At the Jackson Laboratory we’ve been working on a dask-based labeling tool to address chunk-wise labeling of largescale images with objects that “touch”.
In contrast with the label module in dask-image, this should handle the output of deep-learning segmentation methods, such as cellpose, stardist, hover-net, etc. on touching cell images that cannot be addressed by simply merging connected components like in the current label method.
The use case is to segment large-scale images (e.g. Whole slide images) in an efficient manner.
The tool can be found in this repo https://github.com/TheJacksonLaboratory/dask_relabeling.git and there are a pair of examples of the tool.
I would appreciate any feedback and also know if people are interested in doing this kind of large-scale analysis.