FlyWire Proofreading Tips


In this document we’re going to go over some common problems you are likely to come across when proofreading in FlyWire. Using these tips you’ll learn how to better spot cell structure errors, and work around tricky spots in the dataset.

X-Shaped Mergers

These are what you’d traditionally think of as a merger, with a branch from one cell joined to the branch from another cell in a cross or “X” shape. They are usually fairly easy to detach with one cut.

H-Shaped or Parallel Mergers

Think of these types of mergers like the letter H, they are two parallel branches joined together somewhere in the middle.

These mergers can be a little more tricky to dissect, as sometimes they can sometimes have multiple merge points; or “phantom” merge points, where they seem to have a connection point in the 3D, but are actually connected at a different point.

Using the Find Path tool can help you find the merger bridge in these types of cells.

Misalignments

Misalignments are errors created in the EM images where one slide in the dataset is shifted (misaligned) from the previous slide. Sometimes these EM errors can be more complex, and may contain multiple misaligned slides, poorly imaged slides, or blackout spaces between slides. The AI often has trouble traversing these “bad” spots, and will typically terminate early or make a poor connection after the “jump.”

To combat these errors we have figured out a helpful trick that utilizes the annotation tool.

  1. Find a unique object that stands out within your current EM view
  2. Pinpoint that same object in the EM after the slide jump
  3. Drop annotation points into that object and its continuation both before and after the jump
  4. Toggle between the two points by clicking back and forth on each point
  5. This should visually realign the two slides, and help you to find the continuation on your original neuron!

For trickier misalignments without good visual “landmarks” to connect them, you may have to search outside of your current visual field. Watch this advanced misalignment tutorial to see one way to approach these situations.

Twig-to-twig Mergers

Twig-to-twig mergers are two cells, or cell pieces, connected via “twigs” – the skinny termination points on some dendritic arbors.

These types of mergers are usually easy to detach, but can sometimes be tricky to locate, especially if you’re dealing with particularly dense or “twiggy” dendritic arbors.

The “Find Path” tool can be utilized in the case of tricky twig mergers to locate the merger origin point.

Synaptic Invagination

Mergers can often occur at the synapse point between one cell and another. At this junction one cell branch may begin to wrap around its synapse partner, creating somewhat blurry visual boundaries that can be hard for the AI to distinguish.

Disentangling these merger types does not always result in a “clean” cut, but just do your best. Resulting splits are acceptable so long as the majority of the material from each cell is retained.

Using Autofasciculation to Find Somas

Some cell types have autofasciculating soma tracts. This means that the neurite that leads to the soma branches off from the main backbone, but then continues to flow parallel to it for some distance.

Because the soma tract is often quite thin, it has a tendency to get lost in the EM images or confuse the AI.

It can also be difficult for human proofreaders to locate this break point. While you might be tempted to just click around in the 3D trying to “Eye Spy” the continuation, there is a better way.

You can use the EM images to look for an object running parallel to your cell’s backbone. Often times, this is the soma tract! The soma usually connects closer to the dendritic arbor than the axon, so starting your search near the dendrites is usually your best bet.

Gaps in Dendritic Arbors

Before you finish proofreading your cell, you should observe its overall shape in the 3D model. Take note – does the dendritic arbor look complete? Some cells may have just a few dendrites and atypical shapes. However, for cells with larger arbors, it is often apparent if there are sizable “gaps” between branches. These unfilled areas are often spots where a branch is missed.

While you want to take note of suspicious looking attachments or broken spots along branches, it is also important to notice what is not there. Taking a holistic view of your neuron can be just as helpful to diagnose problems as it is to observe the fine details.

MIA Connections

Sometimes very thin neurites go completely MIA. They end up pinched between other structures, and become invisible in the EM images. When this happens it can be very difficult to find a connection.

When observing optic lobe cells in FlyWire, we see these kinds of splits often between the lamina and the medulla neuropils, usually leading to detached somas.

In these situations your best bet is to utilize neurons running parallel to your cell to try and find the soma (or any missing extension) on the other side. You may have to add a number of potential extensions before you find one that makes sense.

Identifying Path Swaps

When proofreading, be on the lookout for path swaps. These are errors that can be created by the AI or another human in which the cell has a false continuation.

These path swaps usually occur because of a visual or technical error in the EM imagery. As we’ve seen before, groups of neurons often travel similar paths, and any small shift in the dataset images can cause the AI (or proofreader) to think the path continues where it doesn’t.

Keep an eye open for these kinds of mistakes. They can be identified in the a few different ways, including looking for visual breaks in the 3D, looking for appropriate directionality of the continuation, and utilizing knowledge of cell types or structures to determine if the continuation looks like it “fits” your cell.

Combining Multiple Techniques (Advanced)

To finish off, let’s take a look at an example of a cell where multiple techniques are used to identify a cell error.

In this tutorial you’ll that having knowledge of cell types, being able to identify similar nearby cells, and using “process of elimination” (adding and subtracting nearby objects) are all employed to work towards a solution for a missing extension.

Thanks for reading!

We hope these tips and tricks were helpful! If you’ve watched all (or even a few) of these videos, you’re well on your way to becoming an expert brainmapper. Check out these cells circuits that were mapped by proofreaders just like you!


Leave a Reply

Your email address will not be published. Required fields are marked *