplasm, extracellular space, etc.) that are not the FOIs.
In the main window, the boxes that belong to this particle set
should now be represented as a different color from the
“ptcls_good” set. Be sure the check box next to the
“ptcls_bad” set is the only one ticked and select “save” to
output the particle stack file. The filename suffix that displays
should be “ptcls_bad”.
10. Close the “Main Window” and select the “Build training set”
tab. This step will combine the positive and negative particle
files as well as the hand-painted segmentation/mask for each
positive particle into one larger file. Under “particles_raw”,
browse for the good/positive particles file. For “particles_label”, find the segmented/painted particles file. In “boxes_negative” locate the bad/negative particles file. Select “Launch”.
11. Select “Train the neural network”. Under “trainset” find the
trainset file (see Note 20). Name the output file in the “nettag”
box. Select “Launch”. The resulting training can be viewed by
opening the e2display browser window using this command:
e2display.py. From there, find the trainout_file and open it
using the “Show Stack” option. An example result is shown
in Fig. 5, where the left-most column shows the region of
interest in the tomogram. The middle column shows the
hand-drawn segmentation while the right-hand column is the
AI training result. The training is considered successful if the
middle and right-hand columns closely match one another.
12. In this step, the trained neural network will be used to perform
the final feature segmentation of the tomogram. Select the
“Apply the neural network” tab. Select the imported,
pre-processed tomogram in the “tomograms” box. Find the
trained neural network from the previous step under “nnet”.
Name the output file and click “Launch”. The output file will
show up in the “segmentations” subdirectory within your
project’s directory. The resulting volumes will be the same
dimensions as the input tomogram, and bright pixels represent
regions that have been segmented by the neural network (see
Fig. 6).
3.9 Visualization
of Segmentation
Results
Visualization and analysis of the resulting segmentation volumes
can be performed in your favorite volume rendering software, but
here we describe a relatively simple workflow for visualization using
the commercially available software Amira. UCSF Chimera is a
42
Ryan K. Hylton et al.
of “negative particles”, which should consist of features in the
tomogram that represent things that are not your FOIs. The
recommended minimum is 100 but more is better, if possible.
Pick as many different aspects of the tomogram as possible
(other cellular components, carbon hole boundaries, gold fiducials, cytoplasm, extracellular space, etc.) that are not the FOIs.
In the main window, the boxes that belong to this particle set
should now be represented as a different color from the
“ptcls_good” set. Be sure the check box next to the
“ptcls_bad” set is the only one ticked and select “save” to
output the particle stack file. The filename suffix that displays
should be “ptcls_bad”.
10. Close the “Main Window” and select the “Build training set”
tab. This step will combine the positive and negative particle
files as well as the hand-painted segmentation/mask for each
positive particle into one larger file. Under “particles_raw”,
browse for the good/positive particles file. For “particles_label”, find the segmented/painted particles file. In “boxes_negative” locate the bad/negative particles file. Select “Launch”.
11. Select “Train the neural network”. Under “trainset” find the
trainset file (see Note 20). Name the output file in the “nettag”
box. Select “Launch”. The resulting training can be viewed by
opening the e2display browser window using this command:
e2display.py. From there, find the trainout_file and open it
using the “Show Stack” option. An example result is shown
in Fig. 5, where the left-most column shows the region of
interest in the tomogram. The middle column shows the
hand-drawn segmentation while the right-hand column is the
AI training result. The training is considered successful if the
middle and right-hand columns closely match one another.
12. In this step, the trained neural network will be used to perform
the final feature segmentation of the tomogram. Select the
“Apply the neural network” tab. Select the imported,
pre-processed tomogram in the “tomograms” box. Find the
trained neural network from the previous step under “nnet”.
Name the output file and click “Launch”. The output file will
show up in the “segmentations” subdirectory within your
project’s directory. The resulting volumes will be the same
dimensions as the input tomogram, and bright pixels represent
regions that have been segmented by the neural network (see
Fig. 6).
3.9 Visualization
of Segmentation
Results
Visualization and analysis of the resulting segmentation volumes
can be performed in your favorite volume rendering software, but
here we describe a relatively simple workflow for visualization using
the commercially available software Amira. UCSF Chimera is a
42
Ryan K. Hylton et al.