Tomogram Annotation/Segmentation

Availability: EMAN2.2+, or daily build after 2016-10 This is the tutorial for a Convolutional neural network(CNN) based semi-automated cellular tomogram annotation protocol.

If you are using this in a publication, please cite:

A stable (thus not up-to-date) online version of the tutorial can also be found at Protocol exchange: https://www.nature.com/protocolexchange/protocols/6171

Updates:

Installation

CNN based tomogram annotation makes use of the TensorFlow deep learning toolkit. At the time of this writing, tensorflow effectively requires CUDA, the compute framework on NVidia GPUs. While future support for OpenCL (an open alternative to CUDA) seems likely, and while it is possible to use TensorFlow on a CPU, at present we cannot recommend running this tool on Mac/Windows. You may do so, but training the networks may be VERY slow.

To make use of accelerated TensorFlow, your machine must be properly configured for CUDA. This is outside the installation of EMAN2, and requires specific drivers, etc. on your machine. Some extra information on how to use CUDA and GPU can be found at http://blake.bcm.edu/emanwiki/EMAN2/Install/BinaryInstallAnaconda#Using_the_GPU

Setup workspace

If you used EMAN2 for your tomogram alignment/reconstruction (EMAN2/e2TomoSmall), simply 'cd' to the project folder and launch e2projectmanager.py. If you will be using tomograms produced by other software:

Make an empty folder and cd into that folder at the command-line. Run e2projectmanager.py.

Select the Tomo Workflow Mode.

Import or Preprocess

If your tomograms were produced using the EMAN2 workflow:

If your tomograms were produced by other software, and you wish to do the segmentation in EMAN2:

Import Tomogram

Overview

The network training process may seem complicated at first, but in reality it is extremely simple. You will be training a neural network to look at your tomogram slice by slice, and for each pixel it will decide whether it does or does not look like part of the feature you are trying to identify. This feature is termed the Region of Interest (ROI). To accomplish this task, we must provide a few manual annotations to teach the network what it should be looking for. We must also provide some regions from the tomgram which do not contain the ROI as negative examples. That is, you are giving the network several examples of: "I am looking for THIS", and several more examples of: "It does NOT look like THIS".

For the network to perform this task well, it is critical that you be absolutely certain about the regions you select for training. The largest error people make when training the network is to include regions containing ambiguous density. If you select a tile containing a clear particle you wish to include, but the tile also contains some density you are unsure of, then you will be in a difficult situation. If you mark the ambiguous region as a particle, and it actually is something else, the trained network will not be very selective for the ROI. If you do not mark the ambiguous density, and it turns out to actually be the ROI, then you are trying to train the network to distinguish between the object you said was the ROI and the object you said was not, which may cause a training failure, or a network which misses many ROIs. The correct solution is not to include any tiles containing ambiguous density! Only include training tiles you are certain that you can annotate accurately in 2-D.

There will be other cases where it may be difficult to distinguish between two features. For example, at low resolution, the strong line of an actin filament, may look somewhat like the edge of a membrane on an organelle. If you train only a single network, and it is trained to recognize actin, then you apply it to a cellular tomogram, you will likely find many other features, such as membrane edges, and possibly the edge of the holy carbon film, get mis-identified as actin. While you may be able to more carefully train the actin network, or alter the threshold when segmenting, usually the better solution is to train a second network to recognize membranes and a third network to recognize the C-film hole, then compete the 3 networks against one another. While the vesicle may look somewhat like actin at low resolution, it should look more like a vesicle than actin, so the result of the competitive networks will be a more accurate segmentation of all 3 features.

We will begin the tutorial by training only a single network. You can then go back and follow the same process to train additional networks for other features, then finally at the end, use a program designed to compete these solutions against one another.

Select Training Examples

In this stage we will be manually selecting a few 64x64 pixel tiles from several tomograms, some of which containin the region of interest (ROI) you wish to train the neural network to identify. We will identify some 2-D tiles which contain the ROI and other tiles which do not contain the ROI at all. In the next step we will annotate the tiles conatining

Select Box training references. Press browse, and select the preproc version of any one imported tomogram. Leave boxsize as -1, and press Launch.

In a moment, three windows will appear: the tomogram view, the selected particles view, and a control-panel (labelled Options). Unlike the 2-D particle picking interface used in single particle analysis, which is almost identical, this program allows you to move through the various Z-slices of the tomogram. The examples you select will be 2-D, drawn from arbitrary slices throughout the 3-D tomogram. Additionally, this program permits selecting multiple classes of tiles from the same tomogram. For example, you may have one class of tiles containing the ROI and another class which does not contain the ROI. It is critical that you use a consistent naming convention for these different classes of particles, since we would like to have training references from multiple tomograms used for the network training.

Particle Output

Manually Annotate Samples

The next step is to manually identify the feature within each ROI.

Navigate to the Segment training references interface in the project manager window. For Particles, browse to the output ROIs you generated in the previous step. Leave Output blank, keep segment checked, and press Launch.

Two windows will appear, one containing images, and the other a control panel. You can navigate through the images similarly to looking through the slices of the tomogram above. The control panel will open with the Draw tab selected. Using the left mouse button, draw on each of the ROIs to exactly cover the feature of interest as best you can. If necessary you can always can go back to the ROI selection window and check the surroundings of each region to aid in segmentation.

Segment all of your ROIs. If there are multiple ROIs in one 2D sample, like multiple microtubules in this case, make sure to segment all of them. If you need to change the size of the pen, change both Pen Size and Pen Size2 to a larger or smaller number. You would like the marked region to match the feature as closely as possible, so reduce the pen size if it is larger than the feature.

When you are finished, simply close the windows. The segmentation file will be saved automatically as "*_seg.hdf" with the same base file name as your ROIs.

Segment Particles

Select Negative Samples

Next you need to identify regions in the tomogram which do not contain the feature of interest at all. Return to the same interface you used to select Positive examples, and press the Clear button in the e2boxer window. This deletes all of your previous selections (the positive examples), so make sure you have finished the previous steps before doing this.

In the tomogram window, select boxes that DON’T contain the feature of interest. You can select as many of these as you like (normally ~100). Try to include a wide variety of different non-feature cellular structures, empty space, gold fiducials and high-contrast carbon.

After finishing picking the negative samples, write the particle output in same way you did for the positive samples, but use a different Output Suffix, like "_bad".

Build Training Set

  1. Select the Build training set option in EMAN2.

  2. In particles_raw, select your "_ptcls" file.

  3. In particles_label, select your "_ptcls_seg" file.

  4. In boxes_negative, select your "_bad" file.

Leave trainset_output blank. Ncopy controls the number of particles in your training set. The default of 10 is fine, unless you want to do a faster run at the expense of accuracy.

Press Launch. The program will print "Done" in your Terminal when it has finished. The training set will be saved as the same name as the positive particles with "_trainset" suffix.

Note that if you have not configured Theano to use CUDA as described above, this will run on a single CPU and take a VERY long time.

Train Neural Network

Open Train the neural network in the project manager. In trainset, browse and choose the "_trainset" file.

The defaults for everything else in this window are sufficient to produce good results. To significantly shorten the length of the training (but potentially reduce the quality), reduce the number of iterations. Write the filename of the trained neural network output in the netout text box, and leave the from_trained box empty if it is the first training process.

Press Launch. The program will print a few numbers quickly at the beginning (this is to monitor the training process. Something is wrong if it prints really huge values), and then will notify you once it's completed each iteration. When it's finished, it will output the trained neural network in the specified netout file and samples of the training result in a file called "trainout_" followed by the netout file name.

After the training is finished, it is recommended to look at the training result before proceeding. Open the "trainout_*.hdf" file from the e2display window (use show stack), and you should see something like this.

Training Results

Zoom in or out a bit so there are 3*N images displayed in each row. For each three images, the first one is the ROI from the tomogram, the second is the corresponding manual segmentation, and the third is the neural network output for the same ROI after training. The neural network is considered well trained if the third image matches the second image. For the negative particles, both the second and the third images should be largely blank (the 3rd may contain some very weak features, this is ok).

If the training result looks somewhat wrong, go back and check your positive and negative training set first. Most significant errors are caused by training set errors, i.e. having some positive particles in the negative training set, or one of the positive training set is not correctly segmented. If the training result looks suboptimal (the segmentation is not clear enough but not totally wrong), you may consider continuing the training for a few rounds. To do this, go back to the Train the neural network panel, choose the previously trained network for the from_trained box and launch the training again. It is usually better to set a smaller learning rate in the continued training. Consider change value in the learnrate to ~0.001, or the displayed learning rate value at the last iteration of the previous training.

If you are satisfied with the result, go to the next step to segment the whole tomogram.

Apply to Tomograms

Finally, open Apply the neural network panel. Choose the tomogram you used to generate the boxes in the tomograms box, choose the saved neural network file (not the "trainout_" file, which is only used for visualization), and set the output filename. You can change the number of threads to use by adjusting the thread option. Keep in mind that using more threads will consume more memory as the tomogram slices are read in at the same time. For example, processing a 1k x 1k x 512 downsampled tomogram on 10 cores would use ~5 GB of RAM. Processing an unscaled 4k x 4k x 1k tomogram would increase RAM usage to ~24 GB. When this process finishes, you can open the output file in your favourite visualization software to view the segmentation.

Segmentation Result

To segment a different feature, just repeat the entire process for the each feature of interest. Make sure to use different file names (eg - _good2 and _bad2)! The trained network should generally work well on other tomograms using a similar specimen with similar microscope settings (clearly the A/pix value must be the same).

Merging multiple annotation results

Merging the results from multiple networks on a single tomogram can help resolve ambiguities, or correct regions which were apparently misassigned. For example, in the microtubule annotation shown above, the carbon edge is falsely recognized as a microtubule. An extra neural network can be trained to specifically recognize the carbon edge and its result can be competed against the microtubule annotation. A multi-level mask is produced after merging multiple annotation result in which the integer values in a voxel identify the type of feature the voxel contains. To merge multiple annotation results, simply run in the terminal:

Tips in selecting training samples

Good vs bad segments

Acknowledgement

Darius Jonasch, the first user of the tomogram segmentation protocol, provided many useful advices to make the workflow user-friendly. He also wrote a tutorial of the earlier version of the protocol, on which this tutorial is based.