Training data, picking and choosing examples

[Rick Giuly]

Hi Alex and Kurt,

If you're using the contour processing stage in Cytoseg, there is a
caveat with marking the positive and negative pixels. For the training
set, it's best to pick *all* the pixels in the training set that are to
be positive.

For the negatives, completeness doesn't matter as much. You can pick and
choose the pixels that are to be negative examples arbitrarily. (And,
picking them carefully may give better results.)


Here's the reason:

This caveat exists because of the "contour processing stage" of
processing where Cytoseg will run the voxel classifier on the whole
training dataset and then figure out what shape of contours likely
belong to mito. It figures out which ones are "good" based on the same
training data you use for pixel classification. It will end up finding a
bunch of "good" contours which be recorded as "bad" unless the positives
are all marked.

PS: Ilastik doesn't have a contour processing step, just the pixel/voxel
processing step, so this caveat doesn't exist in that program.


-Rick