ROIs and segmentation

ROIs and segmentation

Why necessary and why annoying?

  • As scientists we need to talk about specific spatial things
  • It takes a lot of time, potentially (manual vs. automated)
  • It is subjective (if manual)
  • It can really change analysis outcomes
  • De-identification of people's faces is important
  • To reduce multiple comparisons, if you have a priori hypotheses, that is useful and so you need to implement these hypotheses
  • The brain consists of multiple tissue types/stuff, so we obviously need to accurately select the part of the data we care about
  • Issue: criterion for success is currently not objective and subjective (and therefore somewhat of an "art"/"craft"). Eyeballs are important. Expertise is important.
  • If a newbie is doing manual segmentation, they might get it completely wrong?!! (So there is a risk of making things even worse!) Can even happen for non-newbies.

Issues/gripes/challenges:

  • Takes forever
  • Not clear how to interpret noisy functional measures
  • Risk of people suspecting you doing something dirty
  • People disagreeing on what is right
  • This is thankless work, unfortunately. (So, choose your battles.)

ROI formats:

  • Binary values (0 means not in, 1 means in)
  • Integer values (1–N indicate N different regions)
  • Volumes are a lot more straightforward (just a 3D grid); surfaces are a lot harder and generally clunkier to work with.

General "tools" and "methods" for dealing with ROIs:

  • Manually clicking voxels (e.g. for volume data) [e.g. ITK-SNAP, freeview, FSLeyes, mricron, BrainVoyager, etc.].
  • Using a brush size (many voxels at once)
  • Filling polygon shapes
  • Selective drawing (only modify specific voxels)
  • Adaptive brushing ("magic wand")
  • Click and automatically determine all "connected" voxels
  • Post-ROI 'morphological' operations: dilate, erode, union, intersection, smoothing, and other arithmetic operations
  • Experts will know how to interplay manual clicking and automatic morphological operations
  • Trackpad vs. mouse vs. tablets
  • Thresholding operations (e.g. value > 10)
  • Magical smoothing/filtering/operations to "clean up" the data prior to you clicking on it
  • Adjust the contrast/intensity for the specific part of the volume that you are staring at
  • Make sure to look at ALL 3 views to really understand the geometry of the structure. If your current slice is almost parallel to the structure, the visual may look weird.
  • Check your work. The eyeballs are always right.


Various types (scenarios):

  • T1 segmentation (e.g. for surfaces)
  • Defining a brain region from functional data
  • Anatomical segmentation, e.g. vessels
  • Atlas parcellations can deliver you ROIs
  • Identification of subcortical areas from anatomy (REALLY HARD)
  • Brain extraction from T1s (getting rid of face / skull)
  • Quick-and-dirty 3D ellipse to define on a volume


Principles/ideas that guide how you proceed:

  • Partial voluming - your single voxels might mix two different tissue types, due to limited resolution
  • Thermal noise - the data may be noisy
  • Bias field - there may be low-frequency changes in intensity across the volume.
  • One strategy is to correct the bias field before segmenting...
  • For functional data, consider looking both with and without thresholding
  • Be careful of 'kissing' gray matter (nearby in 3D space might be far away in cortical space)
  • Listen to podcasts while you do boring labor.