The method can achieve decent segmentation in cases where you have a solid amount of noisy images and only a few of them are labeled.
Here are the relevant resources:
- for the theory, read the paper
- for Python instructions (training and prediction), read this page
- for Fiji instructions (training and prediction), read this page
- video and FAQ from the NEUBIAS Academy webinar
Timelapse video from a ~7h training + prediction session
- 2D only for now, the 3D option will very likely follow soon
Notes on the preprocessing in Fiji
Please read the
Prepare your data section in the Wiki for how to create the directories used to run the training. Here are some additional notes since the page is currently read-only and I can’t improve it:
You need two directories for the training (raw and labelings) where the names in the labelings directory match the source name in the raw directory, but not all raw images have to be labeled:
… and two directories for the validation (raw and labelings), which should be a smaller set of matching raw images and labelings:
In contrast to what I said in the webinar and what the wiki states, each labeling image of type
intdoes not necessarily need an individual index per object, you can also just use
0for background and
1for foreground. You only need to use different indices for different objects if they touch each other (e.g.
cell2=2) and if you care about distinguishing them. This way you give the training a chance to learn that there should be a border in between.