Different methods were used to segment the structures of interest in the EM-Volume:
- cells: Cell membranes were segmented with a 3d U-Net trained with long-range affinity loss. Based on these predictions, instance segmentation was performed via Lifted Multicut workflow, including priors from the nucleus segmentation.
- chromatin: Ilastik pixel classification was used, restricted to the segmented nuclei.
- cilia: Cilia boundaries were segmented with a 3d U-Net trained with long-range affinity loss. Based on these predictions, instance segmentation was performed via Mutex Watershed and Block-wise Multicut.
- cuticle: Cuticle boundaries were segmented with a 3d U-Net trained with long-range affinity loss. Based on these predictions, instance segmentation was performed via Mutex Watershed and Block-wise Multicut.
- ganglia: The ganglia were segmented by manually selecting the ids of segmented cells.
- nuclei: Nuclear membranes were segmented with a 3d U-Net trained with long-range affinity loss. Based on these predictions, instance segmentation was performed via Mutex Watershed and Block-wise Multicut.
- tissue: Tissue and regions were segmented using the Ilastik carving workflow
If you use any of the segmentation functionality provided, please cite the main publication AND the appropriate methods. For most of these methods, the scalable implementations in cluster tools were used.
Training data and weights for the 3d U-Nets are available on zenodo:
- cells: Training Data, Weights
- cilia: Training Data, Weights
- cuticle: Training Data, Weights
- nuclei: Training Data, Weights
The models are also available on bioimage.io in order to run them in Deep Ilastik (still in beta).
The ilastik project and training data for the chromatin segmentation are also available on zenodo, as well as the ilastik projects for carving out tissue/body parts and the animal outline.
In addition, we provide scripts to validate cell and nucleus segmentations. The validation data, which consists of annotations for nuclei and cell soma from 8 domain experts, is available on zenodo.