Labels collapsed?

[stebo85]

Dear @maxdiefenbach,

I have a weird problem in some datasets where the labels are suddenly "flipped/collapsed" for single slices, here is an example:

Slice 40 - "broken":

Slice 41 - ok:

Labels are:
3 instead of 0
0 instead of 3
0 instead of 1
0 instead of 2
2 instead of 1

Do you have an idea where I could start looking for the problem?

Thank you
Cheers
Steffen

[stebo85]

and this is the line where the error seems to surface: vatsatseg.py:335: RuntimeWarning: invalid value encountered in long_scalars
nvoxels_skin_torso_prev)

[maxdiefenbach]

Dear @maxdiefenbach,

I have a weird problem in some datasets where the labels are suddenly "flipped/collapsed" for single slices, here is an example:

Slice 40 - "broken":

Slice 41 - ok:

Labels are:
3 instead of 0
0 instead of 3
0 instead of 1
0 instead of 2
2 instead of 1

Do you have an idea where I could start looking for the problem?

I have never observed a "flipped" case like you show on the broken slice 40.
The "collapsing" happens when the most outer left and right VAT regions are not connected. In such a case, one should probably join the two disconnected VAT regions of largest bounding box area instead of only taking the one connected region with largest bounding box as happening here:

╭──────── #290 ─ vatsatseg.py ──
│ sat_label = extract_largest_label(f_mask)
│ # TODO: what to do if sat_label is not continuous?
│ # maybe check for a label with a bbox area close to the sat_label?
╰──────── #292 ─

Maybe the two problems are connected, though.

Thank you
Cheers
Steffen

I