Desired output

Each cutout (bounding box) must produce three segmentations:

  1. Cell segmentation with semantic classification
  2. Mitochondria segmentation with uncertain labels
  3. Defect segmentation with semantic classification

Each cutout must be reviewed, unless otherwise noted. The review must inspect the 3D mesh and semantic class of each object in each segmentation. There must be proof-of-work for the review (spreadsheet or Neuroglancer link).

1. Cell segmentation

Rules

  • Paint fully over mitochondria — do not oversegment mitochondria
  • Paint fully over defects, as best as possible
  • Objects that span defects should be treated as a single object (a defect should not "split" an object)
  • Note uncertain segments: make sure any uncertain regions are included later in the defect segmentation as "uncertain" segments

Fat globule handling

  • Oversegment (paint as separate segments) fat globules that are not connected to myelin
  • Make sure there is no boundary between a fat globule and surrounding segments
  • It is better to overpaint the fat globule than to underpaint
  • Leave a boundary between myelin and neighboring objects
  • If the fat globule touches myelin, consider that fat globule part of the myelin object

Semantic classes

Use segment name to label each segment's compartment type based on the mesh of the object.

Class Description
dendrite Dendritic trunk, spine head, or spine neck (try to keep spines attached if possible)
axon Any type of axon, including axon initial segment or terminal
soma Cytosol of the soma other than the nucleus. If soma transitions to neurite within a single segment, mark as "merge between classes"
nucleus Nucleus of any cell
glia All glia cell types
extracellular space Try your best to distinguish this from glia
myelin Myelin sheath; includes fat globules connected to the myelin
myelin inner tongue Glia associated with the myelin on the inside of the sheath
myelin outer tongue Glia associated with the myelin on the outside of the sheath
blood vessel Interior of a blood vessel (mark surrounding cells as soma)
fat globule Dark objects that appear to be a dense tangle of membranes; often span multiple cells. If connected to myelin, do not oversegment.
merge between classes Object spans multiple classes within the bounding box (e.g., soma→dendrite, dendrite-axon merge error). Small overcoloring into glia or extracellular space should NOT be marked as merge.
uncertain Unable to determine semantic category, or unable to determine how an object continues through the volume

2. Mitochondria segmentation

  • Paint the mitochondria and its membrane. Do not overpaint into the cytoplasm of the parent cell.
  • If necessary, underpainting the mitochondria's membrane is okay.
  • If an object is a questionable mitochondria, add it and set segment name to "uncertain".

3. Defect segmentation

Paint the following defects, with a different color for each instance (e.g., each fold gets a different color):

Defect Class Notes
Folds Each fold is a separate segment
Cracks Each crack is a separate segment
Tears Paint the white and black part of any tear as a single segment
Missing section Some of the image in a section is black, but not obviously related to another defect (fold or tear)
Misaligned section A non-black section that noticeably breaks the flow of objects. Do not paint objects in cell or mitochondria segmentation for these sections.
Important: When painting defects, it is better to overpaint than underpaint.

Use segment name to label each segment with its semantic category from the list above.

Changes from v1.1

  • Clarify that cell segmentation should paint over defects as best as possible
  • Change fat globule handling: removed from defect segmentation, added to cell segmentation. Fat globules touching myelin are considered part of the myelin object.
  • Clarify that tears should be marked as a single segment
  • Add "misaligned sections" as a defect class

References

  • Cell Segmentation Visual Glossary v1.0
  • Voxel Painting Cell Segmentation Work Instructions — WebKnossos v1.0