There is no particular theory. It is more of a way to avoid errors in the user annotation. What I had said during the VM is that it avoids creating errors in the geometry of the cells and is actually less effort.
Please see the image here:
What we experienced is that
- Users seeing the edges of the square tended to ‘taper’ the cells by rounding them as they reached the edge.
- It takes longer to make sure that the cell is well segmented and not outside the edge than to draw the entire cell. This has the added benefit that it forces the user not to treat the edge cells somehow differently than the cells inside. It is more a matter of psychology that helps to reach better results, rather than a mathematically founded explanation.
For illustration note that the drawings on the right of the image I posted are also perfectly correct for training. So long as the annotation follows the cell/nuclear boundary properly, there is no issue.
I hope this clears it a bit better.