Cellular membrane prediction model for volume SEM datasets. This model was trained on a FIB-SEM dataset to generically predict membranes (or organelle boundaries) in any volume SEM dataset.
The model was published as major component of the CebraEM workflow in hennies et al. 2023.
CebraNET is available in the Bioimage Model Zoo (bioimage.io) (CebraNET @bioimage.io, CebraNET @zenodo) where it runs in:
To optimize the output of CebraNET for your particular workflow, also see the following section about tweaking the input data.
Dependent on your input data, you might want to consider the following points as applicable:
- Although CebraNET is robust towards image noise sometimes a slight Gaussian smoothing can reduce the amount of false positive predictions
- CebraNET was trained to cope with mis-alignments of the input data stack. However this only works to a certain extent (few pixels), so consider locally re-aligning your data stack if the membrane prediction quality suffers
- Use isotropic data! If your data is anisotropic, re-scale it. For the SBEM dataset used in hennies et al. 2023, we scaled the data from 10 x 10 x 25 nm resolution to 10 nm isotropic resolution.
- Scale your data to alter the level of detail predicted by CebraNET. A resolution of 10 nm is well suited for larger organelles such as mitochondria. A resolution of 5 nm usually yields high level of detail to reconstruct the ER or other small, fine structured organelles.